Module monk.pytorch.finetune.level_8_layers_main
Expand source code
import math
from pytorch.finetune.imports import *
from system.imports import *
from pytorch.finetune.level_7_aux_main import prototype_aux
class prototype_layers(prototype_aux):
@accepts("self", verbose=int, post_trace=True)
@TraceFunction(trace_args=True, trace_rv=True)
def __init__(self, verbose=1):
super().__init__(verbose=verbose);
#####################################################################################################################################
def append_linear(self, num_neurons=False, final_layer=False):
'''
Append dense (fully connected) layer to base network in transfer learning
Args:
num_neurons (int): Number of neurons in the dense layer
final_layer (bool): If True, then number of neurons are directly set as number of classes in dataset for single label type classification
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
if(not num_neurons):
num_neurons = self.system_dict["dataset"]["params"]["num_classes"];
self.system_dict = layer_linear(self.system_dict, num_neurons=num_neurons, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_dropout(self, probability=0.5, final_layer=False):
'''
Append dropout layer to base network in transfer learning
Args:
probability (float): Droping probability of neurons in next layer
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = layer_dropout(self.system_dict, probability=probability, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_elu(self, alpha=1.0, final_layer=False):
'''
Append exponential linear unit activation to base network in transfer learning
Args:
alpha (float): Multiplicatve factor.
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_elu(self.system_dict, alpha=alpha, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_hardshrink(self, lambd=0.5, final_layer=False):
'''
Append hardshrink activation to base network in transfer learning
Args:
lambd (float): thereshold limit value
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_hardshrink(self.system_dict, lambd=lambd, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_hardtanh(self, min_val=-1.0, max_val=1.0, final_layer=False):
'''
Append hardtanh activation to base network in transfer learning
Args:
min_val (float): lower thereshold limit value
max_val (float): upper thereshold limit value
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_hardtanh(self.system_dict, min_val=min_val, max_val=max_val, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_leakyrelu(self, negative_slope=0.01, final_layer=False):
'''
Append Leaky - ReLU activation to base network in transfer learning
Args:
negative_slope (float): Multiplicatve factor towards negative spectrum of real numbers.
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_leakyrelu(self.system_dict, negative_slope=negative_slope, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_logsigmoid(self, final_layer=False):
'''
Append logsigmoid activation to base network in transfer learning
Args:
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_logsigmoid(self.system_dict, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_prelu(self, num_parameters=1, init=0.25, final_layer=False):
'''
Append Learnable parameerized rectified linear unit activation to base network in transfer learning
Args:
init (float): Initialization value for multiplicatve factor towards negative spectrum of real numbers.
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_prelu(self.system_dict, num_parameters=num_parameters, init=init, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_relu(self, final_layer=False):
'''
Append rectified linear unit activation to base network in transfer learning
Args:
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_relu(self.system_dict, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_relu6(self, final_layer=False):
'''
Append rectified linear unit activation - 6 to base network in transfer learning
Args:
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_relu6(self.system_dict, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_rrelu(self, lower=0.125, upper=0.333, final_layer=False):
'''
Append randomized rectified linear unit activation to base network in transfer learning
Args:
lower (float): Lower limit of uniform distribution for the ramdomized multiplicative factor
upper (float): Upper limit of uniform distribution for the ramdomized multiplicative factor
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_rrelu(self.system_dict, lower=lower, upper=upper, final_layer=final_layer)
#####################################################################################################################################
#####################################################################################################################################
def append_selu(self, final_layer=False):
'''
Append scaled exponential linear unit activation to base network in transfer learning
Args:
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_selu(self.system_dict, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_celu(self, alpha=1.0, final_layer=False):
'''
Append continuously differentiable exponential linear unit activation to base network in transfer learning
Args:
alpha (float): Multiplicative factor
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_celu(self.system_dict, alpha=alpha, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_sigmoid(self, final_layer=False):
'''
Append rectified linear unit activation to base network in transfer learning
Args:
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_sigmoid(self.system_dict, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_softplus(self, beta=1, threshold=20, final_layer=False):
'''
Append softplus activation to base network in transfer learning
Args:
beta (float): Multiplicative factor
threshold (int): softplus (thresholded relu) limit
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_softplus(self.system_dict, beta=beta, threshold=threshold, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_softshrink(self, lambd=0.5, final_layer=False):
'''
Append softshrink activation to base network in transfer learning
Args:
lambd (float): thereshold limit value
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_softshrink(self.system_dict, lambd=lambd, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_softsign(self, final_layer=False):
'''
Append tanh activation to base network in transfer learning
Args:
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_softsign(self.system_dict, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_tanh(self, final_layer=False):
'''
Append tanh activation to base network in transfer learning
Args:
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_tanh(self.system_dict, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_tanhshrink(self, final_layer=False):
'''
Append tanhshrink activation to base network in transfer learning
Args:
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_tanhshrink(self.system_dict, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_threshold(self, threshold, value, final_layer=False):
'''
Append threshold activation to base network in transfer learning
Args:
threshold (float): thereshold limit value
value (float): replacement value if input is off-limits
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_threshold(self.system_dict, threshold, value, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_softmin(self, final_layer=False):
'''
Append softmin activation to base network in transfer learning
Args:
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_softmin(self.system_dict, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_softmax(self, final_layer=False):
'''
Append softmax activation to base network in transfer learning
Args:
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_softmax(self.system_dict, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def append_logsoftmax(self, final_layer=False):
'''
Append log-softmax activation to base network in transfer learning
Args:
final_layer (bool): Indicator that this layer marks the end of network.
Returns:
None
'''
if(self.system_dict["model"]["final_layer"]):
msg = "Cannot append more layers.\n";
msg += "Tip: Previously appended layer termed as final layer";
raise ConstraintError(msg);
else:
self.system_dict = activation_logsoftmax(self.system_dict, final_layer=final_layer);
#####################################################################################################################################
#####################################################################################################################################
def convolution1d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out",
groups=1, dilation=1, use_bias=True, layout='NCW', uid=None):
'''
Append 1d-convolution to custom network
Args:
output_channels (int): Number of output features for this layers
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
groups (int): Number of groups for grouped convolution
dilation (int): Factor for dilated convolution
use_bias (bool): If True, learnable bias is added
layout (str): Either of these values (order)
1) 'NCW' - order
2) 'NWC' - order
- N: Number of elements in batches
- C: Number of channels
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp={};
tmp["uid"] = uid;
tmp["name"] = "convolution1d";
tmp["params"] = {};
tmp["params"]["output_channels"] = output_channels;
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["groups"] = groups;
tmp["params"]["dilation"] = dilation;
tmp["params"]["use_bias"] = use_bias;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def convolution2d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out",
groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None):
'''
Append 2d-convolution to custom network
Args:
output_channels (int): Number of output features for this layers
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
groups (int): Number of groups for grouped convolution
dilation (int): Factor for dilated convolution
use_bias (bool): If True, learnable bias is added
layout (str): Either of these values (order)
1) 'NCHW' - Order
2) 'NHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp={};
tmp["uid"] = uid;
tmp["name"] = "convolution2d";
tmp["params"] = {};
tmp["params"]["output_channels"] = output_channels;
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["groups"] = groups;
tmp["params"]["dilation"] = dilation;
tmp["params"]["use_bias"] = use_bias;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def convolution(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out",
groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None):
'''
Append 2d-convolution to custom network
Args:
output_channels (int): Number of output features for this layers
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
groups (int): Number of groups for grouped convolution
dilation (int): Factor for dilated convolution
use_bias (bool): If True, learnable bias is added
layout (str): Either of these values (order)
1) 'NCHW' - Order
2) 'NHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp={};
tmp["uid"] = uid;
tmp["name"] = "convolution2d";
tmp["params"] = {};
tmp["params"]["output_channels"] = output_channels;
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["groups"] = groups;
tmp["params"]["dilation"] = dilation;
tmp["params"]["use_bias"] = use_bias;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def convolution3d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out",
groups=1, dilation=1, use_bias=True, layout='NCDHW', uid=None):
'''
Append 3d-convolution to custom network
Args:
output_channels (int): Number of output features for this layers
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
groups (int): Number of groups for grouped convolution
dilation (int): Factor for dilated convolution
use_bias (bool): If True, learnable bias is added
layout (str): Either of these values (order)
1) 'NCDHW' - Order
2) 'NDHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- D: Depth of features in layers
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp={};
tmp["uid"] = uid;
tmp["name"] = "convolution3d";
tmp["params"] = {};
tmp["params"]["output_channels"] = output_channels;
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["groups"] = groups;
tmp["params"]["dilation"] = dilation;
tmp["params"]["use_bias"] = use_bias;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def transposed_convolution1d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out",
output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCW', uid=None):
'''
Append 1d-transposed-convolution to custom network
Args:
output_channels (int): Number of output features for this layers
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
output_padding (int): Additional padding applied to output
groups (int): Number of groups for grouped convolution
dilation (int): Factor for dilated convolution
use_bias (bool): If True, learnable bias is added
layout (str): Either of these values (order)
1) 'NCW' - order
2) 'NWC' - order
- N: Number of elements in batches
- C: Number of channels
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp={};
tmp["uid"] = uid;
tmp["name"] = "transposed_convolution1d";
tmp["params"] = {};
tmp["params"]["output_channels"] = output_channels;
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["output_padding"] = output_padding;
tmp["params"]["groups"] = groups;
tmp["params"]["dilation"] = dilation;
tmp["params"]["use_bias"] = use_bias;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def transposed_convolution(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out",
output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None):
'''
Append 2d-transposed-convolution to custom network
Args:
output_channels (int): Number of output features for this layers
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
output_padding (int): Additional padding applied to output
groups (int): Number of groups for grouped convolution
dilation (int): Factor for dilated convolution
use_bias (bool): If True, learnable bias is added
layout (str): Either of these values (order)
1) 'NCHW' - Order
2) 'NHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp={};
tmp["uid"] = uid;
tmp["name"] = "transposed_convolution2d";
tmp["params"] = {};
tmp["params"]["output_channels"] = output_channels;
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["output_padding"] = output_padding;
tmp["params"]["groups"] = groups;
tmp["params"]["dilation"] = dilation;
tmp["params"]["use_bias"] = use_bias;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def transposed_convolution2d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out",
output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None):
'''
Append 2d-transposed-convolution to custom network
Args:
output_channels (int): Number of output features for this layers
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
output_padding (int): Additional padding applied to output
groups (int): Number of groups for grouped convolution
dilation (int): Factor for dilated convolution
use_bias (bool): If True, learnable bias is added
layout (str): Either of these values (order)
1) 'NCHW' - Order
2) 'NHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp={};
tmp["uid"] = uid;
tmp["name"] = "transposed_convolution2d";
tmp["params"] = {};
tmp["params"]["output_channels"] = output_channels;
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["output_padding"] = output_padding;
tmp["params"]["groups"] = groups;
tmp["params"]["dilation"] = dilation;
tmp["params"]["use_bias"] = use_bias;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def transposed_convolution3d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out",
output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCDHW', uid=None):
'''
Append 3d-transposed-convolution to custom network
Args:
output_channels (int): Number of output features for this layers
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
output_padding (int): Additional padding applied to output
groups (int): Number of groups for grouped convolution
dilation (int): Factor for dilated convolution
use_bias (bool): If True, learnable bias is added
layout (str): Either of these values (order)
1) 'NCDHW' - Order
2) 'NDHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- D: Depth of features in layers
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp={};
tmp["uid"] = uid;
tmp["name"] = "transposed_convolution3d";
tmp["params"] = {};
tmp["params"]["output_channels"] = output_channels;
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["output_padding"] = output_padding;
tmp["params"]["groups"] = groups;
tmp["params"]["dilation"] = dilation;
tmp["params"]["use_bias"] = use_bias;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def max_pooling1d(self, kernel_size=2, stride=None, padding=0, dilation=1,
return_indices=False, ceil_mode=False, layout='NCW', uid=None):
'''
Append 1d-max-pooling to custom network
Args:
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
dilation (int): Factor for dilated pooling
return_indices (bool): Fixed value set as False
ceil_mode (bool): If True, apply ceil math operation post pooling
layout (str): Either of these values (order)
1) 'NCW' - order
2) 'NWC' - order
- N: Number of elements in batches
- C: Number of channels
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "max_pooling1d";
tmp["params"] = {};
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["dilation"] = dilation;
tmp["params"]["return_indices"] = return_indices;
tmp["params"]["ceil_mode"] = ceil_mode;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def max_pooling2d(self, kernel_size=2, stride=None, padding=0, dilation=1,
return_indices=False, ceil_mode=False, layout='NCHW', uid=None):
'''
Append 2d-max-pooling to custom network
Args:
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
dilation (int): Factor for dilated pooling
return_indices (bool): Fixed value set as False
ceil_mode (bool): If True, apply ceil math operation post pooling
layout (str): Either of these values (order)
1) 'NCHW' - Order
2) 'NHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "max_pooling2d";
tmp["params"] = {};
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["dilation"] = dilation;
tmp["params"]["return_indices"] = return_indices;
tmp["params"]["ceil_mode"] = ceil_mode;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def max_pooling(self, kernel_size=2, stride=None, padding=0, dilation=1,
return_indices=False, ceil_mode=False, layout='NCHW', uid=None):
'''
Append 2d-max-pooling to custom network
Args:
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
dilation (int): Factor for dilated pooling
return_indices (bool): Fixed value set as False
ceil_mode (bool): If True, apply ceil math operation post pooling
layout (str): Either of these values (order)
1) 'NCHW' - Order
2) 'NHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "max_pooling2d";
tmp["params"] = {};
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["dilation"] = dilation;
tmp["params"]["return_indices"] = return_indices;
tmp["params"]["ceil_mode"] = ceil_mode;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def max_pooling3d(self, kernel_size=2, stride=None, padding=0, dilation=1,
return_indices=False, ceil_mode=False, layout='NCDHW', uid=None):
'''
Append 3d-max-pooling to custom network
Args:
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
dilation (int): Factor for dilated pooling
return_indices (bool): Fixed value set as False
ceil_mode (bool): If True, apply ceil math operation post pooling
layout (str): Either of these values (order)
1) 'NCDHW' - Order
2) 'NDHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- D: Depth of features in layers
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "max_pooling3d";
tmp["params"] = {};
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["dilation"] = dilation;
tmp["params"]["return_indices"] = return_indices;
tmp["params"]["ceil_mode"] = ceil_mode;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def average_pooling1d(self, kernel_size=2, stride=None, padding=0, dilation=1,
return_indices=False, ceil_mode=False, include_padding_in_calculation=True,
layout='NCW', uid=None):
'''
Append 1d-average-pooling to custom network
Args:
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
dilation (int): Factor for dilated pooling
return_indices (bool): Fixed value set as False
include_padding_in_calculation (bool): If True, padding will be considered.
ceil_mode (bool): If True, apply ceil math operation post pooling
layout (str): Either of these values (order)
1) 'NCW' - order
2) 'NWC' - order
- N: Number of elements in batches
- C: Number of channels
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "average_pooling1d";
tmp["params"] = {};
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["dilation"] = dilation;
tmp["params"]["return_indices"] = return_indices;
tmp["params"]["ceil_mode"] = ceil_mode;
tmp["params"]["include_padding_in_calculation"] = include_padding_in_calculation;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def average_pooling2d(self, kernel_size=2, stride=None, padding=0, dilation=1,
return_indices=False, ceil_mode=False, include_padding_in_calculation=True,
layout='NCHW', uid=None):
'''
Append 2d-average-pooling to custom network
Args:
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
dilation (int): Factor for dilated pooling
return_indices (bool): Fixed value set as False
ceil_mode (bool): If True, apply ceil math operation post pooling
include_padding_in_calculation (bool): If True, padding will be considered.
layout (str): Either of these values (order)
1) 'NCHW' - Order
2) 'NHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "average_pooling2d";
tmp["params"] = {};
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["dilation"] = dilation;
tmp["params"]["return_indices"] = return_indices;
tmp["params"]["ceil_mode"] = ceil_mode;
tmp["params"]["include_padding_in_calculation"] = include_padding_in_calculation;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def average_pooling(self, kernel_size=2, stride=None, padding=0, dilation=1,
return_indices=False, ceil_mode=False, include_padding_in_calculation=True,
layout='NCHW', uid=None):
'''
Append 2d-average-pooling to custom network
Args:
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
dilation (int): Factor for dilated pooling
return_indices (bool): Fixed value set as False
ceil_mode (bool): If True, apply ceil math operation post pooling
include_padding_in_calculation (bool): If True, padding will be considered.
