def create_detection_losses(cls_score, label_targets, bbox_pred, rois, bbox_targets, bbox_inside_weights, cfg):
# The losses are normalized by the batch size
# classification loss
p_cls_score = placeholder()
p_label_targets = placeholder()
cls_loss = cross_entropy_with_softmax(p_cls_score, p_label_targets, axis=1)
cls_normalization_factor = 1.0 / cfg.NUM_ROI_PROPOSALS
normalized_cls_loss = reduce_sum(cls_loss) * cls_normalization_factor
reduced_cls_loss = cntk.as_block(normalized_cls_loss,
[(p_cls_score, cls_score), (p_label_targets, label_targets)],
'CrossEntropyWithSoftmax', 'norm_cls_loss')
# regression loss
p_bbox_pred = placeholder()
p_bbox_targets = placeholder()
p_bbox_inside_weights = placeholder()
bbox_loss = SmoothL1Loss(cfg.SIGMA_DET_L1, p_bbox_pred, p_bbox_targets, p_bbox_inside_weights, 1.0)
bbox_normalization_factor = 1.0 / cfg.NUM_ROI_PROPOSALS
normalized_bbox_loss = reduce_sum(bbox_loss) * bbox_normalization_factor
reduced_bbox_loss = cntk.as_block(normalized_bbox_loss,
[(p_bbox_pred, bbox_pred), (p_bbox_targets, bbox_targets),
(p_bbox_inside_weights, bbox_inside_weights)],
'SmoothL1Loss', 'norm_bbox_loss')
detection_losses = plus(reduced_cls_loss, reduced_bbox_loss, name="detection_losses")
return detection_losses
def create_detection_losses(cls_score, label_targets, rois, bbox_pred, bbox_targets, bbox_inside_weights):
# classification loss
cls_loss = cross_entropy_with_softmax(cls_score, label_targets, axis=1)
p_cls_loss = placeholder()
p_rois = placeholder()
# The terms that are accounted for in the cls loss are those that correspond to an actual roi proposal --> do not count no-op (all-zero) rois
roi_indicator = reduce_sum(p_rois, axis=1)
cls_num_terms = reduce_sum(cntk.greater_equal(roi_indicator, 0.0))
cls_normalization_factor = 1.0 / cls_num_terms
normalized_cls_loss = reduce_sum(p_cls_loss) * cls_normalization_factor
reduced_cls_loss = cntk.as_block(normalized_cls_loss,
[(p_cls_loss, cls_loss), (p_rois, rois)],
'Normalize', 'norm_cls_loss')
# regression loss
p_bbox_pred = placeholder()
p_bbox_targets = placeholder()
p_bbox_inside_weights = placeholder()
bbox_loss = SmoothL1Loss(cfg["CNTK"].SIGMA_DET_L1, p_bbox_pred, p_bbox_targets, p_bbox_inside_weights, 1.0)
# The bbox loss is normalized by the batch size
bbox_normalization_factor = 1.0 / cfg["TRAIN"].BATCH_SIZE
normalized_bbox_loss = reduce_sum(bbox_loss) * bbox_normalization_factor
reduced_bbox_loss = cntk.as_block(normalized_bbox_loss,
[(p_bbox_pred, bbox_pred), (p_bbox_targets, bbox_targets), (p_bbox_inside_weights, bbox_inside_weights)],
'SmoothL1Loss', 'norm_bbox_loss')
detection_losses = plus(reduced_cls_loss, reduced_bbox_loss, name="detection_losses")
return detection_losses
def create_rpn(conv_out,
scaled_gt_boxes,
im_info,
cfg,
add_loss_functions=True):
'''
Creates a region proposal network for object detection as proposed in the "Faster R-CNN" paper:
Shaoqing Ren and Kaiming He and Ross Girshick and Jian Sun:
"Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks"
Outputs object detection proposals by applying estimated bounding-box
transformations to a set of regular boxes (called "anchors").
Args:
conv_out: The convolutional feature map, i.e. the output of the conv layers from the pretrained classification network
scaled_gt_boxes: The ground truth boxes as (x1, y1, x2, y2, label). Coordinates are absolute pixels wrt. the input image.
