Module 3_mxrcnn.lib.mx-rcnn.symnet.symbol_vgg

Functions

def get_vgg_feature(data)

Expand source code

def get_vgg_feature(data):
    # group 1
    conv1_1 = mx.symbol.Convolution(
        data=data, kernel=(3, 3), pad=(1, 1), num_filter=64, workspace=2048, name="conv1_1")
    relu1_1 = mx.symbol.Activation(data=conv1_1, act_type="relu", name="relu1_1")
    conv1_2 = mx.symbol.Convolution(
        data=relu1_1, kernel=(3, 3), pad=(1, 1), num_filter=64, workspace=2048, name="conv1_2")
    relu1_2 = mx.symbol.Activation(data=conv1_2, act_type="relu", name="relu1_2")
    pool1 = mx.symbol.Pooling(
        data=relu1_2, pool_type="max", kernel=(2, 2), stride=(2, 2), name="pool1")
    # group 2
    conv2_1 = mx.symbol.Convolution(
        data=pool1, kernel=(3, 3), pad=(1, 1), num_filter=128, workspace=2048, name="conv2_1")
    relu2_1 = mx.symbol.Activation(data=conv2_1, act_type="relu", name="relu2_1")
    conv2_2 = mx.symbol.Convolution(
        data=relu2_1, kernel=(3, 3), pad=(1, 1), num_filter=128, workspace=2048, name="conv2_2")
    relu2_2 = mx.symbol.Activation(data=conv2_2, act_type="relu", name="relu2_2")
    pool2 = mx.symbol.Pooling(
        data=relu2_2, pool_type="max", kernel=(2, 2), stride=(2, 2), name="pool2")
    # group 3
    conv3_1 = mx.symbol.Convolution(
        data=pool2, kernel=(3, 3), pad=(1, 1), num_filter=256, workspace=2048, name="conv3_1")
    relu3_1 = mx.symbol.Activation(data=conv3_1, act_type="relu", name="relu3_1")
    conv3_2 = mx.symbol.Convolution(
        data=relu3_1, kernel=(3, 3), pad=(1, 1), num_filter=256, workspace=2048, name="conv3_2")
    relu3_2 = mx.symbol.Activation(data=conv3_2, act_type="relu", name="relu3_2")
    conv3_3 = mx.symbol.Convolution(
        data=relu3_2, kernel=(3, 3), pad=(1, 1), num_filter=256, workspace=2048, name="conv3_3")
    relu3_3 = mx.symbol.Activation(data=conv3_3, act_type="relu", name="relu3_3")
    pool3 = mx.symbol.Pooling(
        data=relu3_3, pool_type="max", kernel=(2, 2), stride=(2, 2), name="pool3")
    # group 4
    conv4_1 = mx.symbol.Convolution(
        data=pool3, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv4_1")
    relu4_1 = mx.symbol.Activation(data=conv4_1, act_type="relu", name="relu4_1")
    conv4_2 = mx.symbol.Convolution(
        data=relu4_1, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv4_2")
    relu4_2 = mx.symbol.Activation(data=conv4_2, act_type="relu", name="relu4_2")
    conv4_3 = mx.symbol.Convolution(
        data=relu4_2, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv4_3")
    relu4_3 = mx.symbol.Activation(data=conv4_3, act_type="relu", name="relu4_3")
    pool4 = mx.symbol.Pooling(
        data=relu4_3, pool_type="max", kernel=(2, 2), stride=(2, 2), name="pool4")
    # group 5
    conv5_1 = mx.symbol.Convolution(
        data=pool4, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv5_1")
    relu5_1 = mx.symbol.Activation(data=conv5_1, act_type="relu", name="relu5_1")
    conv5_2 = mx.symbol.Convolution(
        data=relu5_1, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv5_2")
    relu5_2 = mx.symbol.Activation(data=conv5_2, act_type="relu", name="relu5_2")
    conv5_3 = mx.symbol.Convolution(
        data=relu5_2, kernel=(3, 3), pad=(1, 1), num_filter=512, workspace=2048, name="conv5_3")
    relu5_3 = mx.symbol.Activation(data=conv5_3, act_type="relu", name="relu5_3")

    return relu5_3

def get_vgg_test(anchor_scales, anchor_ratios, rpn_feature_stride, rpn_pre_topk, rpn_post_topk, rpn_nms_thresh, rpn_min_size, num_classes, rcnn_feature_stride, rcnn_pooled_size, rcnn_batch_size)

Expand source code

def get_vgg_test(anchor_scales, anchor_ratios, rpn_feature_stride,
                 rpn_pre_topk, rpn_post_topk, rpn_nms_thresh, rpn_min_size,
                 num_classes, rcnn_feature_stride, rcnn_pooled_size, rcnn_batch_size):
    num_anchors = len(anchor_scales) * len(anchor_ratios)

    data = mx.symbol.Variable(name="data")
    im_info = mx.symbol.Variable(name="im_info")