layout (str): Either of these values (order)
1) 'NCHW' - Order
2) 'NHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "average_pooling2d";
tmp["params"] = {};
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["dilation"] = dilation;
tmp["params"]["return_indices"] = return_indices;
tmp["params"]["ceil_mode"] = ceil_mode;
tmp["params"]["include_padding_in_calculation"] = include_padding_in_calculation;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def average_pooling3d(self, kernel_size=2, stride=None, padding=0, dilation=1,
return_indices=False, ceil_mode=False, include_padding_in_calculation=True,
layout='NCDHW', uid=None):
'''
Append 3d-average-pooling to custom network
Args:
kernel_size (int, tuple): kernel matrix size
stride (int, tuple): kernel movement stride
padding (int, tuple, str): Zero padding applied to input
1) "in_eq_out": Automated padding applied to keep output shape same as input
2) integer or tuple value: Manually add padding
dilation (int): Factor for dilated pooling
return_indices (bool): Fixed value set as False
ceil_mode (bool): If True, apply ceil math operation post pooling
include_padding_in_calculation (bool): If True, padding will be considered.
layout (str): Either of these values (order)
1) 'NCDHW' - Order
2) 'NDHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- D: Depth of features in layers
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "average_pooling3d";
tmp["params"] = {};
tmp["params"]["kernel_size"] = kernel_size;
tmp["params"]["stride"] = stride;
tmp["params"]["padding"] = padding;
tmp["params"]["dilation"] = dilation;
tmp["params"]["return_indices"] = return_indices;
tmp["params"]["ceil_mode"] = ceil_mode;
tmp["params"]["include_padding_in_calculation"] = include_padding_in_calculation;
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def global_max_pooling1d(self, layout='NCW', uid=None):
'''
Append 1d-global-max-pooling to custom network
Args:
layout (str): Either of these values (order)
1) 'NCW' - order
2) 'NWC' - order
- N: Number of elements in batches
- C: Number of channels
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "global_max_pooling1d";
tmp["params"] = {};
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def global_max_pooling2d(self, layout='NCHW', uid=None):
'''
Append 2d-global-max-pooling to custom network
Args:
layout (str): Either of these values (order)
1) 'NCHW' - Order
2) 'NHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "global_max_pooling2d";
tmp["params"] = {};
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def global_max_pooling(self, layout='NCHW', uid=None):
'''
Append 2d-global-max-pooling to custom network
Args:
layout (str): Either of these values (order)
1) 'NCHW' - Order
2) 'NHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "global_max_pooling2d";
tmp["params"] = {};
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def global_max_pooling3d(self, layout='NCDHW', uid=None):
'''
Append 3d-global-max-pooling to custom network
Args:
layout (str): Either of these values (order)
1) 'NCDHW' - Order
2) 'NDHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- D: Depth of features in layers
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "global_max_pooling3d";
tmp["params"] = {};
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def global_average_pooling1d(self, layout='NCW', uid=None):
'''
Append 1d-global-average-pooling to custom network
Args:
layout (str): Either of these values (order)
1) 'NCW' - order
2) 'NWC' - order
- N: Number of elements in batches
- C: Number of channels
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "global_average_pooling1d";
tmp["params"] = {};
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def global_average_pooling2d(self, layout='NCHW', uid=None):
'''
Append 2d-global-average-pooling to custom network
Args:
layout (str): Either of these values (order)
1) 'NCHW' - Order
2) 'NHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "global_average_pooling2d";
tmp["params"] = {};
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def global_average_pooling(self, layout='NCHW', uid=None):
'''
Append 2d-global-average-pooling to custom network
Args:
layout (str): Either of these values (order)
1) 'NCHW' - Order
2) 'NHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "global_average_pooling2d";
tmp["params"] = {};
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def global_average_pooling3d(self, layout='NCDHW', uid=None):
'''
Append 3d-global-average-pooling to custom network
Args:
layout (str): Either of these values (order)
1) 'NCDHW' - Order
2) 'NDHWC' - Order
- N: Number of elements in batches
- C: Number of channels
- D: Depth of features in layers
- H: Height of features in layers
- W: Number of features in layers
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "global_average_pooling3d";
tmp["params"] = {};
tmp["params"]["layout"] = layout;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def fully_connected(self, units=512, use_bias=True, flatten=True, uid=None):
'''
Append fully-connected (dense) layer to custom network
Args:
units (int): Number of neurons in the layer
use_bias (bool): If True, learnable bias is added
flatten (bool): Fixed to True
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "fully_connected";
tmp["params"] = {};
tmp["params"]["units"] = units;
tmp["params"]["use_bias"] = use_bias;
tmp["params"]["flatten"] = flatten;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def dropout(self, drop_probability=0.2, axes=(), uid=None):
'''
Append dropout layer to custom network
Args:
drop_probability (float): Probability for not considering neurons in the output
axes (tuple): Channel axis to implement dropout over
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "dropout";
tmp["params"] = {};
tmp["params"]["drop_probability"] = drop_probability;
tmp["params"]["axes"] = axes;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def flatten(self, uid=None):
'''
Append flatten layer to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "flatten";
tmp["params"] = {};
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def identity(self, uid=None):
'''
Append identity layer to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "identity";
tmp["params"] = {};
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def batch_normalization(self, moving_average_momentum=0.1, epsilon=0.00001, use_trainable_parameters=True,
activate_scale_shift_operation=False, uid=None):
'''
Append batch normalization layer to custom network
Args:
moving_average_momentum (float): Normalization momentum value
epsilon (float): Value to avoid division by zero
use_trainable_paramemetrs (bool): If True, batch norm turns into a trainable layer
activate_scale_shift_operation (bool): Fixed status - False
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "batch_normalization";
tmp["params"] = {};
tmp["params"]["moving_average_momentum"] = moving_average_momentum;
tmp["params"]["epsilon"] = epsilon;
tmp["params"]["use_trainable_parameters"] = use_trainable_parameters;
tmp["params"]["activate_scale_shift_operation"] = activate_scale_shift_operation;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def instance_normalization(self, moving_average_momentum=0.1, epsilon=0.00001, use_trainable_parameters=False,
uid=None):
'''
Append instace normalization layer to custom network
Args:
moving_average_momentum (float): Normalization momentum value
epsilon (float): Value to avoid division by zero
use_trainable_paramemetrs (bool): If True, batch norm turns into a trainable layer
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "instance_normalization";
tmp["params"] = {};
tmp["params"]["moving_average_momentum"] = moving_average_momentum;
tmp["params"]["epsilon"] = epsilon;
tmp["params"]["use_trainable_parameters"] = use_trainable_parameters;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def layer_normalization(self, moving_average_momentum=0.9, epsilon=0.00001, use_trainable_parameters=True,
uid=None):
'''
Append layer normalization layer to custom network
Args:
moving_average_momentum (float): Normalization momentum value
epsilon (float): Value to avoid division by zero
use_trainable_paramemetrs (bool): If True, batch norm turns into a trainable layer
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "layer_normalization";
tmp["params"] = {};
tmp["params"]["epsilon"] = epsilon;
tmp["params"]["use_trainable_parameters"] = use_trainable_parameters;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def add(self, uid=None):
'''
Append elementwise addition layer to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "add";
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def concatenate(self, uid=None):
'''
Append concatenation layer to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "concatenate";
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def relu(self, uid=None):
'''
Append rectified linear unit activation to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "relu";
tmp["params"] = {};
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def sigmoid(self, uid=None):
'''
Append sigmoid activation to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "sigmoid";
tmp["params"] = {};
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def tanh(self, uid=None):
'''
Append tanh activation to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "tanh";
tmp["params"] = {};
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def softplus(self, beta=1, threshold=20, uid=None):
'''
Append softplus activation to custom network
Args:
beta (float): Multiplicative factor
threshold (int): softplus (thresholded relu) limit
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "softplus";
tmp["params"] = {};
tmp["params"]["beta"] = beta;
tmp["params"]["threshold"] = threshold;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def softsign(self, uid=None):
'''
Append softsign activation to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "softsign";
tmp["params"] = {};
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def elu(self, alpha=1.0, uid=None):
'''
Append exponential linear unit activation to custom network
Args:
alpha (float): Multiplicative factor
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "elu";
tmp["params"] = {};
tmp["params"]["alpha"] = alpha;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def leaky_relu(self, alpha=0.01, uid=None):
'''
Append leaky relu activation to custom network
Args:
alpha (float): Multiplicatve factor towards negative spectrum of real numbers.
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "leaky_relu";
tmp["params"] = {};
tmp["params"]["alpha"] = alpha;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def prelu(self, uid=None):
'''
Append paramemeterized rectified linear unit activation to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "prelu";
tmp["params"] = {};
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def selu(self, uid=None):
'''
Append scaled exponential linear unit activation to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "selu";
tmp["params"] = {};
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def hardshrink(self, threshold=0.5, uid=None):
'''
Append hardshrink activation to custom network
Args:
threshold (float): Limit for activation
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "hardshrink";
tmp["params"] = {};
tmp["params"]["threshold"] = threshold;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def hardtanh(self, min_val=-1.0, max_val=1.0, uid=None):
'''
Append Hard-Tanh activation to custom network
Args:
min_val (float): lower thereshold limit value
max_val (float): upper thereshold limit value
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "hardtanh";
tmp["params"] = {};
tmp["params"]["min_val"] = min_val;
tmp["params"]["max_val"] = max_val;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def logsigmoid(self, uid=None):
'''
Append log-sigmoid activation to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "logsigmoid";
tmp["params"] = {};
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def relu6(self, uid=None):
'''
Append relu-6 activation to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "relu6";
tmp["params"] = {};
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def rrelu(self, lower=0.125, upper=0.333, uid=None):
'''
Append randomized rectified linear unit activation to custom network
Args:
lower (float): Lower limit of uniform distribution for the ramdomized multiplicative factor
upper (float): Upper limit of uniform distribution for the ramdomized multiplicative factor
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "rrelu";
tmp["params"] = {};
tmp["params"]["lower"] = lower;
tmp["params"]["upper"] = upper;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def celu(self, alpha=1.0, uid=None):
'''
Append continuously differentiable exponential linear unit activation to custom network
Args:
alpha (float): Multiplicative factor
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "celu";
tmp["params"] = {};
tmp["params"]["alpha"] = alpha;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def softshrink(self, threshold=0.5, uid=None):
'''
Append soft-shrink activation to custom network
Args:
threshold (float): Thresh limit
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "softshrink";
tmp["params"] = {};
tmp["params"]["threshold"] = threshold;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def tanhshrink(self, uid=None):
'''
Append Tanh-Shrink activation to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "tanhshrink";
tmp["params"] = {};
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def threshold(self, threshold=1.0, value=0.01, uid=None):
'''
Append general threshold activation to custom network
Args:
threshold (float): thereshold limit value
value (float): replacement value if input is off-limits
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "threshold";
tmp["params"] = {};
tmp["params"]["threshold"] = threshold;
tmp["params"]["value"] = value;
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def softmin(self, uid=None):
'''
Append softmin activation to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "softmin";
tmp["params"] = {};
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def softmax(self, uid=None):
'''
Append softmax activation to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "softmax";
tmp["params"] = {};
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def logsoftmax(self, uid=None):
'''
Append log-softmax activation to custom network
Args:
uid (str): Unique name for layer, if not mentioned then dynamically assigned
Returns:
dict: Containing all the parameters set as per function arguments
'''
tmp = {};
tmp["uid"] = uid;
tmp["name"] = "logsoftmax";
tmp["params"] = {};
return tmp;
#####################################################################################################################################
#####################################################################################################################################
def resnet_v1_block(self, output_channels=16, stride=1, downsample=True):
'''
Append Resnet V1 Block to custom network
Args:
output_channels (int): Number of output features for this block
stride (int, tuple): kernel movement stride
downsample (bool): If False, residual branch is a shortcut,
Else, residual branch has non-identity layers
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
subnetwork = [];
branch_1 = [];
branch_1.append(self.convolution(output_channels=output_channels, kernel_size=3, stride=stride));
branch_1.append(self.batch_normalization());
branch_1.append(self.relu());
branch_1.append(self.convolution(output_channels=output_channels, kernel_size=3, stride=1));
branch_1.append(self.batch_normalization());
branch_2 = [];
if(downsample):
branch_2.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=stride));
branch_2.append(self.batch_normalization());
else:
branch_2.append(self.identity());
subnetwork.append(branch_1);
subnetwork.append(branch_2);
subnetwork.append(self.add());
network.append(subnetwork)
network.append(self.relu());
return network;
#####################################################################################################################################
#####################################################################################################################################
def resnet_v2_block(self, output_channels=16, stride=1, downsample=True):
'''
Append Resnet V2 Block to custom network
Args:
output_channels (int): Number of output features for this block
stride (int, tuple): kernel movement stride
downsample (bool): If False, residual branch is a shortcut,
Else, residual branch has non-identity layers
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
network.append(self.batch_normalization());
network.append(self.relu());
subnetwork = [];
branch_1 = [];
branch_1.append(self.convolution(output_channels=output_channels, kernel_size=3, stride=stride));
branch_1.append(self.batch_normalization());
branch_1.append(self.relu());
branch_1.append(self.convolution(output_channels=output_channels, kernel_size=3, stride=1));
branch_2 = [];
if(downsample):
branch_2.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=stride));
else:
branch_2.append(self.identity());
subnetwork.append(branch_1);
subnetwork.append(branch_2);
subnetwork.append(self.add());
network.append(subnetwork);
network.append(self.relu());
return network;
#####################################################################################################################################
#####################################################################################################################################
def resnet_v1_bottleneck_block(self, output_channels=16, stride=1, downsample=True):
'''
Append Resnet V1 Bottleneck Block to custom network
Args:
output_channels (int): Number of output features for this block
stride (int, tuple): kernel movement stride
downsample (bool): If False, residual branch is a shortcut,
Else, residual branch has non-identity layers
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
subnetwork = [];
branch_1 = [];
branch_1.append(self.convolution(output_channels=output_channels//4, kernel_size=1, stride=stride));
branch_1.append(self.batch_normalization());
branch_1.append(self.relu());
branch_1.append(self.convolution(output_channels=output_channels//4, kernel_size=3, stride=1));
branch_1.append(self.batch_normalization());
branch_1.append(self.relu());
branch_1.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1));
branch_1.append(self.batch_normalization());
branch_2 = [];
if(downsample):
branch_2.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=stride));
branch_2.append(self.batch_normalization());
else:
branch_2.append(self.identity());
subnetwork.append(branch_1);
subnetwork.append(branch_2);
subnetwork.append(self.add());
network.append(subnetwork)
network.append(self.relu())
return network;
#####################################################################################################################################
#####################################################################################################################################
def resnet_v2_bottleneck_block(self, output_channels=16, stride=1, downsample=True):
'''