im_info: A CNTK variable or constant containing
(pad_width, pad_height, scaled_image_width, scaled_image_height, orig_img_width, orig_img_height)
e.g. (1000, 1000, 1000, 600, 500, 300) for an original image of 600x300 that is scaled and padded to 1000x1000
cfg: The configuration dictionary
add_loss_functions: If set to True rpn_losses will be returned, otherwise None is returned for the losses
Returns:
rpn_rois - the proposed ROIs
rpn_losses - the losses (SmoothL1 loss for bbox regression plus cross entropy for objectness)
'''
# RPN network
# init = 'normal', initValueScale = 0.01, initBias = 0.1
num_channels = cfg["MODEL"].RPN_NUM_CHANNELS
rpn_conv_3x3 = Convolution((3, 3),
num_channels,
activation=relu,
pad=True,
strides=1,
init=normal(scale=0.01),
init_bias=0.0)(conv_out)
rpn_cls_score = Convolution(
(1, 1),
18,
activation=None,
name="rpn_cls_score",
init=normal(scale=0.01),
init_bias=0.0)(rpn_conv_3x3) # 2(bg/fg) * 9(anchors)
rpn_bbox_pred = Convolution(
(1, 1),
36,
activation=None,
name="rpn_bbox_pred",
init=normal(scale=0.01),
init_bias=0.0)(rpn_conv_3x3) # 4(coords) * 9(anchors)
# apply softmax to get (bg, fg) probabilities and reshape predictions back to grid of (18, H, W)
num_predictions = int(rpn_cls_score.shape[0] / 2)
rpn_cls_score_rshp = reshape(
rpn_cls_score,
(2, num_predictions, rpn_cls_score.shape[1], rpn_cls_score.shape[2]),
name="rpn_cls_score_rshp")
p_rpn_cls_score_rshp = cntk.placeholder()
rpn_cls_sm = softmax(p_rpn_cls_score_rshp, axis=0)
rpn_cls_prob = cntk.as_block(rpn_cls_sm,
[(p_rpn_cls_score_rshp, rpn_cls_score_rshp)],
'Softmax', 'rpn_cls_prob')
rpn_cls_prob_reshape = reshape(rpn_cls_prob,
rpn_cls_score.shape,
name="rpn_cls_prob_reshape")
# proposal layer
rpn_rois = create_proposal_layer(rpn_cls_prob_reshape, rpn_bbox_pred,
im_info, cfg)
rpn_losses = None
if (add_loss_functions):
# RPN targets
# Comment: rpn_cls_score is only passed vvv to get width and height of the conv feature map ...
proposal_layer_params = "'feat_stride': {}\n'scales':\n - {}". \
format(cfg["MODEL"].FEATURE_STRIDE, "\n - ".join([str(v) for v in cfg["DATA"].PROPOSAL_LAYER_SCALES]))
atl = user_function(
AnchorTargetLayer(
rpn_cls_score,
scaled_gt_boxes,
im_info,
rpn_batch_size=cfg["TRAIN"].RPN_BATCHSIZE,
rpn_fg_fraction=cfg["TRAIN"].RPN_FG_FRACTION,
clobber_positives=cfg["TRAIN"].RPN_CLOBBER_POSITIVES,
positive_overlap=cfg["TRAIN"].RPN_POSITIVE_OVERLAP,
negative_overlap=cfg["TRAIN"].RPN_NEGATIVE_OVERLAP,
param_str=proposal_layer_params))
rpn_labels = atl.outputs[0]
rpn_bbox_targets = atl.outputs[1]
rpn_bbox_inside_weights = atl.outputs[2]
# classification loss
p_rpn_labels = cntk.placeholder()
p_rpn_cls_score_rshp = cntk.placeholder()
keeps = cntk.greater_equal(p_rpn_labels, 0.0)
fg_labels = element_times(p_rpn_labels, keeps, name="fg_targets")
bg_labels = minus(1, fg_labels, name="bg_targets")
rpn_labels_ignore = splice(bg_labels, fg_labels, axis=0)
rpn_ce = cross_entropy_with_softmax(p_rpn_cls_score_rshp,
rpn_labels_ignore,
axis=0)
rpn_loss_cls = element_times(rpn_ce, keeps)
# The terms that are accounted for in the cls loss are those that have a label >= 0
cls_num_terms = reduce_sum(keeps)
cls_normalization_factor = 1.0 / cls_num_terms
normalized_rpn_cls_loss = reduce_sum(
rpn_loss_cls) * cls_normalization_factor
reduced_rpn_loss_cls = cntk.as_block(
normalized_rpn_cls_loss,
[(p_rpn_labels, rpn_labels),
(p_rpn_cls_score_rshp, rpn_cls_score_rshp)], 'CE_with_ignore',
'norm_rpn_cls_loss')
# regression loss
p_rpn_bbox_pred = cntk.placeholder()
p_rpn_bbox_targets = cntk.placeholder()
p_rpn_bbox_inside_weights = cntk.placeholder()
rpn_loss_bbox = SmoothL1Loss(cfg.SIGMA_RPN_L1, p_rpn_bbox_pred,
p_rpn_bbox_targets,
p_rpn_bbox_inside_weights, 1.0)
# The bbox loss is normalized by the rpn batch size
bbox_normalization_factor = 1.0 / cfg["TRAIN"].RPN_BATCHSIZE
normalized_rpn_bbox_loss = reduce_sum(
rpn_loss_bbox) * bbox_normalization_factor
reduced_rpn_loss_bbox = cntk.as_block(
normalized_rpn_bbox_loss,
[(p_rpn_bbox_pred, rpn_bbox_pred),
(p_rpn_bbox_targets, rpn_bbox_targets),
(p_rpn_bbox_inside_weights, rpn_bbox_inside_weights)],
'SmoothL1Loss', 'norm_rpn_bbox_loss')
rpn_losses = plus(reduced_rpn_loss_cls,
reduced_rpn_loss_bbox,
name="rpn_losses")
return rpn_rois, rpn_losses
def create_rpn(conv_out,
scaled_gt_boxes,
im_info,
add_loss_functions=True,
proposal_layer_param_string=None):
'''
Creates a region proposal network for object detection as proposed in the "Faster R-CNN" paper:
Shaoqing Ren and Kaiming He and Ross Girshick and Jian Sun:
"Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks"
Outputs object detection proposals by applying estimated bounding-box
transformations to a set of regular boxes (called "anchors").