    # shared convolutional layers
    conv_feat = get_vgg_feature(data)

    # rpn feature
    rpn_conv = mx.symbol.Convolution(
        data=conv_feat, kernel=(3, 3), pad=(1, 1), num_filter=512, name="rpn_conv_3x3")
    rpn_relu = mx.symbol.Activation(data=rpn_conv, act_type="relu", name="rpn_relu")

    # rpn classification
    rpn_cls_score = mx.symbol.Convolution(
        data=rpn_relu, kernel=(1, 1), pad=(0, 0), num_filter=2 * num_anchors, name="rpn_cls_score")
    rpn_cls_score_reshape = mx.symbol.Reshape(
        data=rpn_cls_score, shape=(0, 2, -1, 0), name="rpn_cls_score_reshape")
    rpn_cls_act = mx.symbol.softmax(
        data=rpn_cls_score_reshape, axis=1, name="rpn_cls_act")
    rpn_cls_act_reshape = mx.symbol.Reshape(
        data=rpn_cls_act, shape=(0, 2 * num_anchors, -1, 0), name='rpn_cls_act_reshape')

    # rpn bbox regression
    rpn_bbox_pred = mx.symbol.Convolution(
        data=rpn_relu, kernel=(1, 1), pad=(0, 0), num_filter=4 * num_anchors, name="rpn_bbox_pred")

    # rpn proposal
    rois = mx.symbol.contrib.MultiProposal(
        cls_prob=rpn_cls_act_reshape, bbox_pred=rpn_bbox_pred, im_info=im_info, name='rois',
        feature_stride=rpn_feature_stride, scales=anchor_scales, ratios=anchor_ratios,
        rpn_pre_nms_top_n=rpn_pre_topk, rpn_post_nms_top_n=rpn_post_topk,
        threshold=rpn_nms_thresh, rpn_min_size=rpn_min_size)

    # rcnn roi pool
    roi_pool = mx.symbol.ROIPooling(
        name='roi_pool', data=conv_feat, rois=rois, pooled_size=rcnn_pooled_size, spatial_scale=1.0 / rcnn_feature_stride)

    # rcnn top feature
    top_feat = get_vgg_top_feature(roi_pool)

    # rcnn classification
    cls_score = mx.symbol.FullyConnected(name='cls_score', data=top_feat, num_hidden=num_classes)
    cls_prob = mx.symbol.softmax(name='cls_prob', data=cls_score)

    # rcnn bbox regression
    bbox_pred = mx.symbol.FullyConnected(name='bbox_pred', data=top_feat, num_hidden=num_classes * 4)

    # reshape output
    cls_prob = mx.symbol.Reshape(data=cls_prob, shape=(rcnn_batch_size, -1, num_classes), name='cls_prob_reshape')
    bbox_pred = mx.symbol.Reshape(data=bbox_pred, shape=(rcnn_batch_size, -1, 4 * num_classes), name='bbox_pred_reshape')

    # group output
    group = mx.symbol.Group([rois, cls_prob, bbox_pred])
    return group

def get_vgg_top_feature(data)

Expand source code

def get_vgg_top_feature(data):
    # group 6
    flatten = mx.symbol.Flatten(data=data, name="flatten")
    fc6 = mx.symbol.FullyConnected(data=flatten, num_hidden=4096, name="fc6")
    relu6 = mx.symbol.Activation(data=fc6, act_type="relu", name="relu6")
    drop6 = mx.symbol.Dropout(data=relu6, p=0.5, name="drop6")
    # group 7
    fc7 = mx.symbol.FullyConnected(data=drop6, num_hidden=4096, name="fc7")
    relu7 = mx.symbol.Activation(data=fc7, act_type="relu", name="relu7")
    drop7 = mx.symbol.Dropout(data=relu7, p=0.5, name="drop7")
    return drop7

def get_vgg_train(anchor_scales, anchor_ratios, rpn_feature_stride, rpn_pre_topk, rpn_post_topk, rpn_nms_thresh, rpn_min_size, rpn_batch_rois, num_classes, rcnn_feature_stride, rcnn_pooled_size, rcnn_batch_size, rcnn_batch_rois, rcnn_fg_fraction, rcnn_fg_overlap, rcnn_bbox_stds)

Expand source code

def get_vgg_train(anchor_scales, anchor_ratios, rpn_feature_stride,
                  rpn_pre_topk, rpn_post_topk, rpn_nms_thresh, rpn_min_size, rpn_batch_rois,
                  num_classes, rcnn_feature_stride, rcnn_pooled_size, rcnn_batch_size,
                  rcnn_batch_rois, rcnn_fg_fraction, rcnn_fg_overlap, rcnn_bbox_stds):
    num_anchors = len(anchor_scales) * len(anchor_ratios)

    data = mx.symbol.Variable(name="data")
    im_info = mx.symbol.Variable(name="im_info")
    gt_boxes = mx.symbol.Variable(name="gt_boxes")
    rpn_label = mx.symbol.Variable(name='label')
    rpn_bbox_target = mx.symbol.Variable(name='bbox_target')
    rpn_bbox_weight = mx.symbol.Variable(name='bbox_weight')