Append Resnet V2 Bottleneck Block to custom network
Args:
output_channels (int): Number of output features for this block
stride (int, tuple): kernel movement stride
downsample (bool): If False, residual branch is a shortcut,
Else, residual branch has non-identity layers
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
network.append(self.batch_normalization());
network.append(self.relu());
subnetwork = [];
branch_1 = [];
branch_1.append(self.convolution(output_channels=output_channels//4, kernel_size=1, stride=1));
branch_1.append(self.batch_normalization());
branch_1.append(self.relu());
branch_1.append(self.convolution(output_channels=output_channels//4, kernel_size=3, stride=stride));
branch_1.append(self.batch_normalization());
branch_1.append(self.relu());
branch_1.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1));
branch_2 = [];
if(downsample):
branch_2.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=stride));
else:
branch_2.append(self.identity());
subnetwork.append(branch_1);
subnetwork.append(branch_2);
subnetwork.append(self.add());
network.append(subnetwork)
return network;
#####################################################################################################################################
#####################################################################################################################################
def resnext_block(self, output_channels=256, cardinality=8, bottleneck_width=4, stride=1, downsample=True):
'''
Append Resnext Block to custom network
Args:
output_channels (int): Number of output features for this block
cardinality (int): cardinality dimensions for complex transformations
bottleneck_width (int): Bottleneck dimensions for reducing number of features
stride (int): kernel movement stride
downsample (bool): If False, residual branch is a shortcut,
Else, residual branch has non-identity layers
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
channels = output_channels//4;
D = int(math.floor(channels * (bottleneck_width / 64)))
group_width = cardinality * D
subnetwork = [];
branch_1 = [];
branch_1.append(self.convolution(output_channels=group_width, kernel_size=1, stride=1));
branch_1.append(self.batch_normalization());
branch_1.append(self.relu());
branch_1.append(self.convolution(output_channels=group_width, kernel_size=3, stride=stride));
branch_1.append(self.batch_normalization());
branch_1.append(self.relu());
branch_1.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1));
branch_1.append(self.batch_normalization());
branch_2 = [];
if(downsample):
branch_2.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=stride));
branch_2.append(self.batch_normalization());
else:
branch_2.append(self.identity());
subnetwork.append(branch_1);
subnetwork.append(branch_2);
subnetwork.append(self.add());
network.append(subnetwork)
network.append(self.relu());
return network;
#####################################################################################################################################
#####################################################################################################################################
def mobilenet_v2_linear_bottleneck_block(self, output_channels=32, bottleneck_width=4, stride=1):
'''
Append Mobilenet V2 Linear Bottleneck Block to custom network
Args:
output_channels (int): Number of output features for this block
stride (int): kernel movement stride
bottleneck_width (int): Bottleneck dimensions for reducing number of features
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
subnetwork = [];
branch_1 = [];
if(bottleneck_width != 1):
branch_1.append(self.convolution(output_channels=output_channels*bottleneck_width,
kernel_size=1, stride=1));
branch_1.append(self.batch_normalization());
branch_1.append(self.relu());
branch_1.append(self.convolution(output_channels=output_channels*bottleneck_width,
kernel_size=3, stride=stride));
branch_1.append(self.batch_normalization());
branch_1.append(self.relu());
branch_1.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1));
branch_1.append(self.batch_normalization());
branch_2 = [];
branch_2.append(self.identity());
subnetwork.append(branch_1);
subnetwork.append(branch_2);
subnetwork.append(self.add());
network.append(subnetwork);
return network;
#####################################################################################################################################
#####################################################################################################################################
def separable_convolution_block(self, input_channels=16, output_channels=32, kernel_size=3, stride=1, padding=1):
'''
Append Separable convolution Block to custom network
Args:
input_channels (int): Number of input features for this block
output_channels (int): Number of output features for this block
kernel_size (int): Kernel matrix shape for all layers in this block
stride (int): kernel movement stride
padding (int, tuple): external zero padding on input
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
network.append(self.convolution(output_channels=input_channels, kernel_size=kernel_size,
stride=stride, padding=padding, groups=input_channels));
network.append(self.convolution(output_channels=output_channels, kernel_size=1,
stride=1));
return network;
#####################################################################################################################################
#####################################################################################################################################
def mobilenet_v2_inverted_linear_bottleneck_block(self, output_channels=32, bottleneck_width=4, stride=1):
'''
Append Mobilenet V2 Inverted Linear Bottleneck Block to custom network
Args:
output_channels (int): Number of output features for this block
stride (int): kernel movement stride
bottleneck_width (int): Bottleneck dimensions for reducing number of features
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
subnetwork = [];
branch_1 = [];
if(bottleneck_width != 1):
branch_1.append(self.convolution(output_channels=output_channels//bottleneck_width,
kernel_size=1, stride=1));
branch_1.append(self.batch_normalization());
branch_1.append(self.relu());
sep_conv = self.separable_convolution_block(input_channels=output_channels//bottleneck_width,
output_channels=output_channels//bottleneck_width,
kernel_size=3, stride=stride);
branch_1.append(sep_conv);
branch_1.append(self.batch_normalization());
branch_1.append(self.relu());
branch_1.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1));
branch_1.append(self.batch_normalization());
branch_2 = [];
branch_2.append(self.identity());
subnetwork.append(branch_1);
subnetwork.append(branch_2);
subnetwork.append(self.add());
network.append(subnetwork);
return network;
#####################################################################################################################################
#####################################################################################################################################
def squeezenet_fire_block(self, squeeze_channels=16, expand_channels_1x1=32, expand_channels_3x3=64):
'''
Append Squeezenet Fire Block to custom network
Args:
squeeze_channels (int): Number of output features for this block
expand_channels_1x1 (int): Number of convolution_1x1 features for this block
expand_channels_3x3 (int): Number of convolution_3x3 features for this block
bottleneck_width (int): Bottleneck dimensions for reducing number of features
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
network.append(self.convolution(output_channels=squeeze_channels, kernel_size=1, stride=1));
network.append(self.relu());
subnetwork = [];
branch_1 = [];
branch_2 = [];
branch_1.append(self.convolution(output_channels=expand_channels_1x1, kernel_size=1, stride=1));
branch_1.append(self.relu());
branch_2.append(self.convolution(output_channels=expand_channels_3x3, kernel_size=3, stride=1));
branch_2.append(self.relu());
subnetwork.append(branch_1);
subnetwork.append(branch_2);
subnetwork.append(self.concatenate());
network.append(subnetwork);
return network;
#####################################################################################################################################
#####################################################################################################################################
def densenet_block(self, bottleneck_size=4, growth_rate=16, dropout=0.2):
'''
Append Densenet Block to custom network
Args:
bottleneck_size (int): Bottleneck dimensions for reducing number of features
growth_rate (int): Expansion rate for convolution layers for this block
dropout (float): Prbability for dropout layer post convolution
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
subnetwork = [];
branch_1 = [];
branch_2 = [];
branch_1.append(self.batch_normalization());
branch_1.append(self.relu());
branch_1.append(self.convolution(output_channels=bottleneck_size*growth_rate, kernel_size=1, stride=1));
branch_1.append(self.batch_normalization());
branch_1.append(self.relu());
branch_1.append(self.convolution(output_channels=growth_rate, kernel_size=3, stride=1));
branch_1.append(self.dropout(drop_probability=dropout));
branch_2.append(self.identity());
subnetwork.append(branch_1);
subnetwork.append(branch_2);
subnetwork.append(self.concatenate());
network.append(subnetwork);
return network;
#####################################################################################################################################
#####################################################################################################################################
def conv_bn_relu_block(self, output_channels=64, kernel_size=1, stride=1, padding=None):
'''
Append Conv->batch_norm->relu Block to custom network
Args:
output_channels (int): Number of output features for this block
kernel_size (int): Kernel matrix shape for all layers in this block
stride (int): kernel movement stride
padding (int, tuple): external zero padding on input
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
if(padding):
network.append(self.convolution(output_channels=output_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding));
else:
network.append(self.convolution(output_channels=output_channels,
kernel_size=kernel_size,
stride=stride));
network.append(self.batch_normalization());
network.append(self.relu());
return network;
#####################################################################################################################################
#####################################################################################################################################
def inception_a_block(self, pooling_branch_channels=32, pool_type="avg"):
'''
Append Inception-A Block to custom network
Args:
pooling_branch_channels (int): Number of features for conv layers in pooling branch
pool_type (str): Either of these types
- "avg" - Average pooling
- "max" - Max pooling
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
subnetwork = [];
branch_1 = [];
branch_2 = [];
branch_3 = [];
branch_4 = [];
branch_1.append(self.conv_bn_relu_block(output_channels=64, kernel_size=1))
branch_2.append(self.conv_bn_relu_block(output_channels=48, kernel_size=1));
branch_2.append(self.conv_bn_relu_block(output_channels=64, kernel_size=5));
branch_3.append(self.conv_bn_relu_block(output_channels=64, kernel_size=1));
branch_3.append(self.conv_bn_relu_block(output_channels=96, kernel_size=3));
branch_3.append(self.conv_bn_relu_block(output_channels=96, kernel_size=3));
if(pool_type=="avg"):
branch_4.append(self.average_pooling(kernel_size=3, stride=1, padding=1));
else:
branch_4.append(self.max_pooling(kernel_size=3, stride=1, padding=1));
branch_4.append(self.conv_bn_relu_block(output_channels=pooling_branch_channels, kernel_size=1));
subnetwork.append(branch_1);
subnetwork.append(branch_2);
subnetwork.append(branch_3);
subnetwork.append(branch_4);
subnetwork.append(self.concatenate());
network.append(subnetwork);
return network;
#####################################################################################################################################
#####################################################################################################################################
def inception_b_block(self, pool_type="avg"):
'''
Append Inception-B Block to custom network
Args:
pool_type (str): Either of these types
- "avg" - Average pooling
- "max" - Max pooling
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
subnetwork = [];
branch_1 = [];
branch_2 = [];
branch_3 = [];
branch_4 = [];
branch_1.append(self.conv_bn_relu_block(output_channels=384, kernel_size=3))
branch_2.append(self.conv_bn_relu_block(output_channels=64, kernel_size=1));
branch_2.append(self.conv_bn_relu_block(output_channels=96, kernel_size=3));
branch_2.append(self.conv_bn_relu_block(output_channels=96, kernel_size=3));
if(pool_type=="avg"):
branch_3.append(self.average_pooling(kernel_size=3, stride=1, padding=1));
else:
branch_3.append(self.max_pooling(kernel_size=3, stride=1, padding=1));
subnetwork.append(branch_1);
subnetwork.append(branch_2);
subnetwork.append(branch_3);
subnetwork.append(self.concatenate());
network.append(subnetwork);
return network;
#####################################################################################################################################
#####################################################################################################################################
def inception_c_block(self, channels_7x7=3, pool_type="avg"):
'''
Append Inception-C Block to custom network
Args:
channels_7x7 (int): Number of features for conv layers in channels_7x7 branch
pool_type (str): Either of these types
- "avg" - Average pooling
- "max" - Max pooling
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
subnetwork = [];
branch_1 = [];
branch_2 = [];
branch_3 = [];
branch_4 = [];
branch_1.append(self.conv_bn_relu_block(output_channels=192, kernel_size=1))
branch_2.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=1));
branch_2.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=(1, 7), padding=(0, 3)));
branch_2.append(self.conv_bn_relu_block(output_channels=192, kernel_size=(7, 1), padding=(3, 0)));
branch_3.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=1));
branch_3.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=(1, 7), padding=(0, 3)));
branch_3.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=(7, 1), padding=(3, 0)));
branch_3.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=(1, 7), padding=(0, 3)));
branch_3.append(self.conv_bn_relu_block(output_channels=192, kernel_size=(7, 1), padding=(3, 0)));
if(pool_type=="avg"):
branch_4.append(self.average_pooling(kernel_size=3, stride=1, padding=1));
else:
branch_4.append(self.max_pooling(kernel_size=3, stride=1, padding=1));
branch_4.append(self.conv_bn_relu_block(output_channels=192, kernel_size=1));
subnetwork.append(branch_1);
subnetwork.append(branch_2);
subnetwork.append(branch_3);
subnetwork.append(branch_4);
subnetwork.append(self.concatenate());
network.append(subnetwork);
return network;
#####################################################################################################################################
#####################################################################################################################################
def inception_d_block(self, pool_type="avg"):
'''
Append Inception-D Block to custom network
Args:
pool_type (str): Either of these types
- "avg" - Average pooling
- "max" - Max pooling
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
subnetwork = [];
branch_1 = [];
branch_2 = [];
branch_3 = [];
branch_4 = [];
branch_1.append(self.conv_bn_relu_block(output_channels=192, kernel_size=1))
branch_1.append(self.conv_bn_relu_block(output_channels=320, kernel_size=3, stride=2))
branch_2.append(self.conv_bn_relu_block(output_channels=192, kernel_size=1));
branch_2.append(self.conv_bn_relu_block(output_channels=192, kernel_size=(1, 7), padding=(0, 3)));
branch_2.append(self.conv_bn_relu_block(output_channels=192, kernel_size=(7, 1), padding=(3, 0)));
branch_2.append(self.conv_bn_relu_block(output_channels=192, kernel_size=3, stride=2));
if(pool_type=="avg"):
branch_3.append(self.average_pooling(kernel_size=3, stride=2));
else:
branch_3.append(self.max_pooling(kernel_size=3, stride=2));
subnetwork.append(branch_1);
subnetwork.append(branch_2);
subnetwork.append(branch_3);
subnetwork.append(self.concatenate());
network.append(subnetwork);
return network;
#####################################################################################################################################
#####################################################################################################################################
def subbranch_block(self):
'''
Append sub-branch Block to custom network
Args:
None
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
subnetwork = [];
branch_1 = [];
branch_2 = [];
branch_1.append(self.conv_bn_relu_block(output_channels=384, kernel_size=(1, 3), padding=(0, 1)));
branch_2.append(self.conv_bn_relu_block(output_channels=384, kernel_size=(3, 1), padding=(1, 0)));
subnetwork.append(branch_1);
subnetwork.append(branch_2);
subnetwork.append(self.concatenate());
return subnetwork;
#####################################################################################################################################
#####################################################################################################################################
def inception_e_block(self, pool_type="avg"):
'''
Append Inception-C Block to custom network
Args:
pool_type (str): Either of these types
- "avg" - Average pooling
- "max" - Max pooling
Returns:
list: Containing all the layer dictionaries arranged as per function arguments
'''
network = [];
subnetwork = [];
branch_1 = [];
branch_2 = [];
branch_3 = [];
branch_4 = [];
branch_1.append(self.conv_bn_relu_block(output_channels=320, kernel_size=1))
branch_2.append(self.conv_bn_relu_block(output_channels=384, kernel_size=1));
branch_2.append(self.subbranch_block());
branch_3.append(self.conv_bn_relu_block(output_channels=448, kernel_size=1));
branch_3.append(self.conv_bn_relu_block(output_channels=384, kernel_size=3, padding=1));
branch_3.append(self.subbranch_block());
if(pool_type=="avg"):
branch_4.append(self.average_pooling(kernel_size=3, stride=1, padding=1));
else:
branch_4.append(self.max_pooling(kernel_size=3, stride=1, padding=1));
branch_4.append(self.conv_bn_relu_block(output_channels=192, kernel_size=1));
subnetwork.append(branch_1);
subnetwork.append(branch_2);
subnetwork.append(branch_3);
subnetwork.append(branch_4);
subnetwork.append(self.concatenate());
network.append(subnetwork);
return network;
#####################################################################################################################################Classes
class prototype_layers (...)-
Main class for all auxiliary functions - EDA, Estimate Training Time, Resetting params, switching modes, & debugging
Args
verbose:int- Set verbosity levels 0 - Print Nothing 1 - Print desired details
Class that serves as a decorator to trace entry and exit from functions. Used by appending @TraceFunction on top of the definition of the function to trace.
