def rpn(self, image, features, inputs):
'''
1.对features 加上rpn head,提取出proposal box,proposal score,用inputs这个标注
和proposal box,proposal score计算box_loss,label_loss
2.根据image的尺寸,以及上述proposal box,proposal score生成proposal region(绝对坐标),
返回
BoxProposals:封装了选择的(proposal_boxes)
losses:[label_loss,box_loss]
:param image: (1,None,None,3)
:param features:[(1,None,None,1024)] for resnetC4
:param inputs: (None,None,A,2)anchor_label,(None,None,A,4)anchor_boxes,
:return:
'''
featuremap = features[0]
rpn_label_logits, rpn_box_logits = rpn_head('rpn', featuremap, cfg.RPN.HEAD_DIM, cfg.RPN.NUM_ANCHOR)
anchors = RPNAnchors(get_all_anchors(), inputs['anchor_labels'], inputs['anchor_boxes'])
anchors = anchors.narrow_to(featuremap)
image_shape2d = tf.shape(image)[2:] # h,w
pred_boxes_decoded = anchors.decode_logits(rpn_box_logits) # fHxfWxNAx4, floatbox
proposal_boxes, proposal_scores = generate_rpn_proposals(
tf.reshape(pred_boxes_decoded, [-1, 4]),
tf.reshape(rpn_label_logits, [-1]),
image_shape2d,
cfg.RPN.TRAIN_PRE_NMS_TOPK if self.training else cfg.RPN.TEST_PRE_NMS_TOPK,
cfg.RPN.TRAIN_POST_NMS_TOPK if self.training else cfg.RPN.TEST_POST_NMS_TOPK)
if self.training:
losses = rpn_losses(
anchors.gt_labels, anchors.encoded_gt_boxes(), rpn_label_logits, rpn_box_logits)
else:
losses = []
return BoxProposals(proposal_boxes), losses
def rpn(self, image, features, inputs):
featuremap = features[0]
rpn_label_logits, rpn_box_logits = rpn_head('rpn', featuremap,
cfg.RPN.HEAD_DIM,
cfg.RPN.NUM_ANCHOR)
anchors = RPNAnchors(get_all_anchors(), inputs['anchor_labels'],
inputs['anchor_boxes'])
anchors = anchors.narrow_to(featuremap)
image_shape2d = tf.shape(image)[2:] # h,w
pred_boxes_decoded = anchors.decode_logits(
rpn_box_logits) # fHxfWxNAx4, floatbox
proposal_boxes, proposal_scores = generate_rpn_proposals(
tf.reshape(pred_boxes_decoded, [-1, 4]),
tf.reshape(rpn_label_logits,
[-1]), image_shape2d, cfg.RPN.TRAIN_PRE_NMS_TOPK
if self.training else cfg.RPN.TEST_PRE_NMS_TOPK,
cfg.RPN.TRAIN_POST_NMS_TOPK
if self.training else cfg.RPN.TEST_POST_NMS_TOPK)
if self.training:
losses = rpn_losses(anchors.gt_labels, anchors.encoded_gt_boxes(),
rpn_label_logits, rpn_box_logits)
else:
losses = []
return BoxProposals(proposal_boxes), losses
def rpn(self, image, features, inputs):
assert len(cfg.RPN.ANCHOR_SIZES) == len(cfg.FPN.ANCHOR_STRIDES)
image_shape2d = tf.shape(image)[2:] # h,w
all_anchors_fpn = get_all_anchors_fpn()
multilevel_anchors = [RPNAnchors(
all_anchors_fpn[i],
inputs['anchor_labels_lvl{}'.format(i + 2)],
inputs['anchor_boxes_lvl{}'.format(i + 2)]) for i in range(len(all_anchors_fpn))]
self.slice_feature_and_anchors(features, multilevel_anchors)
# Multi-Level RPN Proposals
rpn_outputs = [rpn_head('rpn', pi, cfg.FPN.NUM_CHANNEL, len(cfg.RPN.ANCHOR_RATIOS))
for pi in features]
multilevel_label_logits = [k[0] for k in rpn_outputs]
multilevel_box_logits = [k[1] for k in rpn_outputs]
multilevel_pred_boxes = [anchor.decode_logits(logits)
for anchor, logits in zip(multilevel_anchors, multilevel_box_logits)]
proposal_boxes, proposal_scores = generate_fpn_proposals(
multilevel_pred_boxes, multilevel_label_logits, image_shape2d)
if self.training:
losses = multilevel_rpn_losses(
multilevel_anchors, multilevel_label_logits, multilevel_box_logits)
else:
losses = []
return BoxProposals(proposal_boxes), losses
def build_graph(self, *inputs):
inputs = dict(zip(self.input_names, inputs))
num_fpn_level = len(cfg.FPN.ANCHOR_STRIDES)
assert len(cfg.RPN.ANCHOR_SIZES) == num_fpn_level
is_training = get_current_tower_context().is_training
all_anchors_fpn = get_all_anchors_fpn()
multilevel_anchors = [
RPNAnchors(all_anchors_fpn[i],
inputs['anchor_labels_lvl{}'.format(i + 2)],
inputs['anchor_boxes_lvl{}'.format(i + 2)])
for i in range(len(all_anchors_fpn))
]
image = self.preprocess(inputs['image']) # 1CHW
image_shape2d = tf.shape(image)[2:] # h,w
c2345 = resnet_fpn_backbone(image, cfg.BACKBONE.RESNET_NUM_BLOCK)
p23456 = fpn_model('fpn', c2345)
self.slice_feature_and_anchors(image_shape2d, p23456,
multilevel_anchors)
# Multi-Level RPN Proposals
rpn_outputs = [
rpn_head('rpn', pi, cfg.FPN.NUM_CHANNEL,
len(cfg.RPN.ANCHOR_RATIOS)) for pi in p23456
]
multilevel_label_logits = [k[0] for k in rpn_outputs]
multilevel_box_logits = [k[1] for k in rpn_outputs]
proposal_boxes, proposal_scores = generate_fpn_proposals(
multilevel_anchors, multilevel_label_logits, multilevel_box_logits,
image_shape2d)
gt_boxes, gt_labels = inputs['gt_boxes'], inputs['gt_labels']
if is_training:
proposals = sample_fast_rcnn_targets(proposal_boxes, gt_boxes,
gt_labels)
else:
proposals = BoxProposals(proposal_boxes)
fastrcnn_head_func = getattr(model_frcnn, cfg.FPN.FRCNN_HEAD_FUNC)
if not cfg.FPN.CASCADE:
roi_feature_fastrcnn = multilevel_roi_align(
p23456[:4], proposals.boxes, 7)
head_feature = fastrcnn_head_func('fastrcnn', roi_feature_fastrcnn)
fastrcnn_label_logits, fastrcnn_box_logits = fastrcnn_outputs(
'fastrcnn/outputs', head_feature, cfg.DATA.NUM_CLASS)
fastrcnn_head = FastRCNNHead(
proposals, fastrcnn_box_logits, fastrcnn_label_logits,
tf.constant(cfg.FRCNN.BBOX_REG_WEIGHTS, dtype=tf.float32))
else:
def roi_func(boxes):
return multilevel_roi_align(p23456[:4], boxes, 7)
fastrcnn_head = CascadeRCNNHead(proposals, roi_func,
fastrcnn_head_func, image_shape2d,
cfg.DATA.NUM_CLASS)
if is_training:
all_losses = []
all_losses.extend(
multilevel_rpn_losses(multilevel_anchors,
multilevel_label_logits,
multilevel_box_logits))
all_losses.extend(fastrcnn_head.losses())
if cfg.MODE_MASK:
# maskrcnn loss
roi_feature_maskrcnn = multilevel_roi_align(
p23456[:4],
proposals.fg_boxes(),
14,
name_scope='multilevel_roi_align_mask')
maskrcnn_head_func = getattr(model_mrcnn,
cfg.FPN.MRCNN_HEAD_FUNC)
mask_logits = maskrcnn_head_func(
'maskrcnn', roi_feature_maskrcnn,
cfg.DATA.NUM_CATEGORY) # #fg x #cat x 28 x 28
target_masks_for_fg = crop_and_resize(
tf.expand_dims(inputs['gt_masks'], 1),
proposals.fg_boxes(),
proposals.fg_inds_wrt_gt,
28,
pad_border=False) # fg x 1x28x28
target_masks_for_fg = tf.squeeze(target_masks_for_fg, 1,
'sampled_fg_mask_targets')
all_losses.append(
maskrcnn_loss(mask_logits, proposals.fg_labels(),
target_masks_for_fg))
wd_cost = regularize_cost('.*/W',
l2_regularizer(cfg.TRAIN.WEIGHT_DECAY),
name='wd_cost')
all_losses.append(wd_cost)
total_cost = tf.add_n(all_losses, 'total_cost')
add_moving_summary(total_cost, wd_cost)
return total_cost
else:
decoded_boxes = fastrcnn_head.decoded_output_boxes()
decoded_boxes = clip_boxes(decoded_boxes,
image_shape2d,
name='fastrcnn_all_boxes')
label_scores = fastrcnn_head.output_scores(
name='fastrcnn_all_scores')
final_boxes, final_scores, final_labels = fastrcnn_predictions(
decoded_boxes, label_scores, name_scope='output')
if cfg.MODE_MASK:
# Cascade inference needs roi transform with refined boxes.
roi_feature_maskrcnn = multilevel_roi_align(
p23456[:4], final_boxes, 14)
maskrcnn_head_func = getattr(model_mrcnn,
cfg.FPN.MRCNN_HEAD_FUNC)
mask_logits = maskrcnn_head_func(
'maskrcnn', roi_feature_maskrcnn,
cfg.DATA.NUM_CATEGORY) # #fg x #cat x 28 x 28
indices = tf.stack([
tf.range(tf.size(final_labels)),
tf.to_int32(final_labels) - 1
],
axis=1)
final_mask_logits = tf.gather_nd(mask_logits,
indices) # #resultx28x28
tf.sigmoid(final_mask_logits, name='output/masks')
def build_graph(self, *inputs):
# TODO need to make tensorpack handles dict better
inputs = dict(zip(self.input_names, inputs))
is_training = get_current_tower_context().is_training
image = self.preprocess(inputs['image']) # 1CHW
featuremap = resnet_c4_backbone(image,
cfg.BACKBONE.RESNET_NUM_BLOCK[:3])
rpn_label_logits, rpn_box_logits = rpn_head('rpn', featuremap,
cfg.RPN.HEAD_DIM,
cfg.RPN.NUM_ANCHOR)
anchors = RPNAnchors(get_all_anchors(), inputs['anchor_labels'],
inputs['anchor_boxes'])
anchors = anchors.narrow_to(featuremap)
image_shape2d = tf.shape(image)[2:] # h,w
pred_boxes_decoded = anchors.decode_logits(
rpn_box_logits) # fHxfWxNAx4, floatbox
proposal_boxes, proposal_scores = generate_rpn_proposals(
tf.reshape(pred_boxes_decoded, [-1, 4]),
tf.reshape(rpn_label_logits,
[-1]), image_shape2d, cfg.RPN.TRAIN_PRE_NMS_TOPK
if is_training else cfg.RPN.TEST_PRE_NMS_TOPK,
cfg.RPN.TRAIN_POST_NMS_TOPK
if is_training else cfg.RPN.TEST_POST_NMS_TOPK)
gt_boxes, gt_labels = inputs['gt_boxes'], inputs['gt_labels']
if is_training:
# sample proposal boxes in training
proposals = sample_fast_rcnn_targets(proposal_boxes, gt_boxes,
gt_labels)
else:
# The boxes to be used to crop RoIs.
# Use all proposal boxes in inference
proposals = BoxProposals(proposal_boxes)
boxes_on_featuremap = proposals.boxes * (1.0 / cfg.RPN.ANCHOR_STRIDE)
roi_resized = roi_align(featuremap, boxes_on_featuremap, 14)
feature_fastrcnn = resnet_conv5(
roi_resized, cfg.BACKBONE.RESNET_NUM_BLOCK[-1]) # nxcx7x7
# Keep C5 feature to be shared with mask branch
feature_gap = GlobalAvgPooling('gap',
feature_fastrcnn,
data_format='channels_first')
fastrcnn_label_logits, fastrcnn_box_logits = fastrcnn_outputs(
'fastrcnn', feature_gap, cfg.DATA.NUM_CLASS)
fastrcnn_head = FastRCNNHead(
proposals, fastrcnn_box_logits, fastrcnn_label_logits,
tf.constant(cfg.FRCNN.BBOX_REG_WEIGHTS, dtype=tf.float32))
if is_training:
all_losses = []
# rpn loss
all_losses.extend(
rpn_losses(anchors.gt_labels, anchors.encoded_gt_boxes(),
rpn_label_logits, rpn_box_logits))
# fastrcnn loss
all_losses.extend(fastrcnn_head.losses())
if cfg.MODE_MASK:
# maskrcnn loss
# In training, mask branch shares the same C5 feature.
fg_feature = tf.gather(feature_fastrcnn, proposals.fg_inds())
mask_logits = maskrcnn_upXconv_head(
'maskrcnn', fg_feature, cfg.DATA.NUM_CATEGORY,
num_convs=0) # #fg x #cat x 14x14
target_masks_for_fg = crop_and_resize(
tf.expand_dims(inputs['gt_masks'], 1),
proposals.fg_boxes(),
proposals.fg_inds_wrt_gt,
14,
pad_border=False) # nfg x 1x14x14
target_masks_for_fg = tf.squeeze(target_masks_for_fg, 1,
'sampled_fg_mask_targets')
all_losses.append(
maskrcnn_loss(mask_logits, proposals.fg_labels(),
target_masks_for_fg))
wd_cost = regularize_cost('.*/W',
l2_regularizer(cfg.TRAIN.WEIGHT_DECAY),
name='wd_cost')
all_losses.append(wd_cost)
total_cost = tf.add_n(all_losses, 'total_cost')
add_moving_summary(total_cost, wd_cost)
return total_cost
else:
decoded_boxes = fastrcnn_head.decoded_output_boxes()
decoded_boxes = clip_boxes(decoded_boxes,
image_shape2d,
name='fastrcnn_all_boxes')
label_scores = fastrcnn_head.output_scores(
name='fastrcnn_all_scores')
final_boxes, final_scores, final_labels = fastrcnn_predictions(
decoded_boxes, label_scores, name_scope='output')
if cfg.MODE_MASK:
roi_resized = roi_align(
featuremap, final_boxes * (1.0 / cfg.RPN.ANCHOR_STRIDE),
14)
feature_maskrcnn = resnet_conv5(
roi_resized, cfg.BACKBONE.RESNET_NUM_BLOCK[-1])
mask_logits = maskrcnn_upXconv_head(
'maskrcnn', feature_maskrcnn, cfg.DATA.NUM_CATEGORY,
0) # #result x #cat x 14x14
indices = tf.stack([
tf.range(tf.size(final_labels)),
tf.to_int32(final_labels) - 1
],
axis=1)
final_mask_logits = tf.gather_nd(mask_logits,
indices) # #resultx14x14
tf.sigmoid(final_mask_logits, name='output/masks')
def rpn(self, image, features, inputs, orig_image_dims, seed_gen):
"""
The RPN part of the graph that generate the RPN proposal and losses
Args:
image: BS x NumChannel x H_image x W_image
features: ([tf.Tensor]): A list of 5 FPN feature maps, i.e. level P23456, each with BS x NumChannel x H_feature x W_feature
inputs: dict, contains all input information
orig_image_dims: BS x 3
Returns:
proposal_boxes: top K region proposals, K x 5
losses: scalar, sum of the label loss and box loss
"""
assert len(cfg.RPN.ANCHOR_SIZES) == len(cfg.FPN.ANCHOR_STRIDES)
image_shape2d = orig_image_dims[: ,:2]
all_anchors_fpn = get_all_anchors_fpn()
rpn_outputs = []
if not cfg.TRAIN.RPN_NCHW:
features = [nchw_to_nhwc_transform(c) for c in features]
for pi in features:
# label_logits: BS x H_feaure x W_feature x NA, box_logits: BS x (NA * 4) x H_feature x W_feature
label_logits, box_logits = rpn_head('rpn', pi, cfg.FPN.NUM_CHANNEL, len(cfg.RPN.ANCHOR_RATIOS), seed_gen=seed_gen, fp16=self.fp16)
rpn_outputs.append((label_logits, box_logits))
multilevel_label_logits = [k[0] for k in rpn_outputs] # Num_level * [BS x H_feature x W_feature x NA]
multilevel_box_logits = [k[1] for k in rpn_outputs] # Num_level * [BS x (NA * 4) x H_feature x W_feature]
# proposal_boxes: K x 5, proposal_scores: 1-D K
if cfg.RPN.TOPK_PER_IMAGE:
proposal_boxes, proposal_scores = generate_fpn_proposals_topk_per_image(all_anchors_fpn,
multilevel_box_logits,
multilevel_label_logits,
image_shape2d,
cfg.TRAIN.BATCH_SIZE_PER_GPU)
else:
proposal_boxes, proposal_scores = generate_fpn_proposals(all_anchors_fpn,
multilevel_box_logits,
multilevel_label_logits,
image_shape2d,
cfg.TRAIN.BATCH_SIZE_PER_GPU)
if self.training:
multilevel_anchor_labels = [inputs['anchor_labels_lvl{}'.format(i + 2)] for i in range(len(all_anchors_fpn))]
multilevel_anchor_boxes = [inputs['anchor_boxes_lvl{}'.format(i + 2)] for i in range(len(all_anchors_fpn))]
multilevel_box_logits_reshaped = []
for box_logits in multilevel_box_logits:
shp = tf.shape(box_logits) # BS x (NA * 4) x H_feature x W_feature
box_logits_t = tf.transpose(box_logits, [0, 2, 3, 1]) # BS x H_feature x W_feature x (NA * 4)
box_logits_t = tf.reshape(box_logits_t, tf.stack([shp[0], shp[2], shp[3], -1, 4])) # BS x H_feature x W_feature x NA x 4
multilevel_box_logits_reshaped.append(box_logits_t)
rpn_losses = []
for i in range(cfg.TRAIN.BATCH_SIZE_PER_GPU):
orig_image_hw = orig_image_dims[i, :2]
si_all_anchors_fpn = get_all_anchors_fpn()
si_multilevel_box_logits = [box_logits[i] for box_logits in multilevel_box_logits_reshaped] # [H_feature x W_feature x NA x 4] * Num_levels
si_multilevel_label_logits = [label_logits[i] for label_logits in multilevel_label_logits] # [H_feature x W_feature x NA] * Num_levels
si_multilevel_anchor_labels = [anchor_labels[i] for anchor_labels in multilevel_anchor_labels]
si_multilevel_anchors_boxes = [anchor_boxes[i] for anchor_boxes in multilevel_anchor_boxes]
si_multilevel_anchors = [RPNAnchors(si_all_anchors_fpn[j],
si_multilevel_anchor_labels[j],
si_multilevel_anchors_boxes[j])
for j in range(len(features))]
# Given the original image dims, find what size each layer of the FPN feature map would be (follow FPN padding logic)
mult = float \
(cfg.FPN.RESOLUTION_REQUIREMENT) # the image is padded so that it is a multiple of this (32 with default config).
orig_image_hw_after_fpn_padding = tf.ceil(tf.cast(orig_image_hw, tf.float32) / mult) * mult
featuremap_dims_per_level = []
for lvl, stride in enumerate(cfg.FPN.ANCHOR_STRIDES):
featuremap_dims_float = orig_image_hw_after_fpn_padding / float(stride)
featuremap_dims_per_level.append \
(tf.cast(tf.math.floor(featuremap_dims_float + 0.5), tf.int32)) # Fix bankers rounding
si_multilevel_anchors_narrowed = [anchors.narrow_to_featuremap_dims(dims) for anchors, dims in zip(si_multilevel_anchors, featuremap_dims_per_level)]
si_multilevel_box_logits_narrowed = [box_logits[:dims[0], :dims[1] ,: ,:] for box_logits, dims in zip(si_multilevel_box_logits, featuremap_dims_per_level)]
si_multilevel_label_logits_narrowed = [label_logits[:dims[0], :dims[1] ,:] for label_logits, dims in zip(si_multilevel_label_logits, featuremap_dims_per_level)]
si_losses = multilevel_rpn_losses(si_multilevel_anchors_narrowed,
si_multilevel_label_logits_narrowed,
si_multilevel_box_logits_narrowed)
rpn_losses.extend(si_losses)
with tf.name_scope('rpn_losses'):
total_label_loss = tf.truediv(tf.add_n(rpn_losses[::2]), tf.cast(cfg.TRAIN.BATCH_SIZE_PER_GPU, dtype=tf.float32), name='label_loss')
total_box_loss = tf.truediv(tf.add_n(rpn_losses[1::2]), tf.cast(cfg.TRAIN.BATCH_SIZE_PER_GPU, dtype=tf.float32), name='box_loss')
add_moving_summary(total_label_loss, total_box_loss)
losses = [total_label_loss, total_box_loss]
else:
losses = []
return proposal_boxes, losses
def build_graph(self, *inputs):
is_training = get_current_tower_context().is_training
if cfg.MODE_MASK:
image, anchor_labels, anchor_boxes, gt_boxes, gt_labels, gt_masks = inputs
else:
image, anchor_labels, anchor_boxes, gt_boxes, gt_labels = inputs
image = self.preprocess(image) # 1CHW
#with varreplace.freeze_variables(stop_gradient=True, skip_collection=True):
featuremap = resnet_c4_backbone(image, cfg.BACKBONE.RESNET_NUM_BLOCK[:3])
# freeze
# featuremap = tf.stop_gradient(featuremap)
rpn_label_logits, rpn_box_logits = rpn_head('rpn', featuremap, cfg.RPN.HEAD_DIM, cfg.RPN.NUM_ANCHOR)
anchors = RPNAnchors(get_all_anchors(), anchor_labels, anchor_boxes)
anchors = anchors.narrow_to(featuremap)
image_shape2d = tf.shape(image)[2:] # h,w
pred_boxes_decoded = anchors.decode_logits(rpn_box_logits) # fHxfWxNAx4, floatbox
proposal_boxes, proposal_scores = generate_rpn_proposals(
tf.reshape(pred_boxes_decoded, [-1, 4]),
tf.reshape(rpn_label_logits, [-1]),
image_shape2d,
cfg.RPN.TRAIN_PRE_NMS_TOPK if is_training else cfg.RPN.TEST_PRE_NMS_TOPK,
cfg.RPN.TRAIN_POST_NMS_TOPK if is_training else cfg.RPN.TEST_POST_NMS_TOPK)
if is_training:
# sample proposal boxes in training
rcnn_boxes, rcnn_labels, fg_inds_wrt_gt = sample_fast_rcnn_targets(
proposal_boxes, gt_boxes, gt_labels)
else:
# The boxes to be used to crop RoIs.
# Use all proposal boxes in inference
rcnn_boxes = proposal_boxes
featuremap = resnet_conv5(featuremap, cfg.BACKBONE.RESNET_NUM_BLOCK[-1])
rfcn_cls = Conv2D('rfcn_cls', featuremap, cfg.DATA.NUM_CLASS*3*3, (1, 1), data_format='channels_first')
rfcn_reg = Conv2D('rfcn_reg', featuremap, cfg.DATA.NUM_CLASS*4*3*3, (1, 1), data_format='channels_first')
boxes_on_featuremap = rcnn_boxes * (1.0 / cfg.RPN.ANCHOR_STRIDE)
classify_vote = VotePooling('votepooling_cls', rfcn_cls, boxes_on_featuremap, 3, 3)
classify_regr = VotePooling('votepooling_regr', rfcn_reg, boxes_on_featuremap, 3, 3, isCls=False)
classify_regr = tf.reshape(classify_regr, [-1, cfg.DATA.NUM_CLASS, 4])
if is_training:
# rpn loss
rpn_label_loss, rpn_box_loss = rpn_losses(
anchors.gt_labels, anchors.encoded_gt_boxes(), rpn_label_logits, rpn_box_logits)
# fastrcnn loss
matched_gt_boxes = tf.gather(gt_boxes, fg_inds_wrt_gt)
fg_inds_wrt_sample = tf.reshape(tf.where(rcnn_labels > 0), [-1]) # fg inds w.r.t all samples
fg_sampled_boxes = tf.gather(rcnn_boxes, fg_inds_wrt_sample)
fg_fastrcnn_box_logits = tf.gather(classify_regr, fg_inds_wrt_sample)
fastrcnn_label_loss, fastrcnn_box_loss = self.fastrcnn_training(
image, rcnn_labels, fg_sampled_boxes,
matched_gt_boxes, classify_vote, fg_fastrcnn_box_logits)
if cfg.MODE_MASK:
# maskrcnn loss
fg_labels = tf.gather(rcnn_labels, fg_inds_wrt_sample)
# In training, mask branch shares the same C5 feature.
fg_feature = tf.gather(feature_fastrcnn, fg_inds_wrt_sample)
mask_logits = maskrcnn_upXconv_head(
'maskrcnn', fg_feature, cfg.DATA.NUM_CATEGORY, num_convs=0) # #fg x #cat x 14x14
target_masks_for_fg = crop_and_resize(
tf.expand_dims(gt_masks, 1),
fg_sampled_boxes,
fg_inds_wrt_gt, 14,
pad_border=False) # nfg x 1x14x14
target_masks_for_fg = tf.squeeze(target_masks_for_fg, 1, 'sampled_fg_mask_targets')
mrcnn_loss = maskrcnn_loss(mask_logits, fg_labels, target_masks_for_fg)
else:
mrcnn_loss = 0.0
wd_cost = regularize_cost(
'.*/W', l2_regularizer(cfg.TRAIN.WEIGHT_DECAY), name='wd_cost')
total_cost = tf.add_n([
rpn_label_loss, rpn_box_loss,
fastrcnn_label_loss, fastrcnn_box_loss,
mrcnn_loss, wd_cost], 'total_cost')
add_moving_summary(total_cost, wd_cost)
return total_cost
else:
final_boxes, final_labels = self.fastrcnn_inference(
image_shape2d, rcnn_boxes, classify_vote, classify_regr)
if cfg.MODE_MASK:
roi_resized = roi_align(featuremap, final_boxes * (1.0 / cfg.RPN.ANCHOR_STRIDE), 14)
feature_maskrcnn = resnet_conv5(roi_resized, cfg.BACKBONE.RESNET_NUM_BLOCK[-1])
mask_logits = maskrcnn_upXconv_head(
'maskrcnn', feature_maskrcnn, cfg.DATA.NUM_CATEGORY, 0) # #result x #cat x 14x14
indices = tf.stack([tf.range(tf.size(final_labels)), tf.to_int32(final_labels) - 1], axis=1)
final_mask_logits = tf.gather_nd(mask_logits, indices) # #resultx14x14
tf.sigmoid(final_mask_logits, name='final_masks')
def build_graph(self, *inputs):
num_fpn_level = len(cfg.FPN.ANCHOR_STRIDES)
assert len(cfg.RPN.ANCHOR_SIZES) == num_fpn_level
is_training = get_current_tower_context().is_training
image = inputs[0]
input_anchors = inputs[1: 1 + 2 * num_fpn_level]
multilevel_anchors = [RPNAnchors(*args) for args in
zip(get_all_anchors_fpn(), input_anchors[0::2], input_anchors[1::2])]
gt_boxes, gt_labels = inputs[11], inputs[12]
if cfg.MODE_MASK:
gt_masks = inputs[-1]
image = self.preprocess(image) # 1CHW
image_shape2d = tf.shape(image)[2:] # h,w
c2345 = resnet_fpn_backbone(image, cfg.BACKBONE.RESNET_NUM_BLOCK)
p23456 = fpn_model('fpn', c2345)
self.slice_feature_and_anchors(image_shape2d, p23456, multilevel_anchors)
# Multi-Level RPN Proposals
rpn_outputs = [rpn_head('rpn', pi, cfg.FPN.NUM_CHANNEL, len(cfg.RPN.ANCHOR_RATIOS))
for pi in p23456]
multilevel_label_logits = [k[0] for k in rpn_outputs]
multilevel_box_logits = [k[1] for k in rpn_outputs]
proposal_boxes, proposal_scores = generate_fpn_proposals(
multilevel_anchors, multilevel_label_logits,
multilevel_box_logits, image_shape2d)
if is_training:
rcnn_boxes, rcnn_labels, fg_inds_wrt_gt = sample_fast_rcnn_targets(
proposal_boxes, gt_boxes, gt_labels)
else:
# The boxes to be used to crop RoIs.
rcnn_boxes = proposal_boxes
roi_feature_fastrcnn = multilevel_roi_align(p23456[:4], rcnn_boxes, 7)
fastrcnn_head_func = getattr(model_frcnn, cfg.FPN.FRCNN_HEAD_FUNC)
fastrcnn_label_logits, fastrcnn_box_logits = fastrcnn_head_func(
'fastrcnn', roi_feature_fastrcnn, cfg.DATA.NUM_CLASS)
if is_training:
# rpn loss:
rpn_label_loss, rpn_box_loss = multilevel_rpn_losses(
multilevel_anchors, multilevel_label_logits, multilevel_box_logits)
# fastrcnn loss:
matched_gt_boxes = tf.gather(gt_boxes, fg_inds_wrt_gt)
fg_inds_wrt_sample = tf.reshape(tf.where(rcnn_labels > 0), [-1]) # fg inds w.r.t all samples
fg_sampled_boxes = tf.gather(rcnn_boxes, fg_inds_wrt_sample)
fg_fastrcnn_box_logits = tf.gather(fastrcnn_box_logits, fg_inds_wrt_sample)
fastrcnn_label_loss, fastrcnn_box_loss = self.fastrcnn_training(
image, rcnn_labels, fg_sampled_boxes,
matched_gt_boxes, fastrcnn_label_logits, fg_fastrcnn_box_logits)
if cfg.MODE_MASK:
# maskrcnn loss
fg_labels = tf.gather(rcnn_labels, fg_inds_wrt_sample)
roi_feature_maskrcnn = multilevel_roi_align(
p23456[:4], fg_sampled_boxes, 14,
name_scope='multilevel_roi_align_mask')
maskrcnn_head_func = getattr(model_mrcnn, cfg.FPN.MRCNN_HEAD_FUNC)
mask_logits = maskrcnn_head_func(
'maskrcnn', roi_feature_maskrcnn, cfg.DATA.NUM_CATEGORY) # #fg x #cat x 28 x 28
target_masks_for_fg = crop_and_resize(
tf.expand_dims(gt_masks, 1),
fg_sampled_boxes,
fg_inds_wrt_gt, 28,
pad_border=False) # fg x 1x28x28
target_masks_for_fg = tf.squeeze(target_masks_for_fg, 1, 'sampled_fg_mask_targets')
mrcnn_loss = maskrcnn_loss(mask_logits, fg_labels, target_masks_for_fg)
else:
mrcnn_loss = 0.0
wd_cost =regularize_cost('fastrcnn/.*/W', l2_regularizer(cfg.TRAIN.WEIGHT_DECAY), name='wd_cost')
#wd_cost = regularize_cost(
# '.*/W', l2_regularizer(cfg.TRAIN.WEIGHT_DECAY), name='wd_cost')
total_cost = tf.add_n([rpn_label_loss, rpn_box_loss,
fastrcnn_label_loss, fastrcnn_box_loss,
mrcnn_loss, wd_cost], 'total_cost')
add_moving_summary(total_cost, wd_cost)
return total_cost
else:
final_boxes, final_labels = self.fastrcnn_inference(
image_shape2d, rcnn_boxes, fastrcnn_label_logits, fastrcnn_box_logits)
if cfg.MODE_MASK:
# Cascade inference needs roi transform with refined boxes.
roi_feature_maskrcnn = multilevel_roi_align(p23456[:4], final_boxes, 14)
maskrcnn_head_func = getattr(model_mrcnn, cfg.FPN.MRCNN_HEAD_FUNC)
mask_logits = maskrcnn_head_func(
'maskrcnn', roi_feature_maskrcnn, cfg.DATA.NUM_CATEGORY) # #fg x #cat x 28 x 28
indices = tf.stack([tf.range(tf.size(final_labels)), tf.to_int32(final_labels) - 1], axis=1)
final_mask_logits = tf.gather_nd(mask_logits, indices) # #resultx28x28
tf.sigmoid(final_mask_logits, name='final_masks')
'anchor_boxes')
gt_boxes = tf.placeholder(tf.float32, (None, 4), 'gt_boxes')
gt_labels = tf.placeholder(tf.int64, (None, ), 'gt_labels')
image = preprocess(image_P)
Load_Weights = []
with TowerContext('', is_training=Trainining_is):
featuremap = resnet_c4_backbone(image,
cfg.BACKBONE.RESNET_NUM_BLOCK[:3])
rpn_label_logits, rpn_box_logits = rpn_head('rpn', featuremap,
cfg.RPN.HEAD_DIM,
cfg.RPN.NUM_ANCHOR)
anchors = RPNAnchors(get_all_anchors(), anchor_labels, anchor_boxes)
anchors = anchors.narrow_to(featuremap)
image_shape2d = tf.shape(image)[2:] # h,w
pred_boxes_decoded = anchors.decode_logits(
rpn_box_logits) # fHxfWxNAx4, floatbox
proposal_boxes, proposal_scores = generate_rpn_proposals(
tf.reshape(pred_boxes_decoded, [-1, 4]),
tf.reshape(rpn_label_logits,
[-1]), image_shape2d, cfg.RPN.TRAIN_PRE_NMS_TOPK
if is_training else cfg.RPN.TEST_PRE_NMS_TOPK,
cfg.RPN.TRAIN_POST_NMS_TOPK
if is_training else cfg.RPN.TEST_POST_NMS_TOPK)
if is_training:
# sample proposal boxes in training
rcnn_boxes, rcnn_labels, fg_inds_wrt_gt = sample_fast_rcnn_targets(
def build_graph(self, *inputs):
is_training = get_current_tower_context().is_training
image, anchor_labels, anchor_boxes, gt_boxes, gt_labels, gt_ids, orig_shape = inputs
image = self.preprocess(image) # 1CHW
featuremap = resnet_c4_backbone(image,
cfg.BACKBONE.RESNET_NUM_BLOCK[:3])
rpn_label_logits, rpn_box_logits = rpn_head('rpn', featuremap,
cfg.RPN.HEAD_DIM,
cfg.RPN.NUM_ANCHOR)
anchors = RPNAnchors(get_all_anchors(), anchor_labels, anchor_boxes)
anchors = anchors.narrow_to(featuremap)
image_shape2d = tf.shape(image)[2:] # h,w
# decode into actual image coordinates
pred_boxes_decoded = anchors.decode_logits(
rpn_box_logits) # fHxfWxNAx4, floatbox
proposal_boxes, proposal_scores = generate_rpn_proposals(
tf.reshape(pred_boxes_decoded, [-1, 4]),
tf.reshape(rpn_label_logits,
[-1]), image_shape2d, cfg.RPN.TRAIN_PRE_NMS_TOPK
if is_training else cfg.RPN.TEST_PRE_NMS_TOPK,
cfg.RPN.TRAIN_POST_NMS_TOPK
if is_training else cfg.RPN.TEST_POST_NMS_TOPK)
if is_training:
# sample proposal boxes in training
rcnn_boxes, rcnn_labels, fg_inds_wrt_gt = sample_fast_rcnn_targets(
proposal_boxes, gt_boxes, gt_labels)
else:
# The boxes to be used to crop RoIs.
# Use all proposal boxes in inference
rcnn_boxes = proposal_boxes
boxes_on_featuremap = rcnn_boxes * (1.0 / cfg.RPN.ANCHOR_STRIDE)
# size? #proposals*h*w*c?
roi_resized = roi_align(featuremap, boxes_on_featuremap, 14)
feature_fastrcnn = resnet_conv5(
roi_resized, cfg.BACKBONE.RESNET_NUM_BLOCK[-1]) # nxcx7x7
# Keep C5 feature to be shared with mask branch
feature_gap = GlobalAvgPooling('gap',
feature_fastrcnn,
data_format='channels_first')
fastrcnn_label_logits, fastrcnn_box_logits = fastrcnn_outputs(
'fastrcnn', feature_gap, cfg.DATA.NUM_CLASS)
if is_training:
# rpn loss
rpn_label_loss, rpn_box_loss = rpn_losses(
anchors.gt_labels, anchors.encoded_gt_boxes(),
rpn_label_logits, rpn_box_logits)
# fastrcnn loss
matched_gt_boxes = tf.gather(gt_boxes, fg_inds_wrt_gt)
fg_inds_wrt_sample = tf.reshape(tf.where(rcnn_labels > 0),
[-1]) # fg inds w.r.t all samples
# outputs from fg proposals
fg_sampled_boxes = tf.gather(rcnn_boxes, fg_inds_wrt_sample)
fg_fastrcnn_box_logits = tf.gather(fastrcnn_box_logits,
fg_inds_wrt_sample)
# rcnn_labels: the labels of the proposals
# fg_sampled_boxes: fg proposals
# matched_gt_boxes: just like RPN, the gt boxes
# that match the corresponding fg proposals
fastrcnn_label_loss, fastrcnn_box_loss = self.fastrcnn_training(
image, rcnn_labels, fg_sampled_boxes, matched_gt_boxes,
fastrcnn_label_logits, fg_fastrcnn_box_logits)
# acquire pred for re-id training
# turning NMS off gives re-id branch more training samples
if cfg.RE_ID.NMS:
boxes, final_labels, final_probs = self.fastrcnn_inference(
image_shape2d, rcnn_boxes, fastrcnn_label_logits,
fastrcnn_box_logits)
else:
boxes, final_labels, final_probs = self.fastrcnn_inference_id(
image_shape2d, rcnn_boxes, fastrcnn_label_logits,
fastrcnn_box_logits)
# scale = tf.sqrt(tf.cast(image_shape2d[0], tf.float32) / tf.cast(orig_shape[0], tf.float32) *
# tf.cast(image_shape2d[1], tf.float32) / tf.cast(orig_shape[1], tf.float32))
# final_boxes = boxes / scale
# # boxes are already clipped inside the graph, but after the floating point scaling, this may not be true any more.
# final_boxes = tf_clip_boxes(final_boxes, orig_shape)
# IOU, discard bad dets, assign re-id labels
# the results are already NMS so no need to NMS again
# crop from conv4 with dets (maybe plus gts)
# feedforward re-id branch
# resizing during ROIalign?
iou = pairwise_iou(boxes, gt_boxes) # are the gt boxes resized?
tp_mask = tf.reduce_max(iou, axis=1) >= cfg.RE_ID.IOU_THRESH
iou = tf.boolean_mask(iou, tp_mask)
# return iou to debug
def re_id_loss(pred_boxes, pred_matching_gt_ids, featuremap):
with tf.variable_scope('id_head'):
num_of_samples_used = tf.get_variable(
'num_of_samples_used', initializer=0, trainable=False)
num_of_samples_used = num_of_samples_used.assign_add(
tf.shape(pred_boxes)[0])
boxes_on_featuremap = pred_boxes * (1.0 /
cfg.RPN.ANCHOR_STRIDE)
# name scope?
# stop gradient
roi_resized = roi_align(featuremap, boxes_on_featuremap,
14)
feature_idhead = resnet_conv5(
roi_resized,
cfg.BACKBONE.RESNET_NUM_BLOCK[-1]) # nxcx7x7
feature_gap = GlobalAvgPooling(
'gap', feature_idhead, data_format='channels_first')
init = tf.variance_scaling_initializer()
hidden = FullyConnected('fc6',
feature_gap,
1024,
kernel_initializer=init,
activation=tf.nn.relu)
hidden = FullyConnected('fc7',
hidden,
1024,
kernel_initializer=init,
activation=tf.nn.relu)
hidden = FullyConnected('fc8',
hidden,
256,
kernel_initializer=init,
activation=tf.nn.relu)
id_logits = FullyConnected(
'class',
hidden,
cfg.DATA.NUM_ID,
kernel_initializer=tf.random_normal_initializer(
stddev=0.01))
label_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=pred_matching_gt_ids, logits=id_logits)
label_loss = tf.reduce_mean(label_loss, name='label_loss')
return label_loss, num_of_samples_used
def check_unid_pedes(iou, gt_ids, boxes, tp_mask, featuremap):
pred_gt_ind = tf.argmax(iou, axis=1)
# output following tensors
# pick out the -2 class here
pred_matching_gt_ids = tf.gather(gt_ids, pred_gt_ind)
pred_boxes = tf.boolean_mask(boxes, tp_mask)
# label 1 corresponds to unid pedes
unid_ind = tf.not_equal(pred_matching_gt_ids, 1)
pred_matching_gt_ids = tf.boolean_mask(pred_matching_gt_ids,
unid_ind)
pred_boxes = tf.boolean_mask(pred_boxes, unid_ind)
ret = tf.cond(
tf.equal(tf.size(pred_boxes), 0), lambda:
(tf.constant(cfg.RE_ID.STABLE_LOSS), tf.constant(0)),
lambda: re_id_loss(pred_boxes, pred_matching_gt_ids,
featuremap))
return ret
with tf.name_scope('id_head'):
# no detection has IOU > 0.7, re-id returns 0 loss
re_id_loss, num_of_samples_used = tf.cond(
tf.equal(tf.size(iou), 0), lambda:
(tf.constant(cfg.RE_ID.STABLE_LOSS), tf.constant(0)),
lambda: check_unid_pedes(iou, gt_ids, boxes, tp_mask,
featuremap))
add_tensor_summary(num_of_samples_used, ['scalar'],
name='num_of_samples_used')
# for debug, use tensor name to take out the handle
# return re_id_loss
# pred_gt_ind = tf.argmax(iou, axis=1)
# # output following tensors
# # pick out the -2 class here
# pred_gt_ids = tf.gather(gt_ids, pred_gt_ind)
# pred_boxes = tf.boolean_mask(boxes, tp_mask)
# unid_ind = pred_gt_ids != 1
# return unid_ind
# return tf.shape(boxes)[0]
unnormed_id_loss = tf.identity(re_id_loss, name='unnormed_id_loss')
re_id_loss = tf.divide(re_id_loss, cfg.RE_ID.LOSS_NORMALIZATION,
're_id_loss')
add_moving_summary(unnormed_id_loss)
add_moving_summary(re_id_loss)
wd_cost = regularize_cost('.*/W',
l2_regularizer(cfg.TRAIN.WEIGHT_DECAY),
name='wd_cost')
# weights on the losses?
total_cost = tf.add_n([
rpn_label_loss, rpn_box_loss, fastrcnn_label_loss,
fastrcnn_box_loss, re_id_loss, wd_cost
], 'total_cost')
add_moving_summary(total_cost, wd_cost)
return total_cost
else:
if cfg.RE_ID.QUERY_EVAL:
# resize the gt_boxes in dataflow
final_boxes = gt_boxes
else:
final_boxes, final_labels, _ = self.fastrcnn_inference(
image_shape2d, rcnn_boxes, fastrcnn_label_logits,
fastrcnn_box_logits)
with tf.variable_scope('id_head'):
preds_on_featuremap = final_boxes * (1.0 /
cfg.RPN.ANCHOR_STRIDE)