Args:
conv_out: The convolutional feature map, i.e. the output of the conv layers from the pretrained classification network
scaled_gt_boxes: The ground truth boxes as (x1, y1, x2, y2, label). Coordinates are absolute pixels wrt. the input image.
im_info: (image_widht, image_height, image_scale) as CNTK variable or constant
add_loss_functions: If set to True rpn_losses will be returned, otherwise None is returned for the losses
proposal_layer_param_string: A yaml parameter string that is passed to the proposal layer.
Returns:
rpn_rois - the proposed ROIs
rpn_losses - the losses (SmoothL1 loss for bbox regression plus cross entropy for objectness)
'''
# RPN network
# init = 'normal', initValueScale = 0.01, initBias = 0.1
rpn_conv_3x3 = Convolution((3, 3),
256,
activation=relu,
pad=True,
strides=1,
init=normal(scale=0.01),
init_bias=0.1)(conv_out)
rpn_cls_score = Convolution(
(1, 1),
18,
activation=None,
name="rpn_cls_score",
init=normal(scale=0.01),
init_bias=0.1)(rpn_conv_3x3) # 2(bg/fg) * 9(anchors)
rpn_bbox_pred = Convolution(
(1, 1),
36,
activation=None,
name="rpn_bbox_pred",
init=normal(scale=0.01),
init_bias=0.1)(rpn_conv_3x3) # 4(coords) * 9(anchors)
# apply softmax to get (bg, fg) probabilities and reshape predictions back to grid of (18, H, W)
num_predictions = int(np.prod(rpn_cls_score.shape) / 2)
rpn_cls_score_rshp = reshape(rpn_cls_score, (2, num_predictions))
rpn_cls_prob = softmax(rpn_cls_score_rshp, axis=0, name="objness_softmax")
rpn_cls_prob_reshape = reshape(rpn_cls_prob, rpn_cls_score.shape)
# proposal layer
rpn_rois_raw = user_function(
ProposalLayer(rpn_cls_prob_reshape,
rpn_bbox_pred,
im_info,
param_str=proposal_layer_param_string))
rpn_rois = alias(rpn_rois_raw, name='rpn_rois')
rpn_losses = None
if (add_loss_functions):
# RPN targets
# Comment: rpn_cls_score is only passed vvv to get width and height of the conv feature map ...
atl = user_function(
AnchorTargetLayer(rpn_cls_score,
scaled_gt_boxes,
im_info,
param_str=proposal_layer_param_string))
rpn_labels = atl.outputs[0]
rpn_bbox_targets = atl.outputs[1]
rpn_bbox_inside_weights = atl.outputs[2]
# For loss functions: ignore label predictions for the 'ignore label',
# i.e. set target and prediction to 0 --> needs to be softmaxed before
rpn_labels_rshp = reshape(rpn_labels, (1, num_predictions))
ignore = user_function(
IgnoreLabel(rpn_cls_prob, rpn_labels_rshp, ignore_label=-1))
rpn_cls_prob_ignore = ignore.outputs[0]
fg_targets = ignore.outputs[1]
bg_targets = 1 - fg_targets
rpn_labels_ignore = splice(bg_targets, fg_targets, axis=0)
# RPN losses
rpn_loss_cls = cross_entropy_with_softmax(rpn_cls_prob_ignore,
rpn_labels_ignore,
axis=0)
rpn_loss_bbox = user_function(
SmoothL1Loss(rpn_bbox_pred, rpn_bbox_targets,
rpn_bbox_inside_weights))
rpn_losses = plus(reduce_sum(rpn_loss_cls),
reduce_sum(rpn_loss_bbox),
name="rpn_losses")
return rpn_rois, rpn_losses