Expand source code
class prototype_layers(prototype_aux): @accepts("self", verbose=int, post_trace=True) @TraceFunction(trace_args=True, trace_rv=True) def __init__(self, verbose=1): super().__init__(verbose=verbose); ##################################################################################################################################### def append_linear(self, num_neurons=False, final_layer=False): ''' Append dense (fully connected) layer to base network in transfer learning Args: num_neurons (int): Number of neurons in the dense layer final_layer (bool): If True, then number of neurons are directly set as number of classes in dataset for single label type classification Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: if(not num_neurons): num_neurons = self.system_dict["dataset"]["params"]["num_classes"]; self.system_dict = layer_linear(self.system_dict, num_neurons=num_neurons, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_dropout(self, probability=0.5, final_layer=False): ''' Append dropout layer to base network in transfer learning Args: probability (float): Droping probability of neurons in next layer final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = layer_dropout(self.system_dict, probability=probability, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_elu(self, alpha=1.0, final_layer=False): ''' Append exponential linear unit activation to base network in transfer learning Args: alpha (float): Multiplicatve factor. final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_elu(self.system_dict, alpha=alpha, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_hardshrink(self, lambd=0.5, final_layer=False): ''' Append hardshrink activation to base network in transfer learning Args: lambd (float): thereshold limit value final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_hardshrink(self.system_dict, lambd=lambd, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_hardtanh(self, min_val=-1.0, max_val=1.0, final_layer=False): ''' Append hardtanh activation to base network in transfer learning Args: min_val (float): lower thereshold limit value max_val (float): upper thereshold limit value final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_hardtanh(self.system_dict, min_val=min_val, max_val=max_val, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_leakyrelu(self, negative_slope=0.01, final_layer=False): ''' Append Leaky - ReLU activation to base network in transfer learning Args: negative_slope (float): Multiplicatve factor towards negative spectrum of real numbers. final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_leakyrelu(self.system_dict, negative_slope=negative_slope, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_logsigmoid(self, final_layer=False): ''' Append logsigmoid activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_logsigmoid(self.system_dict, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_prelu(self, num_parameters=1, init=0.25, final_layer=False): ''' Append Learnable parameerized rectified linear unit activation to base network in transfer learning Args: init (float): Initialization value for multiplicatve factor towards negative spectrum of real numbers. final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_prelu(self.system_dict, num_parameters=num_parameters, init=init, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_relu(self, final_layer=False): ''' Append rectified linear unit activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_relu(self.system_dict, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_relu6(self, final_layer=False): ''' Append rectified linear unit activation - 6 to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_relu6(self.system_dict, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_rrelu(self, lower=0.125, upper=0.333, final_layer=False): ''' Append randomized rectified linear unit activation to base network in transfer learning Args: lower (float): Lower limit of uniform distribution for the ramdomized multiplicative factor upper (float): Upper limit of uniform distribution for the ramdomized multiplicative factor final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_rrelu(self.system_dict, lower=lower, upper=upper, final_layer=final_layer) ##################################################################################################################################### ##################################################################################################################################### def append_selu(self, final_layer=False): ''' Append scaled exponential linear unit activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_selu(self.system_dict, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_celu(self, alpha=1.0, final_layer=False): ''' Append continuously differentiable exponential linear unit activation to base network in transfer learning Args: alpha (float): Multiplicative factor final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_celu(self.system_dict, alpha=alpha, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_sigmoid(self, final_layer=False): ''' Append rectified linear unit activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_sigmoid(self.system_dict, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_softplus(self, beta=1, threshold=20, final_layer=False): ''' Append softplus activation to base network in transfer learning Args: beta (float): Multiplicative factor threshold (int): softplus (thresholded relu) limit final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_softplus(self.system_dict, beta=beta, threshold=threshold, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_softshrink(self, lambd=0.5, final_layer=False): ''' Append softshrink activation to base network in transfer learning Args: lambd (float): thereshold limit value final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_softshrink(self.system_dict, lambd=lambd, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_softsign(self, final_layer=False): ''' Append tanh activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_softsign(self.system_dict, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_tanh(self, final_layer=False): ''' Append tanh activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_tanh(self.system_dict, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_tanhshrink(self, final_layer=False): ''' Append tanhshrink activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_tanhshrink(self.system_dict, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_threshold(self, threshold, value, final_layer=False): ''' Append threshold activation to base network in transfer learning Args: threshold (float): thereshold limit value value (float): replacement value if input is off-limits final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_threshold(self.system_dict, threshold, value, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_softmin(self, final_layer=False): ''' Append softmin activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_softmin(self.system_dict, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_softmax(self, final_layer=False): ''' Append softmax activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_softmax(self.system_dict, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def append_logsoftmax(self, final_layer=False): ''' Append log-softmax activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_logsoftmax(self.system_dict, final_layer=final_layer); ##################################################################################################################################### ##################################################################################################################################### def convolution1d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", groups=1, dilation=1, use_bias=True, layout='NCW', uid=None): ''' Append 1d-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "convolution1d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def convolution2d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None): ''' Append 2d-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "convolution2d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def convolution(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None): ''' Append 2d-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "convolution2d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def convolution3d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", groups=1, dilation=1, use_bias=True, layout='NCDHW', uid=None): ''' Append 3d-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "convolution3d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def transposed_convolution1d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCW', uid=None): ''' Append 1d-transposed-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding output_padding (int): Additional padding applied to output groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "transposed_convolution1d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["output_padding"] = output_padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def transposed_convolution(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None): ''' Append 2d-transposed-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding output_padding (int): Additional padding applied to output groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "transposed_convolution2d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["output_padding"] = output_padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def transposed_convolution2d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None): ''' Append 2d-transposed-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding output_padding (int): Additional padding applied to output groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "transposed_convolution2d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["output_padding"] = output_padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def transposed_convolution3d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCDHW', uid=None): ''' Append 3d-transposed-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding output_padding (int): Additional padding applied to output groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "transposed_convolution3d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["output_padding"] = output_padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def max_pooling1d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, layout='NCW', uid=None): ''' Append 1d-max-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False ceil_mode (bool): If True, apply ceil math operation post pooling layout (str): Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "max_pooling1d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def max_pooling2d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, layout='NCHW', uid=None): ''' Append 2d-max-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False ceil_mode (bool): If True, apply ceil math operation post pooling layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "max_pooling2d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def max_pooling(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, layout='NCHW', uid=None): ''' Append 2d-max-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False ceil_mode (bool): If True, apply ceil math operation post pooling layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "max_pooling2d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def max_pooling3d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, layout='NCDHW', uid=None): ''' Append 3d-max-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False ceil_mode (bool): If True, apply ceil math operation post pooling layout (str): Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "max_pooling3d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def average_pooling1d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, include_padding_in_calculation=True, layout='NCW', uid=None): ''' Append 1d-average-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False include_padding_in_calculation (bool): If True, padding will be considered. ceil_mode (bool): If True, apply ceil math operation post pooling layout (str): Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "average_pooling1d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["include_padding_in_calculation"] = include_padding_in_calculation; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def average_pooling2d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, include_padding_in_calculation=True, layout='NCHW', uid=None): ''' Append 2d-average-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False ceil_mode (bool): If True, apply ceil math operation post pooling include_padding_in_calculation (bool): If True, padding will be considered. layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "average_pooling2d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["include_padding_in_calculation"] = include_padding_in_calculation; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def average_pooling(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, include_padding_in_calculation=True, layout='NCHW', uid=None): ''' Append 2d-average-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False ceil_mode (bool): If True, apply ceil math operation post pooling include_padding_in_calculation (bool): If True, padding will be considered. layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "average_pooling2d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["include_padding_in_calculation"] = include_padding_in_calculation; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def average_pooling3d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, include_padding_in_calculation=True, layout='NCDHW', uid=None): ''' Append 3d-average-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False ceil_mode (bool): If True, apply ceil math operation post pooling include_padding_in_calculation (bool): If True, padding will be considered. layout (str): Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "average_pooling3d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["include_padding_in_calculation"] = include_padding_in_calculation; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def global_max_pooling1d(self, layout='NCW', uid=None): ''' Append 1d-global-max-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_max_pooling1d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def global_max_pooling2d(self, layout='NCHW', uid=None): ''' Append 2d-global-max-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_max_pooling2d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def global_max_pooling(self, layout='NCHW', uid=None): ''' Append 2d-global-max-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_max_pooling2d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def global_max_pooling3d(self, layout='NCDHW', uid=None): ''' Append 3d-global-max-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_max_pooling3d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def global_average_pooling1d(self, layout='NCW', uid=None): ''' Append 1d-global-average-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_average_pooling1d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def global_average_pooling2d(self, layout='NCHW', uid=None): ''' Append 2d-global-average-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_average_pooling2d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def global_average_pooling(self, layout='NCHW', uid=None): ''' Append 2d-global-average-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_average_pooling2d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def global_average_pooling3d(self, layout='NCDHW', uid=None): ''' Append 3d-global-average-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_average_pooling3d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; ##################################################################################################################################### ##################################################################################################################################### def fully_connected(self, units=512, use_bias=True, flatten=True, uid=None): ''' Append fully-connected (dense) layer to custom network Args: units (int): Number of neurons in the layer use_bias (bool): If True, learnable bias is added flatten (bool): Fixed to True uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "fully_connected"; tmp["params"] = {}; tmp["params"]["units"] = units; tmp["params"]["use_bias"] = use_bias; tmp["params"]["flatten"] = flatten; return tmp; ##################################################################################################################################### ##################################################################################################################################### def dropout(self, drop_probability=0.2, axes=(), uid=None): ''' Append dropout layer to custom network Args: drop_probability (float): Probability for not considering neurons in the output axes (tuple): Channel axis to implement dropout over uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "dropout"; tmp["params"] = {}; tmp["params"]["drop_probability"] = drop_probability; tmp["params"]["axes"] = axes; return tmp; ##################################################################################################################################### ##################################################################################################################################### def flatten(self, uid=None): ''' Append flatten layer to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "flatten"; tmp["params"] = {}; return tmp; ##################################################################################################################################### ##################################################################################################################################### def identity(self, uid=None): ''' Append identity layer to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "identity"; tmp["params"] = {}; return tmp; ##################################################################################################################################### ##################################################################################################################################### def batch_normalization(self, moving_average_momentum=0.1, epsilon=0.00001, use_trainable_parameters=True, activate_scale_shift_operation=False, uid=None): ''' Append batch normalization layer to custom network Args: moving_average_momentum (float): Normalization momentum value epsilon (float): Value to avoid division by zero use_trainable_paramemetrs (bool): If True, batch norm turns into a trainable layer activate_scale_shift_operation (bool): Fixed status - False uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "batch_normalization"; tmp["params"] = {}; tmp["params"]["moving_average_momentum"] = moving_average_momentum; tmp["params"]["epsilon"] = epsilon; tmp["params"]["use_trainable_parameters"] = use_trainable_parameters; tmp["params"]["activate_scale_shift_operation"] = activate_scale_shift_operation; return tmp; ##################################################################################################################################### ##################################################################################################################################### def instance_normalization(self, moving_average_momentum=0.1, epsilon=0.00001, use_trainable_parameters=False, uid=None): ''' Append instace normalization layer to custom network Args: moving_average_momentum (float): Normalization momentum value epsilon (float): Value to avoid division by zero use_trainable_paramemetrs (bool): If True, batch norm turns into a trainable layer uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "instance_normalization"; tmp["params"] = {}; tmp["params"]["moving_average_momentum"] = moving_average_momentum; tmp["params"]["epsilon"] = epsilon; tmp["params"]["use_trainable_parameters"] = use_trainable_parameters; return tmp; ##################################################################################################################################### ##################################################################################################################################### def layer_normalization(self, moving_average_momentum=0.9, epsilon=0.00001, use_trainable_parameters=True, uid=None): ''' Append layer normalization layer to custom network Args: moving_average_momentum (float): Normalization momentum value epsilon (float): Value to avoid division by zero use_trainable_paramemetrs (bool): If True, batch norm turns into a trainable layer uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "layer_normalization"; tmp["params"] = {}; tmp["params"]["epsilon"] = epsilon; tmp["params"]["use_trainable_parameters"] = use_trainable_parameters; return tmp; ##################################################################################################################################### ##################################################################################################################################### def add(self, uid=None): ''' Append elementwise addition layer to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "add"; return tmp; ##################################################################################################################################### ##################################################################################################################################### def concatenate(self, uid=None): ''' Append concatenation layer to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "concatenate"; return tmp; ##################################################################################################################################### ##################################################################################################################################### def relu(self, uid=None): ''' Append rectified linear unit activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "relu"; tmp["params"] = {}; return tmp; ##################################################################################################################################### ##################################################################################################################################### def sigmoid(self, uid=None): ''' Append sigmoid activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "sigmoid"; tmp["params"] = {}; return tmp; ##################################################################################################################################### ##################################################################################################################################### def tanh(self, uid=None): ''' Append tanh activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "tanh"; tmp["params"] = {}; return tmp; ##################################################################################################################################### ##################################################################################################################################### def softplus(self, beta=1, threshold=20, uid=None): ''' Append softplus activation to custom network Args: beta (float): Multiplicative factor threshold (int): softplus (thresholded relu) limit uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "softplus"; tmp["params"] = {}; tmp["params"]["beta"] = beta; tmp["params"]["threshold"] = threshold; return tmp; ##################################################################################################################################### ##################################################################################################################################### def softsign(self, uid=None): ''' Append softsign activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "softsign"; tmp["params"] = {}; return tmp; ##################################################################################################################################### ##################################################################################################################################### def elu(self, alpha=1.0, uid=None): ''' Append exponential linear unit activation to custom network Args: alpha (float): Multiplicative factor uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "elu"; tmp["params"] = {}; tmp["params"]["alpha"] = alpha; return tmp; ##################################################################################################################################### ##################################################################################################################################### def leaky_relu(self, alpha=0.01, uid=None): ''' Append leaky relu activation to custom network Args: alpha (float): Multiplicatve factor towards negative spectrum of real numbers. uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "leaky_relu"; tmp["params"] = {}; tmp["params"]["alpha"] = alpha; return tmp; ##################################################################################################################################### ##################################################################################################################################### def prelu(self, uid=None): ''' Append paramemeterized rectified linear unit activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "prelu"; tmp["params"] = {}; return tmp; ##################################################################################################################################### ##################################################################################################################################### def selu(self, uid=None): ''' Append scaled exponential linear unit activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "selu"; tmp["params"] = {}; return tmp; ##################################################################################################################################### ##################################################################################################################################### def hardshrink(self, threshold=0.5, uid=None): ''' Append hardshrink activation to custom network Args: threshold (float): Limit for activation uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "hardshrink"; tmp["params"] = {}; tmp["params"]["threshold"] = threshold; return tmp; ##################################################################################################################################### ##################################################################################################################################### def hardtanh(self, min_val=-1.0, max_val=1.0, uid=None): ''' Append Hard-Tanh activation to custom network Args: min_val (float): lower thereshold limit value max_val (float): upper thereshold limit value uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "hardtanh"; tmp["params"] = {}; tmp["params"]["min_val"] = min_val; tmp["params"]["max_val"] = max_val; return tmp; ##################################################################################################################################### ##################################################################################################################################### def logsigmoid(self, uid=None): ''' Append log-sigmoid activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "logsigmoid"; tmp["params"] = {}; return tmp; ##################################################################################################################################### ##################################################################################################################################### def relu6(self, uid=None): ''' Append relu-6 activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "relu6"; tmp["params"] = {}; return tmp; ##################################################################################################################################### ##################################################################################################################################### def