    # shared convolutional layers
    conv_feat = get_vgg_feature(data)

    # RPN layers
    rpn_conv = mx.symbol.Convolution(
        data=conv_feat, kernel=(3, 3), pad=(1, 1), num_filter=512, name="rpn_conv_3x3")
    rpn_relu = mx.symbol.Activation(data=rpn_conv, act_type="relu", name="rpn_relu")

    # rpn classification
    rpn_cls_score = mx.symbol.Convolution(
        data=rpn_relu, kernel=(1, 1), pad=(0, 0), num_filter=2 * num_anchors, name="rpn_cls_score")
    rpn_cls_score_reshape = mx.symbol.Reshape(
        data=rpn_cls_score, shape=(0, 2, -1, 0), name="rpn_cls_score_reshape")
    rpn_cls_prob = mx.symbol.SoftmaxOutput(data=rpn_cls_score_reshape, label=rpn_label, multi_output=True,
                                           normalization='valid', use_ignore=True, ignore_label=-1, name="rpn_cls_prob")
    rpn_cls_act = mx.symbol.softmax(
        data=rpn_cls_score_reshape, axis=1, name="rpn_cls_act")
    rpn_cls_act_reshape = mx.symbol.Reshape(
        data=rpn_cls_act, shape=(0, 2 * num_anchors, -1, 0), name='rpn_cls_act_reshape')

    # rpn bbox regression
    rpn_bbox_pred = mx.symbol.Convolution(
        data=rpn_relu, kernel=(1, 1), pad=(0, 0), num_filter=4 * num_anchors, name="rpn_bbox_pred")
    rpn_bbox_loss_ = rpn_bbox_weight * mx.symbol.smooth_l1(name='rpn_bbox_loss_', scalar=3.0, data=(rpn_bbox_pred - rpn_bbox_target))
    rpn_bbox_loss = mx.sym.MakeLoss(name='rpn_bbox_loss', data=rpn_bbox_loss_, grad_scale=1.0 / rpn_batch_rois)

    # rpn proposal
    rois = mx.symbol.contrib.MultiProposal(
        cls_prob=rpn_cls_act_reshape, bbox_pred=rpn_bbox_pred, im_info=im_info, name='rois',
        feature_stride=rpn_feature_stride, scales=anchor_scales, ratios=anchor_ratios,
        rpn_pre_nms_top_n=rpn_pre_topk, rpn_post_nms_top_n=rpn_post_topk,
        threshold=rpn_nms_thresh, rpn_min_size=rpn_min_size)

    # rcnn roi proposal target
    group = mx.symbol.Custom(rois=rois, gt_boxes=gt_boxes, op_type='proposal_target',
                             num_classes=num_classes, batch_images=rcnn_batch_size,
                             batch_rois=rcnn_batch_rois, fg_fraction=rcnn_fg_fraction,
                             fg_overlap=rcnn_fg_overlap, box_stds=rcnn_bbox_stds)
    rois = group[0]
    label = group[1]
    bbox_target = group[2]
    bbox_weight = group[3]

    # rcnn roi pool
    roi_pool = mx.symbol.ROIPooling(
        name='roi_pool', data=conv_feat, rois=rois, pooled_size=rcnn_pooled_size, spatial_scale=1.0 / rcnn_feature_stride)

    # rcnn top feature
    top_feat = get_vgg_top_feature(roi_pool)

    # rcnn classification
    cls_score = mx.symbol.FullyConnected(name='cls_score', data=top_feat, num_hidden=num_classes)
    cls_prob = mx.symbol.SoftmaxOutput(name='cls_prob', data=cls_score, label=label, normalization='batch')

    # rcnn bbox regression
    bbox_pred = mx.symbol.FullyConnected(name='bbox_pred', data=top_feat, num_hidden=num_classes * 4)
    bbox_loss_ = bbox_weight * mx.symbol.smooth_l1(name='bbox_loss_', scalar=1.0, data=(bbox_pred - bbox_target))
    bbox_loss = mx.sym.MakeLoss(name='bbox_loss', data=bbox_loss_, grad_scale=1.0 / rcnn_batch_rois)

    # reshape output
    label = mx.symbol.Reshape(data=label, shape=(rcnn_batch_size, -1), name='label_reshape')
    cls_prob = mx.symbol.Reshape(data=cls_prob, shape=(rcnn_batch_size, -1, num_classes), name='cls_prob_reshape')
    bbox_loss = mx.symbol.Reshape(data=bbox_loss, shape=(rcnn_batch_size, -1, 4 * num_classes), name='bbox_loss_reshape')

    # group output
    group = mx.symbol.Group([rpn_cls_prob, rpn_bbox_loss, cls_prob, bbox_loss, mx.symbol.BlockGrad(label)])
    return group