rrelu(self, lower=0.125, upper=0.333, uid=None): ''' Append randomized rectified linear unit activation to custom network Args: lower (float): Lower limit of uniform distribution for the ramdomized multiplicative factor upper (float): Upper limit of uniform distribution for the ramdomized multiplicative factor uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "rrelu"; tmp["params"] = {}; tmp["params"]["lower"] = lower; tmp["params"]["upper"] = upper; return tmp; ##################################################################################################################################### ##################################################################################################################################### def celu(self, alpha=1.0, uid=None): ''' Append continuously differentiable exponential linear unit activation to custom network Args: alpha (float): Multiplicative factor uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "celu"; tmp["params"] = {}; tmp["params"]["alpha"] = alpha; return tmp; ##################################################################################################################################### ##################################################################################################################################### def softshrink(self, threshold=0.5, uid=None): ''' Append soft-shrink activation to custom network Args: threshold (float): Thresh limit uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "softshrink"; tmp["params"] = {}; tmp["params"]["threshold"] = threshold; return tmp; ##################################################################################################################################### ##################################################################################################################################### def tanhshrink(self, uid=None): ''' Append Tanh-Shrink activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "tanhshrink"; tmp["params"] = {}; return tmp; ##################################################################################################################################### ##################################################################################################################################### def threshold(self, threshold=1.0, value=0.01, uid=None): ''' Append general threshold activation to custom network Args: threshold (float): thereshold limit value value (float): replacement value if input is off-limits uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "threshold"; tmp["params"] = {}; tmp["params"]["threshold"] = threshold; tmp["params"]["value"] = value; return tmp; ##################################################################################################################################### ##################################################################################################################################### def softmin(self, uid=None): ''' Append softmin activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "softmin"; tmp["params"] = {}; return tmp; ##################################################################################################################################### ##################################################################################################################################### def softmax(self, uid=None): ''' Append softmax activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "softmax"; tmp["params"] = {}; return tmp; ##################################################################################################################################### ##################################################################################################################################### def logsoftmax(self, uid=None): ''' Append log-softmax activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "logsoftmax"; tmp["params"] = {}; return tmp; ##################################################################################################################################### ##################################################################################################################################### def resnet_v1_block(self, output_channels=16, stride=1, downsample=True): ''' Append Resnet V1 Block to custom network Args: output_channels (int): Number of output features for this block stride (int, tuple): kernel movement stride downsample (bool): If False, residual branch is a shortcut, Else, residual branch has non-identity layers Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_1.append(self.convolution(output_channels=output_channels, kernel_size=3, stride=stride)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels, kernel_size=3, stride=1)); branch_1.append(self.batch_normalization()); branch_2 = []; if(downsample): branch_2.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=stride)); branch_2.append(self.batch_normalization()); else: branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.add()); network.append(subnetwork) network.append(self.relu()); return network; ##################################################################################################################################### ##################################################################################################################################### def resnet_v2_block(self, output_channels=16, stride=1, downsample=True): ''' Append Resnet V2 Block to custom network Args: output_channels (int): Number of output features for this block stride (int, tuple): kernel movement stride downsample (bool): If False, residual branch is a shortcut, Else, residual branch has non-identity layers Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; network.append(self.batch_normalization()); network.append(self.relu()); subnetwork = []; branch_1 = []; branch_1.append(self.convolution(output_channels=output_channels, kernel_size=3, stride=stride)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels, kernel_size=3, stride=1)); branch_2 = []; if(downsample): branch_2.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=stride)); else: branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.add()); network.append(subnetwork); network.append(self.relu()); return network; ##################################################################################################################################### ##################################################################################################################################### def resnet_v1_bottleneck_block(self, output_channels=16, stride=1, downsample=True): ''' Append Resnet V1 Bottleneck Block to custom network Args: output_channels (int): Number of output features for this block stride (int, tuple): kernel movement stride downsample (bool): If False, residual branch is a shortcut, Else, residual branch has non-identity layers Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_1.append(self.convolution(output_channels=output_channels//4, kernel_size=1, stride=stride)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels//4, kernel_size=3, stride=1)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_2 = []; if(downsample): branch_2.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=stride)); branch_2.append(self.batch_normalization()); else: branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.add()); network.append(subnetwork) network.append(self.relu()) return network; ##################################################################################################################################### ##################################################################################################################################### def resnet_v2_bottleneck_block(self, output_channels=16, stride=1, downsample=True): ''' Append Resnet V2 Bottleneck Block to custom network Args: output_channels (int): Number of output features for this block stride (int, tuple): kernel movement stride downsample (bool): If False, residual branch is a shortcut, Else, residual branch has non-identity layers Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; network.append(self.batch_normalization()); network.append(self.relu()); subnetwork = []; branch_1 = []; branch_1.append(self.convolution(output_channels=output_channels//4, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels//4, kernel_size=3, stride=stride)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1)); branch_2 = []; if(downsample): branch_2.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=stride)); else: branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.add()); network.append(subnetwork) return network; ##################################################################################################################################### ##################################################################################################################################### def resnext_block(self, output_channels=256, cardinality=8, bottleneck_width=4, stride=1, downsample=True): ''' Append Resnext Block to custom network Args: output_channels (int): Number of output features for this block cardinality (int): cardinality dimensions for complex transformations bottleneck_width (int): Bottleneck dimensions for reducing number of features stride (int): kernel movement stride downsample (bool): If False, residual branch is a shortcut, Else, residual branch has non-identity layers Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; channels = output_channels//4; D = int(math.floor(channels * (bottleneck_width / 64))) group_width = cardinality * D subnetwork = []; branch_1 = []; branch_1.append(self.convolution(output_channels=group_width, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=group_width, kernel_size=3, stride=stride)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_2 = []; if(downsample): branch_2.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=stride)); branch_2.append(self.batch_normalization()); else: branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.add()); network.append(subnetwork) network.append(self.relu()); return network; ##################################################################################################################################### ##################################################################################################################################### def mobilenet_v2_linear_bottleneck_block(self, output_channels=32, bottleneck_width=4, stride=1): ''' Append Mobilenet V2 Linear Bottleneck Block to custom network Args: output_channels (int): Number of output features for this block stride (int): kernel movement stride bottleneck_width (int): Bottleneck dimensions for reducing number of features Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; if(bottleneck_width != 1): branch_1.append(self.convolution(output_channels=output_channels*bottleneck_width, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels*bottleneck_width, kernel_size=3, stride=stride)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_2 = []; branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.add()); network.append(subnetwork); return network; ##################################################################################################################################### ##################################################################################################################################### def separable_convolution_block(self, input_channels=16, output_channels=32, kernel_size=3, stride=1, padding=1): ''' Append Separable convolution Block to custom network Args: input_channels (int): Number of input features for this block output_channels (int): Number of output features for this block kernel_size (int): Kernel matrix shape for all layers in this block stride (int): kernel movement stride padding (int, tuple): external zero padding on input Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; network.append(self.convolution(output_channels=input_channels, kernel_size=kernel_size, stride=stride, padding=padding, groups=input_channels)); network.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1)); return network; ##################################################################################################################################### ##################################################################################################################################### def mobilenet_v2_inverted_linear_bottleneck_block(self, output_channels=32, bottleneck_width=4, stride=1): ''' Append Mobilenet V2 Inverted Linear Bottleneck Block to custom network Args: output_channels (int): Number of output features for this block stride (int): kernel movement stride bottleneck_width (int): Bottleneck dimensions for reducing number of features Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; if(bottleneck_width != 1): branch_1.append(self.convolution(output_channels=output_channels//bottleneck_width, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); sep_conv = self.separable_convolution_block(input_channels=output_channels//bottleneck_width, output_channels=output_channels//bottleneck_width, kernel_size=3, stride=stride); branch_1.append(sep_conv); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_2 = []; branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.add()); network.append(subnetwork); return network; ##################################################################################################################################### ##################################################################################################################################### def squeezenet_fire_block(self, squeeze_channels=16, expand_channels_1x1=32, expand_channels_3x3=64): ''' Append Squeezenet Fire Block to custom network Args: squeeze_channels (int): Number of output features for this block expand_channels_1x1 (int): Number of convolution_1x1 features for this block expand_channels_3x3 (int): Number of convolution_3x3 features for this block bottleneck_width (int): Bottleneck dimensions for reducing number of features Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; network.append(self.convolution(output_channels=squeeze_channels, kernel_size=1, stride=1)); network.append(self.relu()); subnetwork = []; branch_1 = []; branch_2 = []; branch_1.append(self.convolution(output_channels=expand_channels_1x1, kernel_size=1, stride=1)); branch_1.append(self.relu()); branch_2.append(self.convolution(output_channels=expand_channels_3x3, kernel_size=3, stride=1)); branch_2.append(self.relu()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.concatenate()); network.append(subnetwork); return network; ##################################################################################################################################### ##################################################################################################################################### def densenet_block(self, bottleneck_size=4, growth_rate=16, dropout=0.2): ''' Append Densenet Block to custom network Args: bottleneck_size (int): Bottleneck dimensions for reducing number of features growth_rate (int): Expansion rate for convolution layers for this block dropout (float): Prbability for dropout layer post convolution Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_2 = []; branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=bottleneck_size*growth_rate, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=growth_rate, kernel_size=3, stride=1)); branch_1.append(self.dropout(drop_probability=dropout)); branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.concatenate()); network.append(subnetwork); return network; ##################################################################################################################################### ##################################################################################################################################### def conv_bn_relu_block(self, output_channels=64, kernel_size=1, stride=1, padding=None): ''' Append Conv->batch_norm->relu Block to custom network Args: output_channels (int): Number of output features for this block kernel_size (int): Kernel matrix shape for all layers in this block stride (int): kernel movement stride padding (int, tuple): external zero padding on input Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; if(padding): network.append(self.convolution(output_channels=output_channels, kernel_size=kernel_size, stride=stride, padding=padding)); else: network.append(self.convolution(output_channels=output_channels, kernel_size=kernel_size, stride=stride)); network.append(self.batch_normalization()); network.append(self.relu()); return network; ##################################################################################################################################### ##################################################################################################################################### def inception_a_block(self, pooling_branch_channels=32, pool_type="avg"): ''' Append Inception-A Block to custom network Args: pooling_branch_channels (int): Number of features for conv layers in pooling branch pool_type (str): Either of these types - "avg" - Average pooling - "max" - Max pooling Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_2 = []; branch_3 = []; branch_4 = []; branch_1.append(self.conv_bn_relu_block(output_channels=64, kernel_size=1)) branch_2.append(self.conv_bn_relu_block(output_channels=48, kernel_size=1)); branch_2.append(self.conv_bn_relu_block(output_channels=64, kernel_size=5)); branch_3.append(self.conv_bn_relu_block(output_channels=64, kernel_size=1)); branch_3.append(self.conv_bn_relu_block(output_channels=96, kernel_size=3)); branch_3.append(self.conv_bn_relu_block(output_channels=96, kernel_size=3)); if(pool_type=="avg"): branch_4.append(self.average_pooling(kernel_size=3, stride=1, padding=1)); else: branch_4.append(self.max_pooling(kernel_size=3, stride=1, padding=1)); branch_4.append(self.conv_bn_relu_block(output_channels=pooling_branch_channels, kernel_size=1)); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(branch_3); subnetwork.append(branch_4); subnetwork.append(self.concatenate()); network.append(subnetwork); return network; ##################################################################################################################################### ##################################################################################################################################### def inception_b_block(self, pool_type="avg"): ''' Append Inception-B Block to custom network Args: pool_type (str): Either of these types - "avg" - Average pooling - "max" - Max pooling Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_2 = []; branch_3 = []; branch_4 = []; branch_1.append(self.conv_bn_relu_block(output_channels=384, kernel_size=3)) branch_2.append(self.conv_bn_relu_block(output_channels=64, kernel_size=1)); branch_2.append(self.conv_bn_relu_block(output_channels=96, kernel_size=3)); branch_2.append(self.conv_bn_relu_block(output_channels=96, kernel_size=3)); if(pool_type=="avg"): branch_3.append(self.average_pooling(kernel_size=3, stride=1, padding=1)); else: branch_3.append(self.max_pooling(kernel_size=3, stride=1, padding=1)); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(branch_3); subnetwork.append(self.concatenate()); network.append(subnetwork); return network; ##################################################################################################################################### ##################################################################################################################################### def inception_c_block(self, channels_7x7=3, pool_type="avg"): ''' Append Inception-C Block to custom network Args: channels_7x7 (int): Number of features for conv layers in channels_7x7 branch pool_type (str): Either of these types - "avg" - Average pooling - "max" - Max pooling Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_2 = []; branch_3 = []; branch_4 = []; branch_1.append(self.conv_bn_relu_block(output_channels=192, kernel_size=1)) branch_2.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=1)); branch_2.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=(1, 7), padding=(0, 3))); branch_2.append(self.conv_bn_relu_block(output_channels=192, kernel_size=(7, 1), padding=(3, 0))); branch_3.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=1)); branch_3.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=(1, 7), padding=(0, 3))); branch_3.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=(7, 1), padding=(3, 0))); branch_3.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=(1, 7), padding=(0, 3))); branch_3.append(self.conv_bn_relu_block(output_channels=192, kernel_size=(7, 1), padding=(3, 0))); if(pool_type=="avg"): branch_4.append(self.average_pooling(kernel_size=3, stride=1, padding=1)); else: branch_4.append(self.max_pooling(kernel_size=3, stride=1, padding=1)); branch_4.append(self.conv_bn_relu_block(output_channels=192, kernel_size=1)); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(branch_3); subnetwork.append(branch_4); subnetwork.append(self.concatenate()); network.append(subnetwork); return network; ##################################################################################################################################### ##################################################################################################################################### def inception_d_block(self, pool_type="avg"): ''' Append Inception-D Block to custom network Args: pool_type (str): Either of these types - "avg" - Average pooling - "max" - Max pooling Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_2 = []; branch_3 = []; branch_4 = []; branch_1.append(self.conv_bn_relu_block(output_channels=192, kernel_size=1)) branch_1.append(self.conv_bn_relu_block(output_channels=320, kernel_size=3, stride=2)) branch_2.append(self.conv_bn_relu_block(output_channels=192, kernel_size=1)); branch_2.append(self.conv_bn_relu_block(output_channels=192, kernel_size=(1, 7), padding=(0, 3))); branch_2.append(self.conv_bn_relu_block(output_channels=192, kernel_size=(7, 1), padding=(3, 0))); branch_2.append(self.conv_bn_relu_block(output_channels=192, kernel_size=3, stride=2)); if(pool_type=="avg"): branch_3.append(self.average_pooling(kernel_size=3, stride=2)); else: branch_3.append(self.max_pooling(kernel_size=3, stride=2)); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(branch_3); subnetwork.append(self.concatenate()); network.append(subnetwork); return network; ##################################################################################################################################### ##################################################################################################################################### def subbranch_block(self): ''' Append sub-branch Block to custom network Args: None Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' subnetwork = []; branch_1 = []; branch_2 = []; branch_1.append(self.conv_bn_relu_block(output_channels=384, kernel_size=(1, 3), padding=(0, 1))); branch_2.append(self.conv_bn_relu_block(output_channels=384, kernel_size=(3, 1), padding=(1, 0))); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.concatenate()); return subnetwork; ##################################################################################################################################### ##################################################################################################################################### def inception_e_block(self, pool_type="avg"): ''' Append Inception-C Block to custom network Args: pool_type (str): Either of these types - "avg" - Average pooling - "max" - Max pooling Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_2 = []; branch_3 = []; branch_4 = []; branch_1.append(self.conv_bn_relu_block(output_channels=320, kernel_size=1)) branch_2.append(self.conv_bn_relu_block(output_channels=384, kernel_size=1)); branch_2.append(self.subbranch_block()); branch_3.append(self.conv_bn_relu_block(output_channels=448, kernel_size=1)); branch_3.append(self.conv_bn_relu_block(output_channels=384, kernel_size=3, padding=1)); branch_3.append(self.subbranch_block()); if(pool_type=="avg"): branch_4.append(self.average_pooling(kernel_size=3, stride=1, padding=1)); else: branch_4.append(self.max_pooling(kernel_size=3, stride=1, padding=1)); branch_4.append(self.conv_bn_relu_block(output_channels=192, kernel_size=1)); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(branch_3); subnetwork.append(branch_4); subnetwork.append(self.concatenate()); network.append(subnetwork); return network;Ancestors
- pytorch.finetune.level_7_aux_main.prototype_aux
- pytorch.finetune.level_6_params_main.prototype_params
- pytorch.finetune.level_5_state_base.finetune_state
- pytorch.finetune.level_4_evaluation_base.finetune_evaluation
- pytorch.finetune.level_3_training_base.finetune_training
- pytorch.finetune.level_2_model_base.finetune_model
- pytorch.finetune.level_1_dataset_base.finetune_dataset
- system.base_class.system
Methods
def add(self, uid=None)-
Append elementwise addition layer to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def add(self, uid=None): ''' Append elementwise addition layer to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "add"; return tmp; def append_celu(self, alpha=1.0, final_layer=False)-
Append continuously differentiable exponential linear unit activation to base network in transfer learning
Args
alpha:float- Multiplicative factor
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_celu(self, alpha=1.0, final_layer=False): ''' Append continuously differentiable exponential linear unit activation to base network in transfer learning Args: alpha (float): Multiplicative factor final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_celu(self.system_dict, alpha=alpha, final_layer=final_layer); def append_dropout(self, probability=0.5, final_layer=False)-
Append dropout layer to base network in transfer learning
Args
probability:float- Droping probability of neurons in next layer
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_dropout(self, probability=0.5, final_layer=False): ''' Append dropout layer to base network in transfer learning Args: probability (float): Droping probability of neurons in next layer final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = layer_dropout(self.system_dict, probability=probability, final_layer=final_layer); def append_elu(self, alpha=1.0, final_layer=False)-
Append exponential linear unit activation to base network in transfer learning
Args
alpha:float- Multiplicatve factor.
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_elu(self, alpha=1.0, final_layer=False): ''' Append exponential linear unit activation to base network in transfer learning Args: alpha (float): Multiplicatve factor. final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_elu(self.system_dict, alpha=alpha, final_layer=final_layer); def append_hardshrink(self, lambd=0.5, final_layer=False)-
Append hardshrink activation to base network in transfer learning
Args
lambd:float- thereshold limit value
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_hardshrink(self, lambd=0.5, final_layer=False): ''' Append hardshrink activation to base network in transfer learning Args: lambd (float): thereshold limit value final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_hardshrink(self.system_dict, lambd=lambd, final_layer=final_layer); def append_hardtanh(self, min_val=-1.0, max_val=1.0, final_layer=False)-
Append hardtanh activation to base network in transfer learning
Args
min_val:float- lower thereshold limit value
max_val:float- upper thereshold limit value
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_hardtanh(self, min_val=-1.0, max_val=1.0, final_layer=False): ''' Append hardtanh activation to base network in transfer learning Args: min_val (float): lower thereshold limit value max_val (float): upper thereshold limit value final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_hardtanh(self.system_dict, min_val=min_val, max_val=max_val, final_layer=final_layer); def append_leakyrelu(self, negative_slope=0.01, final_layer=False)-
Append Leaky - ReLU activation to base network in transfer learning
Args
negative_slope:float- Multiplicatve factor towards negative spectrum of real numbers.
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_leakyrelu(self, negative_slope=0.01, final_layer=False): ''' Append Leaky - ReLU activation to base network in transfer learning Args: negative_slope (float): Multiplicatve factor towards negative spectrum of real numbers. final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_leakyrelu(self.system_dict, negative_slope=negative_slope, final_layer=final_layer); def append_linear(self, num_neurons=False, final_layer=False)-
Append dense (fully connected) layer to base network in transfer learning
Args
num_neurons:int- Number of neurons in the dense layer
final_layer:bool- If True, then number of neurons are directly set as number of classes in dataset for single label type classification
Returns
None
Expand source code
def append_linear(self, num_neurons=False, final_layer=False): ''' Append dense (fully connected) layer to base network in transfer learning Args: num_neurons (int): Number of neurons in the dense layer final_layer (bool): If True, then number of neurons are directly set as number of classes in dataset for single label type classification Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: if(not num_neurons): num_neurons = self.system_dict["dataset"]["params"]["num_classes"]; self.system_dict = layer_linear(self.system_dict, num_neurons=num_neurons, final_layer=final_layer); def append_logsigmoid(self, final_layer=False)-
Append logsigmoid activation to base network in transfer learning
Args
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_logsigmoid(self, final_layer=False): ''' Append logsigmoid activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_logsigmoid(self.system_dict, final_layer=final_layer); def append_logsoftmax(self, final_layer=False)-
Append log-softmax activation to base network in transfer learning
Args
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_logsoftmax(self, final_layer=False): ''' Append log-softmax activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_logsoftmax(self.system_dict, final_layer=final_layer); def append_prelu(self, num_parameters=1, init=0.25, final_layer=False)-
Append Learnable parameerized rectified linear unit activation to base network in transfer learning
Args
init:float- Initialization value for multiplicatve factor towards negative spectrum of real numbers.
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_prelu(self, num_parameters=1, init=0.25, final_layer=False): ''' Append Learnable parameerized rectified linear unit activation to base network in transfer learning Args: init (float): Initialization value for multiplicatve factor towards negative spectrum of real numbers. final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_prelu(self.system_dict, num_parameters=num_parameters, init=init, final_layer=final_layer); def append_relu(self, final_layer=False)-
Append rectified linear unit activation to base network in transfer learning
Args
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_relu(self, final_layer=False): ''' Append rectified linear unit activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_relu(self.system_dict, final_layer=final_layer); def append_relu6(self, final_layer=False)-
Append rectified linear unit activation - 6 to base network in transfer learning
Args
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_relu6(self, final_layer=False): ''' Append rectified linear unit activation - 6 to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_relu6(self.system_dict, final_layer=final_layer); def append_rrelu(self, lower=0.125, upper=0.333, final_layer=False)-
Append randomized rectified linear unit activation to base network in transfer learning
Args
lower:float- Lower limit of uniform distribution for the ramdomized multiplicative factor
upper:float- Upper limit of uniform distribution for the ramdomized multiplicative factor
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_rrelu(self, lower=0.125, upper=0.333, final_layer=False): ''' Append randomized rectified linear unit activation to base network in transfer learning Args: lower (float): Lower limit of uniform distribution for the ramdomized multiplicative factor upper (float): Upper limit of uniform distribution for the ramdomized multiplicative factor final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_rrelu(self.system_dict, lower=lower, upper=upper, final_layer=final_layer) def append_selu(self, final_layer=False)-
Append scaled exponential linear unit activation to base network in transfer learning
Args
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_selu(self, final_layer=False): ''' Append scaled exponential linear unit activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_selu(self.system_dict, final_layer=final_layer); def append_sigmoid(self, final_layer=False)-
Append rectified linear unit activation to base network in transfer learning
Args
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_sigmoid(self, final_layer=False): ''' Append rectified linear unit activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_sigmoid(self.system_dict, final_layer=final_layer); def append_softmax(self, final_layer=False)-
Append softmax activation to base network in transfer learning
Args
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_softmax(self, final_layer=False): ''' Append softmax activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_softmax(self.system_dict, final_layer=final_layer); def append_softmin(self, final_layer=False)-
Append softmin activation to base network in transfer learning
Args
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_softmin(self, final_layer=False): ''' Append softmin activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_softmin(self.system_dict, final_layer=final_layer); def append_softplus(self, beta=1, threshold=20, final_layer=False)-
Append softplus activation to base network in transfer learning
Args
beta:float- Multiplicative factor
threshold:int- softplus (thresholded relu) limit
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_softplus(self, beta=1, threshold=20, final_layer=False): ''' Append softplus activation to base network in transfer learning Args: beta (float): Multiplicative factor threshold (int): softplus (thresholded relu) limit final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_softplus(self.system_dict, beta=beta, threshold=threshold, final_layer=final_layer); def append_softshrink(self, lambd=0.5, final_layer=False)-
Append softshrink activation to base network in transfer learning
Args
lambd:float- thereshold limit value
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_softshrink(self, lambd=0.5, final_layer=False): ''' Append softshrink activation to base network in transfer learning Args: lambd (float): thereshold limit value final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_softshrink(self.system_dict, lambd=lambd, final_layer=final_layer); def append_softsign(self, final_layer=False)-
Append tanh activation to base network in transfer learning
Args
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_softsign(self, final_layer=False): ''' Append tanh activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_softsign(self.system_dict, final_layer=final_layer); def append_tanh(self, final_layer=False)-
Append tanh activation to base network in transfer learning
Args
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_tanh(self, final_layer=False): ''' Append tanh activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_tanh(self.system_dict, final_layer=final_layer); def append_tanhshrink(self, final_layer=False)-
Append tanhshrink activation to base network in transfer learning
Args
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_tanhshrink(self, final_layer=False): ''' Append tanhshrink activation to base network in transfer learning Args: final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_tanhshrink(self.system_dict, final_layer=final_layer); def append_threshold(self, threshold, value, final_layer=False)-
Append threshold activation to base network in transfer learning
Args
threshold:float- thereshold limit value
value:float- replacement value if input is off-limits
final_layer:bool- Indicator that this layer marks the end of network.
Returns
None
Expand source code
def append_threshold(self, threshold, value, final_layer=False): ''' Append threshold activation to base network in transfer learning Args: threshold (float): thereshold limit value value (float): replacement value if input is off-limits final_layer (bool): Indicator that this layer marks the end of network. Returns: None ''' if(self.system_dict["model"]["final_layer"]): msg = "Cannot append more layers.\n"; msg += "Tip: Previously appended layer termed as final layer"; raise ConstraintError(msg); else: self.system_dict = activation_threshold(self.system_dict, threshold, value, final_layer=final_layer); def average_pooling(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, include_padding_in_calculation=True, layout='NCHW', uid=None)-
Append 2d-average-pooling to custom network
Args
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
dilation:int- Factor for dilated pooling
return_indices:bool- Fixed value set as False
ceil_mode:bool- If True, apply ceil math operation post pooling
include_padding_in_calculation:bool- If True, padding will be considered.
layout:str- Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def average_pooling(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, include_padding_in_calculation=True, layout='NCHW', uid=None): ''' Append 2d-average-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False ceil_mode (bool): If True, apply ceil math operation post pooling include_padding_in_calculation (bool): If True, padding will be considered. layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "average_pooling2d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["include_padding_in_calculation"] = include_padding_in_calculation; tmp["params"]["layout"] = layout; return tmp; def average_pooling1d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, include_padding_in_calculation=True, layout='NCW', uid=None)-
Append 1d-average-pooling to custom network
Args
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
dilation:int- Factor for dilated pooling
return_indices:bool- Fixed value set as False
include_padding_in_calculation:bool- If True, padding will be considered.
ceil_mode:bool- If True, apply ceil math operation post pooling
layout:str- Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def average_pooling1d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, include_padding_in_calculation=True, layout='NCW', uid=None): ''' Append 1d-average-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False include_padding_in_calculation (bool): If True, padding will be considered. ceil_mode (bool): If True, apply ceil math operation post pooling layout (str): Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "average_pooling1d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["include_padding_in_calculation"] = include_padding_in_calculation; tmp["params"]["layout"] = layout; return tmp; def average_pooling2d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, include_padding_in_calculation=True, layout='NCHW', uid=None)-
Append 2d-average-pooling to custom network
Args
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
dilation:int- Factor for dilated pooling
return_indices:bool- Fixed value set as False
ceil_mode:bool- If True, apply ceil math operation post pooling
include_padding_in_calculation:bool- If True, padding will be considered.
layout:str- Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def average_pooling2d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, include_padding_in_calculation=True, layout='NCHW', uid=None): ''' Append 2d-average-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False ceil_mode (bool): If True, apply ceil math operation post pooling include_padding_in_calculation (bool): If True, padding will be considered. layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "average_pooling2d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["include_padding_in_calculation"] = include_padding_in_calculation; tmp["params"]["layout"] = layout; return tmp; def average_pooling3d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, include_padding_in_calculation=True, layout='NCDHW', uid=None)-
Append 3d-average-pooling to custom network
Args
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
dilation:int- Factor for dilated pooling
return_indices:bool- Fixed value set as False
ceil_mode:bool- If True, apply ceil math operation post pooling
include_padding_in_calculation:bool- If True, padding will be considered.
layout:str- Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def average_pooling3d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, include_padding_in_calculation=True, layout='NCDHW', uid=None): ''' Append 3d-average-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False ceil_mode (bool): If True, apply ceil math operation post pooling include_padding_in_calculation (bool): If True, padding will be considered. layout (str): Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "average_pooling3d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["include_padding_in_calculation"] = include_padding_in_calculation; tmp["params"]["layout"] = layout; return tmp; def batch_normalization(self, moving_average_momentum=0.1, epsilon=1e-05, use_trainable_parameters=True, activate_scale_shift_operation=False, uid=None)-
Append batch normalization layer to custom network
Args
moving_average_momentum:float- Normalization momentum value
epsilon:float- Value to avoid division by zero
use_trainable_paramemetrs:bool- If True, batch norm turns into a trainable layer
activate_scale_shift_operation:bool- Fixed status - False
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def batch_normalization(self, moving_average_momentum=0.1, epsilon=0.00001, use_trainable_parameters=True, activate_scale_shift_operation=False, uid=None): ''' Append batch normalization layer to custom network Args: moving_average_momentum (float): Normalization momentum value epsilon (float): Value to avoid division by zero use_trainable_paramemetrs (bool): If True, batch norm turns into a trainable layer activate_scale_shift_operation (bool): Fixed status - False uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "batch_normalization"; tmp["params"] = {}; tmp["params"]["moving_average_momentum"] = moving_average_momentum; tmp["params"]["epsilon"] = epsilon; tmp["params"]["use_trainable_parameters"] = use_trainable_parameters; tmp["params"]["activate_scale_shift_operation"] = activate_scale_shift_operation; return tmp; def celu(self, alpha=1.0, uid=None)-
Append continuously differentiable exponential linear unit activation to custom network
Args
alpha:float- Multiplicative factor
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def celu(self, alpha=1.0, uid=None): ''' Append continuously differentiable exponential linear unit activation to custom network Args: alpha (float): Multiplicative factor uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "celu"; tmp["params"] = {}; tmp["params"]["alpha"] = alpha; return tmp; def concatenate(self, uid=None)-
Append concatenation layer to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def concatenate(self, uid=None): ''' Append concatenation layer to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "concatenate"; return tmp; def conv_bn_relu_block(self, output_channels=64, kernel_size=1, stride=1, padding=None)-
Append Conv->batch_norm->relu Block to custom network
Args
output_channels:int- Number of output features for this block
kernel_size:int- Kernel matrix shape for all layers in this block
stride:int- kernel movement stride
padding:int,tuple- external zero padding on input
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def conv_bn_relu_block(self, output_channels=64, kernel_size=1, stride=1, padding=None): ''' Append Conv->batch_norm->relu Block to custom network Args: output_channels (int): Number of output features for this block kernel_size (int): Kernel matrix shape for all layers in this block stride (int): kernel movement stride padding (int, tuple): external zero padding on input Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; if(padding): network.append(self.convolution(output_channels=output_channels, kernel_size=kernel_size, stride=stride, padding=padding)); else: network.append(self.convolution(output_channels=output_channels, kernel_size=kernel_size, stride=stride)); network.append(self.batch_normalization()); network.append(self.relu()); return network; def convolution(self, output_channels=3, kernel_size=3, stride=1, padding='in_eq_out', groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None)-
Append 2d-convolution to custom network
Args
output_channels:int- Number of output features for this layers
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
groups:int- Number of groups for grouped convolution
dilation:int- Factor for dilated convolution
use_bias:bool- If True, learnable bias is added
layout:str- Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def convolution(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None): ''' Append 2d-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "convolution2d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp; def convolution1d(self, output_channels=3, kernel_size=3, stride=1, padding='in_eq_out', groups=1, dilation=1, use_bias=True, layout='NCW', uid=None)-
Append 1d-convolution to custom network
Args
output_channels:int- Number of output features for this layers
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
groups:int- Number of groups for grouped convolution
dilation:int- Factor for dilated convolution
use_bias:bool- If True, learnable bias is added
layout:str- Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def convolution1d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", groups=1, dilation=1, use_bias=True, layout='NCW', uid=None): ''' Append 1d-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "convolution1d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp; def convolution2d(self, output_channels=3, kernel_size=3, stride=1, padding='in_eq_out', groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None)-
Append 2d-convolution to custom network
Args
output_channels:int- Number of output features for this layers
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
groups:int- Number of groups for grouped convolution
dilation:int- Factor for dilated convolution
use_bias:bool- If True, learnable bias is added
layout:str- Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def convolution2d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None): ''' Append 2d-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "convolution2d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp; def convolution3d(self, output_channels=3, kernel_size=3, stride=1, padding='in_eq_out', groups=1, dilation=1, use_bias=True, layout='NCDHW', uid=None)-
Append 3d-convolution to custom network
Args
output_channels:int- Number of output features for this layers
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
groups:int- Number of groups for grouped convolution
dilation:int- Factor for dilated convolution
use_bias:bool- If True, learnable bias is added
layout:str- Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def convolution3d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", groups=1, dilation=1, use_bias=True, layout='NCDHW', uid=None): ''' Append 3d-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "convolution3d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp; def densenet_block(self, bottleneck_size=4, growth_rate=16, dropout=0.2)-
Append Densenet Block to custom network
Args
bottleneck_size:int- Bottleneck dimensions for reducing number of features
growth_rate:int- Expansion rate for convolution layers for this block
dropout:float- Prbability for dropout layer post convolution
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def densenet_block(self, bottleneck_size=4, growth_rate=16, dropout=0.2): ''' Append Densenet Block to custom network Args: bottleneck_size (int): Bottleneck dimensions for reducing number of features growth_rate (int): Expansion rate for convolution layers for this block dropout (float): Prbability for dropout layer post convolution Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_2 = []; branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=bottleneck_size*growth_rate, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=growth_rate, kernel_size=3, stride=1)); branch_1.append(self.dropout(drop_probability=dropout)); branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.concatenate()); network.append(subnetwork); return network; def dropout(self, drop_probability=0.2, axes=(), uid=None)-
Append dropout layer to custom network
Args
drop_probability:float- Probability for not considering neurons in the output
axes:tuple- Channel axis to implement dropout over
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def dropout(self, drop_probability=0.2, axes=(), uid=None): ''' Append dropout layer to custom network Args: drop_probability (float): Probability for not considering neurons in the output axes (tuple): Channel axis to implement dropout over uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "dropout"; tmp["params"] = {}; tmp["params"]["drop_probability"] = drop_probability; tmp["params"]["axes"] = axes; return tmp; def elu(self, alpha=1.0, uid=None)-
Append exponential linear unit activation to custom network
Args
alpha:float- Multiplicative factor
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def elu(self, alpha=1.0, uid=None): ''' Append exponential linear unit activation to custom network Args: alpha (float): Multiplicative factor uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "elu"; tmp["params"] = {}; tmp["params"]["alpha"] = alpha; return tmp; def flatten(self, uid=None)-
Append flatten layer to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def flatten(self, uid=None): ''' Append flatten layer to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "flatten"; tmp["params"] = {}; return tmp; def fully_connected(self, units=512, use_bias=True, flatten=True, uid=None)-
Append fully-connected (dense) layer to custom network
Args
units:int- Number of neurons in the layer
use_bias:bool- If True, learnable bias is added
flatten:bool- Fixed to True
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def fully_connected(self, units=512, use_bias=True, flatten=True, uid=None): ''' Append fully-connected (dense) layer to custom network Args: units (int): Number of neurons in the layer use_bias (bool): If True, learnable bias is added flatten (bool): Fixed to True uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "fully_connected"; tmp["params"] = {}; tmp["params"]["units"] = units; tmp["params"]["use_bias"] = use_bias; tmp["params"]["flatten"] = flatten; return tmp; def global_average_pooling(self, layout='NCHW', uid=None)-
Append 2d-global-average-pooling to custom network
Args
layout:str- Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def global_average_pooling(self, layout='NCHW', uid=None): ''' Append 2d-global-average-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_average_pooling2d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; def global_average_pooling1d(self, layout='NCW', uid=None)-
Append 1d-global-average-pooling to custom network
Args
layout:str- Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def global_average_pooling1d(self, layout='NCW', uid=None): ''' Append 1d-global-average-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_average_pooling1d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; def global_average_pooling2d(self, layout='NCHW', uid=None)-
Append 2d-global-average-pooling to custom network
Args
layout:str- Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def global_average_pooling2d(self, layout='NCHW', uid=None): ''' Append 2d-global-average-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_average_pooling2d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; def global_average_pooling3d(self, layout='NCDHW', uid=None)-
Append 3d-global-average-pooling to custom network
Args
layout:str- Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def global_average_pooling3d(self, layout='NCDHW', uid=None): ''' Append 3d-global-average-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_average_pooling3d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; def global_max_pooling(self, layout='NCHW', uid=None)-
Append 2d-global-max-pooling to custom network
Args
layout:str- Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def global_max_pooling(self, layout='NCHW', uid=None): ''' Append 2d-global-max-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_max_pooling2d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; def global_max_pooling1d(self, layout='NCW', uid=None)-
Append 1d-global-max-pooling to custom network
Args
layout:str- Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def global_max_pooling1d(self, layout='NCW', uid=None): ''' Append 1d-global-max-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_max_pooling1d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; def global_max_pooling2d(self, layout='NCHW', uid=None)-
Append 2d-global-max-pooling to custom network
Args
layout:str- Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def global_max_pooling2d(self, layout='NCHW', uid=None): ''' Append 2d-global-max-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_max_pooling2d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; def global_max_pooling3d(self, layout='NCDHW', uid=None)-
Append 3d-global-max-pooling to custom network
Args
layout:str- Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def global_max_pooling3d(self, layout='NCDHW', uid=None): ''' Append 3d-global-max-pooling to custom network Args: layout (str): Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "global_max_pooling3d"; tmp["params"] = {}; tmp["params"]["layout"] = layout; return tmp; def hardshrink(self, threshold=0.5, uid=None)-
Append hardshrink activation to custom network
Args
threshold:float- Limit for activation
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def hardshrink(self, threshold=0.5, uid=None): ''' Append hardshrink activation to custom network Args: threshold (float): Limit for activation uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "hardshrink"; tmp["params"] = {}; tmp["params"]["threshold"] = threshold; return tmp; def hardtanh(self, min_val=-1.0, max_val=1.0, uid=None)-
Append Hard-Tanh activation to custom network
Args
min_val:float- lower thereshold limit value
max_val:float- upper thereshold limit value
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def hardtanh(self, min_val=-1.0, max_val=1.0, uid=None): ''' Append Hard-Tanh activation to custom network Args: min_val (float): lower thereshold limit value max_val (float): upper thereshold limit value uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "hardtanh"; tmp["params"] = {}; tmp["params"]["min_val"] = min_val; tmp["params"]["max_val"] = max_val; return tmp; def identity(self, uid=None)-
Append identity layer to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def identity(self, uid=None): ''' Append identity layer to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "identity"; tmp["params"] = {}; return tmp; def inception_a_block(self, pooling_branch_channels=32, pool_type='avg')-
Append Inception-A Block to custom network
Args
pooling_branch_channels:int- Number of features for conv layers in pooling branch
pool_type:str- Either of these types - "avg" - Average pooling - "max" - Max pooling
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def inception_a_block(self, pooling_branch_channels=32, pool_type="avg"): ''' Append Inception-A Block to custom network Args: pooling_branch_channels (int): Number of features for conv layers in pooling branch pool_type (str): Either of these types - "avg" - Average pooling - "max" - Max pooling Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_2 = []; branch_3 = []; branch_4 = []; branch_1.append(self.conv_bn_relu_block(output_channels=64, kernel_size=1)) branch_2.append(self.conv_bn_relu_block(output_channels=48, kernel_size=1)); branch_2.append(self.conv_bn_relu_block(output_channels=64, kernel_size=5)); branch_3.append(self.conv_bn_relu_block(output_channels=64, kernel_size=1)); branch_3.append(self.conv_bn_relu_block(output_channels=96, kernel_size=3)); branch_3.append(self.conv_bn_relu_block(output_channels=96, kernel_size=3)); if(pool_type=="avg"): branch_4.append(self.average_pooling(kernel_size=3, stride=1, padding=1)); else: branch_4.append(self.max_pooling(kernel_size=3, stride=1, padding=1)); branch_4.append(self.conv_bn_relu_block(output_channels=pooling_branch_channels, kernel_size=1)); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(branch_3); subnetwork.append(branch_4); subnetwork.append(self.concatenate()); network.append(subnetwork); return network; def inception_b_block(self, pool_type='avg')-
Append Inception-B Block to custom network
Args
pool_type:str- Either of these types - "avg" - Average pooling - "max" - Max pooling
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def inception_b_block(self, pool_type="avg"): ''' Append Inception-B Block to custom network Args: pool_type (str): Either of these types - "avg" - Average pooling - "max" - Max pooling Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_2 = []; branch_3 = []; branch_4 = []; branch_1.append(self.conv_bn_relu_block(output_channels=384, kernel_size=3)) branch_2.append(self.conv_bn_relu_block(output_channels=64, kernel_size=1)); branch_2.append(self.conv_bn_relu_block(output_channels=96, kernel_size=3)); branch_2.append(self.conv_bn_relu_block(output_channels=96, kernel_size=3)); if(pool_type=="avg"): branch_3.append(self.average_pooling(kernel_size=3, stride=1, padding=1)); else: branch_3.append(self.max_pooling(kernel_size=3, stride=1, padding=1)); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(branch_3); subnetwork.append(self.concatenate()); network.append(subnetwork); return network; def inception_c_block(self, channels_7x7=3, pool_type='avg')-
Append Inception-C Block to custom network
Args
channels_7x7:int- Number of features for conv layers in channels_7x7 branch
pool_type:str- Either of these types - "avg" - Average pooling - "max" - Max pooling
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def inception_c_block(self, channels_7x7=3, pool_type="avg"): ''' Append Inception-C Block to custom network Args: channels_7x7 (int): Number of features for conv layers in channels_7x7 branch pool_type (str): Either of these types - "avg" - Average pooling - "max" - Max pooling Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_2 = []; branch_3 = []; branch_4 = []; branch_1.append(self.conv_bn_relu_block(output_channels=192, kernel_size=1)) branch_2.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=1)); branch_2.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=(1, 7), padding=(0, 3))); branch_2.append(self.conv_bn_relu_block(output_channels=192, kernel_size=(7, 1), padding=(3, 0))); branch_3.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=1)); branch_3.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=(1, 7), padding=(0, 3))); branch_3.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=(7, 1), padding=(3, 0))); branch_3.append(self.conv_bn_relu_block(output_channels=channels_7x7, kernel_size=(1, 7), padding=(0, 3))); branch_3.append(self.conv_bn_relu_block(output_channels=192, kernel_size=(7, 1), padding=(3, 0))); if(pool_type=="avg"): branch_4.append(self.average_pooling(kernel_size=3, stride=1, padding=1)); else: branch_4.append(self.max_pooling(kernel_size=3, stride=1, padding=1)); branch_4.append(self.conv_bn_relu_block(output_channels=192, kernel_size=1)); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(branch_3); subnetwork.append(branch_4); subnetwork.append(self.concatenate()); network.append(subnetwork); return network; def inception_d_block(self, pool_type='avg')-
Append Inception-D Block to custom network
Args
pool_type:str- Either of these types - "avg" - Average pooling - "max" - Max pooling
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def inception_d_block(self, pool_type="avg"): ''' Append Inception-D Block to custom network Args: pool_type (str): Either of these types - "avg" - Average pooling - "max" - Max pooling Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_2 = []; branch_3 = []; branch_4 = []; branch_1.append(self.conv_bn_relu_block(output_channels=192, kernel_size=1)) branch_1.append(self.conv_bn_relu_block(output_channels=320, kernel_size=3, stride=2)) branch_2.append(self.conv_bn_relu_block(output_channels=192, kernel_size=1)); branch_2.append(self.conv_bn_relu_block(output_channels=192, kernel_size=(1, 7), padding=(0, 3))); branch_2.append(self.conv_bn_relu_block(output_channels=192, kernel_size=(7, 1), padding=(3, 0))); branch_2.append(self.conv_bn_relu_block(output_channels=192, kernel_size=3, stride=2)); if(pool_type=="avg"): branch_3.append(self.average_pooling(kernel_size=3, stride=2)); else: branch_3.append(self.max_pooling(kernel_size=3, stride=2)); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(branch_3); subnetwork.append(self.concatenate()); network.append(subnetwork); return network; def inception_e_block(self, pool_type='avg')-
Append Inception-C Block to custom network
Args
pool_type:str- Either of these types - "avg" - Average pooling - "max" - Max pooling
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def inception_e_block(self, pool_type="avg"): ''' Append Inception-C Block to custom network Args: pool_type (str): Either of these types - "avg" - Average pooling - "max" - Max pooling Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_2 = []; branch_3 = []; branch_4 = []; branch_1.append(self.conv_bn_relu_block(output_channels=320, kernel_size=1)) branch_2.append(self.conv_bn_relu_block(output_channels=384, kernel_size=1)); branch_2.append(self.subbranch_block()); branch_3.append(self.conv_bn_relu_block(output_channels=448, kernel_size=1)); branch_3.append(self.conv_bn_relu_block(output_channels=384, kernel_size=3, padding=1)); branch_3.append(self.subbranch_block()); if(pool_type=="avg"): branch_4.append(self.average_pooling(kernel_size=3, stride=1, padding=1)); else: branch_4.append(self.max_pooling(kernel_size=3, stride=1, padding=1)); branch_4.append(self.conv_bn_relu_block(output_channels=192, kernel_size=1)); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(branch_3); subnetwork.append(branch_4); subnetwork.append(self.concatenate()); network.append(subnetwork); return network; def instance_normalization(self, moving_average_momentum=0.1, epsilon=1e-05, use_trainable_parameters=False, uid=None)-
Append instace normalization layer to custom network
Args
moving_average_momentum:float- Normalization momentum value
epsilon:float- Value to avoid division by zero
use_trainable_paramemetrs:bool- If True, batch norm turns into a trainable layer
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def instance_normalization(self, moving_average_momentum=0.1, epsilon=0.00001, use_trainable_parameters=False, uid=None): ''' Append instace normalization layer to custom network Args: moving_average_momentum (float): Normalization momentum value epsilon (float): Value to avoid division by zero use_trainable_paramemetrs (bool): If True, batch norm turns into a trainable layer uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "instance_normalization"; tmp["params"] = {}; tmp["params"]["moving_average_momentum"] = moving_average_momentum; tmp["params"]["epsilon"] = epsilon; tmp["params"]["use_trainable_parameters"] = use_trainable_parameters; return tmp; def layer_normalization(self, moving_average_momentum=0.9, epsilon=1e-05, use_trainable_parameters=True, uid=None)-
Append layer normalization layer to custom network
Args
moving_average_momentum:float- Normalization momentum value
epsilon:float- Value to avoid division by zero
use_trainable_paramemetrs:bool- If True, batch norm turns into a trainable layer
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def layer_normalization(self, moving_average_momentum=0.9, epsilon=0.00001, use_trainable_parameters=True, uid=None): ''' Append layer normalization layer to custom network Args: moving_average_momentum (float): Normalization momentum value epsilon (float): Value to avoid division by zero use_trainable_paramemetrs (bool): If True, batch norm turns into a trainable layer uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "layer_normalization"; tmp["params"] = {}; tmp["params"]["epsilon"] = epsilon; tmp["params"]["use_trainable_parameters"] = use_trainable_parameters; return tmp; def leaky_relu(self, alpha=0.01, uid=None)-
Append leaky relu activation to custom network
Args
alpha:float- Multiplicatve factor towards negative spectrum of real numbers.
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def leaky_relu(self, alpha=0.01, uid=None): ''' Append leaky relu activation to custom network Args: alpha (float): Multiplicatve factor towards negative spectrum of real numbers. uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "leaky_relu"; tmp["params"] = {}; tmp["params"]["alpha"] = alpha; return tmp; def logsigmoid(self, uid=None)-
Append log-sigmoid activation to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def logsigmoid(self, uid=None): ''' Append log-sigmoid activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "logsigmoid"; tmp["params"] = {}; return tmp; def logsoftmax(self, uid=None)-
Append log-softmax activation to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def logsoftmax(self, uid=None): ''' Append log-softmax activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "logsoftmax"; tmp["params"] = {}; return tmp; def max_pooling(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, layout='NCHW', uid=None)-
Append 2d-max-pooling to custom network
Args
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
dilation:int- Factor for dilated pooling
return_indices:bool- Fixed value set as False
ceil_mode:bool- If True, apply ceil math operation post pooling
layout:str- Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def max_pooling(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, layout='NCHW', uid=None): ''' Append 2d-max-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False ceil_mode (bool): If True, apply ceil math operation post pooling layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "max_pooling2d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["layout"] = layout; return tmp; def max_pooling1d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, layout='NCW', uid=None)-
Append 1d-max-pooling to custom network
Args
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
dilation:int- Factor for dilated pooling
return_indices:bool- Fixed value set as False
ceil_mode:bool- If True, apply ceil math operation post pooling
layout:str- Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def max_pooling1d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, layout='NCW', uid=None): ''' Append 1d-max-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False ceil_mode (bool): If True, apply ceil math operation post pooling layout (str): Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "max_pooling1d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["layout"] = layout; return tmp; def max_pooling2d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, layout='NCHW', uid=None)-
Append 2d-max-pooling to custom network
Args
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
dilation:int- Factor for dilated pooling
return_indices:bool- Fixed value set as False
ceil_mode:bool- If True, apply ceil math operation post pooling
layout:str- Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def max_pooling2d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, layout='NCHW', uid=None): ''' Append 2d-max-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False ceil_mode (bool): If True, apply ceil math operation post pooling layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "max_pooling2d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["layout"] = layout; return tmp; def max_pooling3d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, layout='NCDHW', uid=None)-
Append 3d-max-pooling to custom network
Args
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
dilation:int- Factor for dilated pooling
return_indices:bool- Fixed value set as False
ceil_mode:bool- If True, apply ceil math operation post pooling
layout:str- Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def max_pooling3d(self, kernel_size=2, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False, layout='NCDHW', uid=None): ''' Append 3d-max-pooling to custom network Args: kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding dilation (int): Factor for dilated pooling return_indices (bool): Fixed value set as False ceil_mode (bool): If True, apply ceil math operation post pooling layout (str): Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "max_pooling3d"; tmp["params"] = {}; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["dilation"] = dilation; tmp["params"]["return_indices"] = return_indices; tmp["params"]["ceil_mode"] = ceil_mode; tmp["params"]["layout"] = layout; return tmp; def mobilenet_v2_inverted_linear_bottleneck_block(self, output_channels=32, bottleneck_width=4, stride=1)-
Append Mobilenet V2 Inverted Linear Bottleneck Block to custom network
Args
output_channels:int- Number of output features for this block
stride:int- kernel movement stride
bottleneck_width:int- Bottleneck dimensions for reducing number of features
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def mobilenet_v2_inverted_linear_bottleneck_block(self, output_channels=32, bottleneck_width=4, stride=1): ''' Append Mobilenet V2 Inverted Linear Bottleneck Block to custom network Args: output_channels (int): Number of output features for this block stride (int): kernel movement stride bottleneck_width (int): Bottleneck dimensions for reducing number of features Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; if(bottleneck_width != 1): branch_1.append(self.convolution(output_channels=output_channels//bottleneck_width, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); sep_conv = self.separable_convolution_block(input_channels=output_channels//bottleneck_width, output_channels=output_channels//bottleneck_width, kernel_size=3, stride=stride); branch_1.append(sep_conv); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_2 = []; branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.add()); network.append(subnetwork); return network; def mobilenet_v2_linear_bottleneck_block(self, output_channels=32, bottleneck_width=4, stride=1)-
Append Mobilenet V2 Linear Bottleneck Block to custom network
Args
output_channels:int- Number of output features for this block
stride:int- kernel movement stride
bottleneck_width:int- Bottleneck dimensions for reducing number of features
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def mobilenet_v2_linear_bottleneck_block(self, output_channels=32, bottleneck_width=4, stride=1): ''' Append Mobilenet V2 Linear Bottleneck Block to custom network Args: output_channels (int): Number of output features for this block stride (int): kernel movement stride bottleneck_width (int): Bottleneck dimensions for reducing number of features Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; if(bottleneck_width != 1): branch_1.append(self.convolution(output_channels=output_channels*bottleneck_width, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels*bottleneck_width, kernel_size=3, stride=stride)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_2 = []; branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.add()); network.append(subnetwork); return network; def prelu(self, uid=None)-
Append paramemeterized rectified linear unit activation to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def prelu(self, uid=None): ''' Append paramemeterized rectified linear unit activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "prelu"; tmp["params"] = {}; return tmp; def relu(self, uid=None)-
Append rectified linear unit activation to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def relu(self, uid=None): ''' Append rectified linear unit activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "relu"; tmp["params"] = {}; return tmp; def relu6(self, uid=None)-
Append relu-6 activation to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def relu6(self, uid=None): ''' Append relu-6 activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "relu6"; tmp["params"] = {}; return tmp; def resnet_v1_block(self, output_channels=16, stride=1, downsample=True)-
Append Resnet V1 Block to custom network
Args
output_channels:int- Number of output features for this block
stride:int,tuple- kernel movement stride
downsample:bool- If False, residual branch is a shortcut, Else, residual branch has non-identity layers
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def resnet_v1_block(self, output_channels=16, stride=1, downsample=True): ''' Append Resnet V1 Block to custom network Args: output_channels (int): Number of output features for this block stride (int, tuple): kernel movement stride downsample (bool): If False, residual branch is a shortcut, Else, residual branch has non-identity layers Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_1.append(self.convolution(output_channels=output_channels, kernel_size=3, stride=stride)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels, kernel_size=3, stride=1)); branch_1.append(self.batch_normalization()); branch_2 = []; if(downsample): branch_2.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=stride)); branch_2.append(self.batch_normalization()); else: branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.add()); network.append(subnetwork) network.append(self.relu()); return network; def resnet_v1_bottleneck_block(self, output_channels=16, stride=1, downsample=True)-
Append Resnet V1 Bottleneck Block to custom network
Args
output_channels:int- Number of output features for this block
stride:int,tuple- kernel movement stride
downsample:bool- If False, residual branch is a shortcut, Else, residual branch has non-identity layers
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def resnet_v1_bottleneck_block(self, output_channels=16, stride=1, downsample=True): ''' Append Resnet V1 Bottleneck Block to custom network Args: output_channels (int): Number of output features for this block stride (int, tuple): kernel movement stride downsample (bool): If False, residual branch is a shortcut, Else, residual branch has non-identity layers Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; subnetwork = []; branch_1 = []; branch_1.append(self.convolution(output_channels=output_channels//4, kernel_size=1, stride=stride)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels//4, kernel_size=3, stride=1)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_2 = []; if(downsample): branch_2.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=stride)); branch_2.append(self.batch_normalization()); else: branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.add()); network.append(subnetwork) network.append(self.relu()) return network; def resnet_v2_block(self, output_channels=16, stride=1, downsample=True)-
Append Resnet V2 Block to custom network
Args
output_channels:int- Number of output features for this block
stride:int,tuple- kernel movement stride
downsample:bool- If False, residual branch is a shortcut, Else, residual branch has non-identity layers
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def resnet_v2_block(self, output_channels=16, stride=1, downsample=True): ''' Append Resnet V2 Block to custom network Args: output_channels (int): Number of output features for this block stride (int, tuple): kernel movement stride downsample (bool): If False, residual branch is a shortcut, Else, residual branch has non-identity layers Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; network.append(self.batch_normalization()); network.append(self.relu()); subnetwork = []; branch_1 = []; branch_1.append(self.convolution(output_channels=output_channels, kernel_size=3, stride=stride)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels, kernel_size=3, stride=1)); branch_2 = []; if(downsample): branch_2.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=stride)); else: branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.add()); network.append(subnetwork); network.append(self.relu()); return network; def resnet_v2_bottleneck_block(self, output_channels=16, stride=1, downsample=True)-
Append Resnet V2 Bottleneck Block to custom network
Args
output_channels:int- Number of output features for this block
stride:int,tuple- kernel movement stride
downsample:bool- If False, residual branch is a shortcut, Else, residual branch has non-identity layers
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def resnet_v2_bottleneck_block(self, output_channels=16, stride=1, downsample=True): ''' Append Resnet V2 Bottleneck Block to custom network Args: output_channels (int): Number of output features for this block stride (int, tuple): kernel movement stride downsample (bool): If False, residual branch is a shortcut, Else, residual branch has non-identity layers Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; network.append(self.batch_normalization()); network.append(self.relu()); subnetwork = []; branch_1 = []; branch_1.append(self.convolution(output_channels=output_channels//4, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels//4, kernel_size=3, stride=stride)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1)); branch_2 = []; if(downsample): branch_2.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=stride)); else: branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.add()); network.append(subnetwork) return network; def resnext_block(self, output_channels=256, cardinality=8, bottleneck_width=4, stride=1, downsample=True)-
Append Resnext Block to custom network
Args
output_channels:int- Number of output features for this block
cardinality:int- cardinality dimensions for complex transformations
bottleneck_width:int- Bottleneck dimensions for reducing number of features
stride:int- kernel movement stride
downsample:bool- If False, residual branch is a shortcut, Else, residual branch has non-identity layers
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def resnext_block(self, output_channels=256, cardinality=8, bottleneck_width=4, stride=1, downsample=True): ''' Append Resnext Block to custom network Args: output_channels (int): Number of output features for this block cardinality (int): cardinality dimensions for complex transformations bottleneck_width (int): Bottleneck dimensions for reducing number of features stride (int): kernel movement stride downsample (bool): If False, residual branch is a shortcut, Else, residual branch has non-identity layers Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; channels = output_channels//4; D = int(math.floor(channels * (bottleneck_width / 64))) group_width = cardinality * D subnetwork = []; branch_1 = []; branch_1.append(self.convolution(output_channels=group_width, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=group_width, kernel_size=3, stride=stride)); branch_1.append(self.batch_normalization()); branch_1.append(self.relu()); branch_1.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1)); branch_1.append(self.batch_normalization()); branch_2 = []; if(downsample): branch_2.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=stride)); branch_2.append(self.batch_normalization()); else: branch_2.append(self.identity()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.add()); network.append(subnetwork) network.append(self.relu()); return network; def rrelu(self, lower=0.125, upper=0.333, uid=None)-
Append randomized rectified linear unit activation to custom network
Args
lower:float- Lower limit of uniform distribution for the ramdomized multiplicative factor
upper:float- Upper limit of uniform distribution for the ramdomized multiplicative factor
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def rrelu(self, lower=0.125, upper=0.333, uid=None): ''' Append randomized rectified linear unit activation to custom network Args: lower (float): Lower limit of uniform distribution for the ramdomized multiplicative factor upper (float): Upper limit of uniform distribution for the ramdomized multiplicative factor uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "rrelu"; tmp["params"] = {}; tmp["params"]["lower"] = lower; tmp["params"]["upper"] = upper; return tmp; def selu(self, uid=None)-
Append scaled exponential linear unit activation to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def selu(self, uid=None): ''' Append scaled exponential linear unit activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "selu"; tmp["params"] = {}; return tmp; def separable_convolution_block(self, input_channels=16, output_channels=32, kernel_size=3, stride=1, padding=1)-
Append Separable convolution Block to custom network
Args
input_channels:int- Number of input features for this block
output_channels:int- Number of output features for this block
kernel_size:int- Kernel matrix shape for all layers in this block
stride:int- kernel movement stride
padding:int,tuple- external zero padding on input
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def separable_convolution_block(self, input_channels=16, output_channels=32, kernel_size=3, stride=1, padding=1): ''' Append Separable convolution Block to custom network Args: input_channels (int): Number of input features for this block output_channels (int): Number of output features for this block kernel_size (int): Kernel matrix shape for all layers in this block stride (int): kernel movement stride padding (int, tuple): external zero padding on input Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; network.append(self.convolution(output_channels=input_channels, kernel_size=kernel_size, stride=stride, padding=padding, groups=input_channels)); network.append(self.convolution(output_channels=output_channels, kernel_size=1, stride=1)); return network; def sigmoid(self, uid=None)-
Append sigmoid activation to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def sigmoid(self, uid=None): ''' Append sigmoid activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "sigmoid"; tmp["params"] = {}; return tmp; def softmax(self, uid=None)-
Append softmax activation to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def softmax(self, uid=None): ''' Append softmax activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "softmax"; tmp["params"] = {}; return tmp; def softmin(self, uid=None)-
Append softmin activation to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def softmin(self, uid=None): ''' Append softmin activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "softmin"; tmp["params"] = {}; return tmp; def softplus(self, beta=1, threshold=20, uid=None)-
Append softplus activation to custom network
Args
beta:float- Multiplicative factor
threshold:int- softplus (thresholded relu) limit
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def softplus(self, beta=1, threshold=20, uid=None): ''' Append softplus activation to custom network Args: beta (float): Multiplicative factor threshold (int): softplus (thresholded relu) limit uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "softplus"; tmp["params"] = {}; tmp["params"]["beta"] = beta; tmp["params"]["threshold"] = threshold; return tmp; def softshrink(self, threshold=0.5, uid=None)-
Append soft-shrink activation to custom network
Args
threshold:float- Thresh limit
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def softshrink(self, threshold=0.5, uid=None): ''' Append soft-shrink activation to custom network Args: threshold (float): Thresh limit uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "softshrink"; tmp["params"] = {}; tmp["params"]["threshold"] = threshold; return tmp; def softsign(self, uid=None)-
Append softsign activation to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def softsign(self, uid=None): ''' Append softsign activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "softsign"; tmp["params"] = {}; return tmp; def squeezenet_fire_block(self, squeeze_channels=16, expand_channels_1x1=32, expand_channels_3x3=64)-
Append Squeezenet Fire Block to custom network
Args
squeeze_channels:int- Number of output features for this block
expand_channels_1x1:int- Number of convolution_1x1 features for this block
expand_channels_3x3:int- Number of convolution_3x3 features for this block
bottleneck_width:int- Bottleneck dimensions for reducing number of features
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def squeezenet_fire_block(self, squeeze_channels=16, expand_channels_1x1=32, expand_channels_3x3=64): ''' Append Squeezenet Fire Block to custom network Args: squeeze_channels (int): Number of output features for this block expand_channels_1x1 (int): Number of convolution_1x1 features for this block expand_channels_3x3 (int): Number of convolution_3x3 features for this block bottleneck_width (int): Bottleneck dimensions for reducing number of features Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' network = []; network.append(self.convolution(output_channels=squeeze_channels, kernel_size=1, stride=1)); network.append(self.relu()); subnetwork = []; branch_1 = []; branch_2 = []; branch_1.append(self.convolution(output_channels=expand_channels_1x1, kernel_size=1, stride=1)); branch_1.append(self.relu()); branch_2.append(self.convolution(output_channels=expand_channels_3x3, kernel_size=3, stride=1)); branch_2.append(self.relu()); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.concatenate()); network.append(subnetwork); return network; def subbranch_block(self)-
Append sub-branch Block to custom network
Args
None
Returns
list- Containing all the layer dictionaries arranged as per function arguments
Expand source code
def subbranch_block(self): ''' Append sub-branch Block to custom network Args: None Returns: list: Containing all the layer dictionaries arranged as per function arguments ''' subnetwork = []; branch_1 = []; branch_2 = []; branch_1.append(self.conv_bn_relu_block(output_channels=384, kernel_size=(1, 3), padding=(0, 1))); branch_2.append(self.conv_bn_relu_block(output_channels=384, kernel_size=(3, 1), padding=(1, 0))); subnetwork.append(branch_1); subnetwork.append(branch_2); subnetwork.append(self.concatenate()); return subnetwork; def tanh(self, uid=None)-
Append tanh activation to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def tanh(self, uid=None): ''' Append tanh activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "tanh"; tmp["params"] = {}; return tmp; def tanhshrink(self, uid=None)-
Append Tanh-Shrink activation to custom network
Args
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def tanhshrink(self, uid=None): ''' Append Tanh-Shrink activation to custom network Args: uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "tanhshrink"; tmp["params"] = {}; return tmp; def threshold(self, threshold=1.0, value=0.01, uid=None)-
Append general threshold activation to custom network
Args
threshold:float- thereshold limit value
value:float- replacement value if input is off-limits
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def threshold(self, threshold=1.0, value=0.01, uid=None): ''' Append general threshold activation to custom network Args: threshold (float): thereshold limit value value (float): replacement value if input is off-limits uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp = {}; tmp["uid"] = uid; tmp["name"] = "threshold"; tmp["params"] = {}; tmp["params"]["threshold"] = threshold; tmp["params"]["value"] = value; return tmp; def transposed_convolution(self, output_channels=3, kernel_size=3, stride=1, padding='in_eq_out', output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None)-
Append 2d-transposed-convolution to custom network
Args
output_channels:int- Number of output features for this layers
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
output_padding:int- Additional padding applied to output
groups:int- Number of groups for grouped convolution
dilation:int- Factor for dilated convolution
use_bias:bool- If True, learnable bias is added
layout:str- Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def transposed_convolution(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None): ''' Append 2d-transposed-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding output_padding (int): Additional padding applied to output groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "transposed_convolution2d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["output_padding"] = output_padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp; def transposed_convolution1d(self, output_channels=3, kernel_size=3, stride=1, padding='in_eq_out', output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCW', uid=None)-
Append 1d-transposed-convolution to custom network
Args
output_channels:int- Number of output features for this layers
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
output_padding:int- Additional padding applied to output
groups:int- Number of groups for grouped convolution
dilation:int- Factor for dilated convolution
use_bias:bool- If True, learnable bias is added
layout:str- Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def transposed_convolution1d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCW', uid=None): ''' Append 1d-transposed-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding output_padding (int): Additional padding applied to output groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCW' - order 2) 'NWC' - order - N: Number of elements in batches - C: Number of channels - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "transposed_convolution1d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["output_padding"] = output_padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp; def transposed_convolution2d(self, output_channels=3, kernel_size=3, stride=1, padding='in_eq_out', output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None)-
Append 2d-transposed-convolution to custom network
Args
output_channels:int- Number of output features for this layers
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
output_padding:int- Additional padding applied to output
groups:int- Number of groups for grouped convolution
dilation:int- Factor for dilated convolution
use_bias:bool- If True, learnable bias is added
layout:str- Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def transposed_convolution2d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCHW', uid=None): ''' Append 2d-transposed-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding output_padding (int): Additional padding applied to output groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCHW' - Order 2) 'NHWC' - Order - N: Number of elements in batches - C: Number of channels - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "transposed_convolution2d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["output_padding"] = output_padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp; def transposed_convolution3d(self, output_channels=3, kernel_size=3, stride=1, padding='in_eq_out', output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCDHW', uid=None)-
Append 3d-transposed-convolution to custom network
Args
output_channels:int- Number of output features for this layers
kernel_size:int,tuple- kernel matrix size
stride:int,tuple- kernel movement stride
padding:int,tuple,str- Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding
output_padding:int- Additional padding applied to output
groups:int- Number of groups for grouped convolution
dilation:int- Factor for dilated convolution
use_bias:bool- If True, learnable bias is added
layout:str- Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers
uid:str- Unique name for layer, if not mentioned then dynamically assigned
Returns
dict- Containing all the parameters set as per function arguments
Expand source code
def transposed_convolution3d(self, output_channels=3, kernel_size=3, stride=1, padding="in_eq_out", output_padding=0, groups=1, dilation=1, use_bias=True, layout='NCDHW', uid=None): ''' Append 3d-transposed-convolution to custom network Args: output_channels (int): Number of output features for this layers kernel_size (int, tuple): kernel matrix size stride (int, tuple): kernel movement stride padding (int, tuple, str): Zero padding applied to input 1) "in_eq_out": Automated padding applied to keep output shape same as input 2) integer or tuple value: Manually add padding output_padding (int): Additional padding applied to output groups (int): Number of groups for grouped convolution dilation (int): Factor for dilated convolution use_bias (bool): If True, learnable bias is added layout (str): Either of these values (order) 1) 'NCDHW' - Order 2) 'NDHWC' - Order - N: Number of elements in batches - C: Number of channels - D: Depth of features in layers - H: Height of features in layers - W: Number of features in layers uid (str): Unique name for layer, if not mentioned then dynamically assigned Returns: dict: Containing all the parameters set as per function arguments ''' tmp={}; tmp["uid"] = uid; tmp["name"] = "transposed_convolution3d"; tmp["params"] = {}; tmp["params"]["output_channels"] = output_channels; tmp["params"]["kernel_size"] = kernel_size; tmp["params"]["stride"] = stride; tmp["params"]["padding"] = padding; tmp["params"]["output_padding"] = output_padding; tmp["params"]["groups"] = groups; tmp["params"]["dilation"] = dilation; tmp["params"]["use_bias"] = use_bias; tmp["params"]["layout"] = layout; return tmp;