# name scope?
# stop gradient
roi_resized = roi_align(featuremap, preds_on_featuremap, 14)
feature_idhead = resnet_conv5(
roi_resized, cfg.BACKBONE.RESNET_NUM_BLOCK[-1]) # nxcx7x7
feature_gap = GlobalAvgPooling('gap',
feature_idhead,
data_format='channels_first')
hidden = FullyConnected('fc6',
feature_gap,
1024,
activation=tf.nn.relu)
hidden = FullyConnected('fc7',
hidden,
1024,
activation=tf.nn.relu)
fv = FullyConnected('fc8', hidden, 256, activation=tf.nn.relu)
id_logits = FullyConnected(
'class',
fv,
cfg.DATA.NUM_ID,
kernel_initializer=tf.random_normal_initializer(
stddev=0.01))
scale = tf.sqrt(
tf.cast(image_shape2d[0], tf.float32) /
tf.cast(orig_shape[0], tf.float32) *
tf.cast(image_shape2d[1], tf.float32) /
tf.cast(orig_shape[1], tf.float32))
rescaled_final_boxes = final_boxes / scale
# boxes are already clipped inside the graph, but after the floating point scaling, this may not be true any more.
# rescaled_final_boxes_pre_clip = tf.identity(rescaled_final_boxes, name='re_boxes_pre_clip')
rescaled_final_boxes = tf_clip_boxes(rescaled_final_boxes,
orig_shape)
rescaled_final_boxes = tf.identity(rescaled_final_boxes,
'rescaled_final_boxes')
fv = tf.identity(fv, name='feature_vector')
prob = tf.nn.softmax(id_logits, name='re_id_probs')
def build_graph(self, *inputs):
inputs = dict(zip(self.input_names, inputs))
is_training = get_current_tower_context().is_training
image = self.preprocess(inputs['image']) # 1CHW
featuremap = resnet_c4_backbone(image,
cfg.BACKBONE.RESNET_NUM_BLOCK[:3])
rpn_label_logits, rpn_box_logits = rpn_head('rpn', featuremap,
cfg.RPN.HEAD_DIM,
cfg.RPN.NUM_ANCHOR)
anchors = RPNAnchors(get_all_anchors(), inputs['anchor_labels'],
inputs['anchor_boxes'])
anchors = anchors.narrow_to(featuremap)
image_shape2d = tf.shape(image)[2:] # h,w
pred_boxes_decoded = anchors.decode_logits(
rpn_box_logits) # fHxfWxNAx4, floatbox
proposal_boxes, proposal_scores = generate_rpn_proposals(
tf.reshape(pred_boxes_decoded, [-1, 4]),
tf.reshape(rpn_label_logits,
[-1]), image_shape2d, cfg.RPN.TRAIN_PRE_NMS_TOPK
if is_training else cfg.RPN.TEST_PRE_NMS_TOPK,
cfg.RPN.TRAIN_POST_NMS_TOPK
if is_training else cfg.RPN.TEST_POST_NMS_TOPK)
gt_boxes, gt_labels, gt_masks = inputs['gt_boxes'], inputs[
'gt_labels'], inputs['gt_masks']
if is_training:
# sample proposal boxes in training
rcnn_boxes, rcnn_labels, fg_inds_wrt_gt, rcnn_masks = sample_fast_rcnn_targets(
proposal_boxes, gt_boxes, gt_labels, gt_masks)
matched_gt_boxes = tf.gather(gt_boxes,
fg_inds_wrt_gt,
name='gt_boxes_per_fg_proposal')
matched_gt_masks = tf.gather(gt_masks,
fg_inds_wrt_gt,
name='gt_masks_per_fg_proposal')
else:
# The boxes to be used to crop RoIs.
# Use all proposal boxes in inference
rcnn_boxes = proposal_boxes
angles = tf.ones((tf.shape(proposal_boxes)[0], 1)) * (-45.)
x1y1, x2y2 = proposal_boxes[:, 0:2], proposal_boxes[:, 2:4]
wh = x2y2 - x1y1
xy = (x2y2 + x1y1) * 0.5
rcnn_masks = tf.concat([xy, wh, angles], axis=1)
rcnn_labels, matched_gt_boxes, matched_gt_masks = None, None, None
# ToDo
boxes_on_featuremap = rcnn_boxes * (1.0 / cfg.RPN.ANCHOR_STRIDE)
roi_resized = roi_align(featuremap, boxes_on_featuremap, 14)
feature_fastrcnn = resnet_conv5(
roi_resized, cfg.BACKBONE.RESNET_NUM_BLOCK[-1]) # nxcx7x7
# Keep C5 feature to be shared with mask branch
feature_gap = GlobalAvgPooling('gap',
feature_fastrcnn,
data_format='channels_first')
fastrcnn_label_logits, fastrcnn_box_logits, fastrcnn_mask_logits = fastrcnn_outputs(
'fastrcnn', feature_gap, cfg.DATA.NUM_CLASS)
fastrcnn_head = FastRCNNHead(rcnn_boxes, rcnn_masks,
fastrcnn_box_logits, fastrcnn_mask_logits,
fastrcnn_label_logits, rcnn_labels,
matched_gt_boxes, matched_gt_masks)
if is_training:
# rpn loss
rpn_label_loss, rpn_box_loss = rpn_losses(
anchors.gt_labels, anchors.encoded_gt_boxes(),
rpn_label_logits, rpn_box_logits)
# fastrcnn loss
fastrcnn_label_loss, fastrcnn_box_loss, mask_loss = fastrcnn_head.losses(
)
wd_cost = regularize_cost('.*/W',
l2_regularizer(cfg.TRAIN.WEIGHT_DECAY),
name='wd_cost')
total_cost = tf.add_n([
rpn_label_loss, rpn_box_loss, fastrcnn_label_loss,
fastrcnn_box_loss, mask_loss, wd_cost
], 'total_cost')
add_moving_summary(total_cost, wd_cost)
return total_cost
else:
# ToDo
final_boxes, final_labels = self.fastrcnn_inference(
image_shape2d, fastrcnn_head)
indices = tf.stack([
tf.range(tf.size(final_labels)),
tf.to_int32(final_labels) - 1
],
axis=1)
final_mask_logits = tf.gather_nd(fastrcnn_mask_logits,
indices,
name='final_masks')
def build_graph(self, *inputs):
is_training = get_current_tower_context().is_training
# if cfg.MODE_MASK:
# image, anchor_labels, anchor_boxes, gt_boxes, gt_labels, gt_masks = inputs
# else:
# image, anchor_labels, anchor_boxes, gt_boxes, gt_labels, scale_index = inputs
image, anchor_labels, anchor_boxes, gt_boxes, gt_labels = inputs
image = self.preprocess(image) # 1CHW
featuremap = resnet_c4_backbone(image,
cfg.BACKBONE.RESNET_NUM_BLOCK[:3])
rpn_label_logits, rpn_box_logits = rpn_head(
'rpn', featuremap, cfg.RPN.HEAD_DIM, cfg.RPN.NUM_ANCHOR)
anchors = RPNAnchors(get_all_anchors(), anchor_labels, anchor_boxes)
anchors = anchors.narrow_to(featuremap)
image_shape2d = tf.shape(image)[2:] # h,w
pred_boxes_decoded = anchors.decode_logits(
rpn_box_logits) # fHxfWxNAx4, floatbox
proposal_boxes, proposal_scores = generate_rpn_proposals(
tf.reshape(pred_boxes_decoded, [-1, 4]),
tf.reshape(rpn_label_logits,
[-1]), image_shape2d, cfg.RPN.TRAIN_PRE_NMS_TOPK
if is_training else cfg.RPN.TEST_PRE_NMS_TOPK,
cfg.RPN.TRAIN_POST_NMS_TOPK
if is_training else cfg.RPN.TEST_POST_NMS_TOPK)
if is_training:
# sample proposal boxes in training
# rcnn_boxes, rcnn_labels, fg_inds_wrt_gt = sample_sniper_targets(
# proposal_boxes, gt_boxes, gt_labels, scale_index)
rcnn_boxes, rcnn_labels, fg_inds_wrt_gt = sample_sniper_targets(
proposal_boxes, gt_boxes, gt_labels)
else:
# The boxes to be used to crop RoIs.
# Use all proposal boxes in inference
rcnn_boxes = proposal_boxes
boxes_on_featuremap = rcnn_boxes * (1.0 / cfg.RPN.ANCHOR_STRIDE)
roi_resized = roi_align(featuremap, boxes_on_featuremap, 14)
feature_fastrcnn = resnet_conv5(
roi_resized, cfg.BACKBONE.RESNET_NUM_BLOCK[-1]) # nxcx7x7
# Keep C5 feature to be shared with mask branch
feature_gap = GlobalAvgPooling(
'gap', feature_fastrcnn, data_format='channels_first')
fastrcnn_label_logits, fastrcnn_box_logits = fastrcnn_outputs(
'fastrcnn', feature_gap, cfg.DATA.NUM_CLASS)
if is_training:
# rpn loss
rpn_label_loss, rpn_box_loss = rpn_losses(
anchors.gt_labels, anchors.encoded_gt_boxes(),
rpn_label_logits, rpn_box_logits)
# fastrcnn loss
matched_gt_boxes = tf.gather(gt_boxes, fg_inds_wrt_gt)
fg_inds_wrt_sample = tf.reshape(tf.where(rcnn_labels > 0),
[-1]) # fg inds w.r.t all samples
fg_sampled_boxes = tf.gather(rcnn_boxes, fg_inds_wrt_sample)
fg_fastrcnn_box_logits = tf.gather(fastrcnn_box_logits,
fg_inds_wrt_sample)
fastrcnn_label_loss, fastrcnn_box_loss = self.fastrcnn_training(
image, rcnn_labels, fg_sampled_boxes, matched_gt_boxes,
fastrcnn_label_logits, fg_fastrcnn_box_logits)
if cfg.MODE_MASK:
# maskrcnn loss
fg_labels = tf.gather(rcnn_labels, fg_inds_wrt_sample)
# In training, mask branch shares the same C5 feature.
fg_feature = tf.gather(feature_fastrcnn, fg_inds_wrt_sample)
mask_logits = maskrcnn_upXconv_head(
'maskrcnn', fg_feature, cfg.DATA.NUM_CATEGORY,
num_convs=0) # #fg x #cat x 14x14
target_masks_for_fg = crop_and_resize(
tf.expand_dims(gt_masks, 1),
fg_sampled_boxes,
fg_inds_wrt_gt,
14,
pad_border=False) # nfg x 1x14x14
target_masks_for_fg = tf.squeeze(target_masks_for_fg, 1,
'sampled_fg_mask_targets')
mrcnn_loss = maskrcnn_loss(mask_logits, fg_labels,
target_masks_for_fg)
else:
mrcnn_loss = 0.0
wd_cost = regularize_cost(
'(?:group1|group2|group3|rpn|fastrcnn|maskrcnn)/.*W',
l2_regularizer(cfg.TRAIN.WEIGHT_DECAY),
name='wd_cost')
total_cost = tf.add_n([
rpn_label_loss, rpn_box_loss, fastrcnn_label_loss,
fastrcnn_box_loss, mrcnn_loss, wd_cost
], 'total_cost')
add_moving_summary(total_cost, wd_cost)
return total_cost * (1. / cfg.TRAIN.NUM_GPUS)
else:
final_boxes, final_labels = self.fastrcnn_inference(
image_shape2d, rcnn_boxes, fastrcnn_label_logits,
fastrcnn_box_logits)
if cfg.MODE_MASK:
roi_resized = roi_align(
featuremap, final_boxes * (1.0 / cfg.RPN.ANCHOR_STRIDE),
14)
feature_maskrcnn = resnet_conv5(
roi_resized, cfg.BACKBONE.RESNET_NUM_BLOCK[-1])
mask_logits = maskrcnn_upXconv_head(
'maskrcnn', feature_maskrcnn, cfg.DATA.NUM_CATEGORY,
0) # #result x #cat x 14x14
indices = tf.stack(
[
tf.range(tf.size(final_labels)),
tf.to_int32(final_labels) - 1
],
axis=1)
final_mask_logits = tf.gather_nd(mask_logits,
indices) # #resultx14x14
tf.sigmoid(final_mask_logits, name='final_masks')
def build_graph(self, *inputs):
num_fpn_level = len(cfg.FPN.ANCHOR_STRIDES)
assert len(cfg.RPN.ANCHOR_SIZES) == num_fpn_level
is_training = get_current_tower_context().is_training
image = inputs[0]
input_anchors = inputs[1:1 + 2 * num_fpn_level]
multilevel_anchors = [
RPNAnchors(*args)
for args in zip(get_all_anchors_fpn(), input_anchors[0::2],
input_anchors[1::2])
]
gt_boxes, gt_labels = inputs[11], inputs[12]
if cfg.MODE_MASK:
gt_masks = inputs[-1]
image = self.preprocess(image) # 1CHW
image_shape2d = tf.shape(image)[2:] # h,w
c2345 = resnet_fpn_backbone(image, cfg.BACKBONE.RESNET_NUM_BLOCK)
p23456 = fpn_model('fpn', c2345)
self.slice_feature_and_anchors(image_shape2d, p23456,
multilevel_anchors)
# Multi-Level RPN Proposals
multilevel_proposals = []
rpn_loss_collection = []
for lvl in range(num_fpn_level):
rpn_label_logits, rpn_box_logits = rpn_head(
'rpn', p23456[lvl], cfg.FPN.NUM_CHANNEL,
len(cfg.RPN.ANCHOR_RATIOS))
with tf.name_scope('FPN_lvl{}'.format(lvl + 2)):
anchors = multilevel_anchors[lvl]
pred_boxes_decoded = anchors.decode_logits(rpn_box_logits)
proposal_boxes, proposal_scores = generate_rpn_proposals(
tf.reshape(pred_boxes_decoded, [-1, 4]),
tf.reshape(rpn_label_logits, [-1]), image_shape2d,
cfg.RPN.TRAIN_FPN_NMS_TOPK
if is_training else cfg.RPN.TEST_FPN_NMS_TOPK)
multilevel_proposals.append((proposal_boxes, proposal_scores))
if is_training:
label_loss, box_loss = rpn_losses(
anchors.gt_labels, anchors.encoded_gt_boxes(),
rpn_label_logits, rpn_box_logits)
rpn_loss_collection.extend([label_loss, box_loss])
# Merge proposals from multi levels, pick top K
proposal_boxes = tf.concat([x[0] for x in multilevel_proposals],
axis=0) # nx4
proposal_scores = tf.concat([x[1] for x in multilevel_proposals],
axis=0) # n
proposal_topk = tf.minimum(
tf.size(proposal_scores), cfg.RPN.TRAIN_FPN_NMS_TOPK
if is_training else cfg.RPN.TEST_FPN_NMS_TOPK)
proposal_scores, topk_indices = tf.nn.top_k(proposal_scores,
k=proposal_topk,
sorted=False)
proposal_boxes = tf.gather(proposal_boxes, topk_indices)
if is_training:
rcnn_boxes, rcnn_labels, fg_inds_wrt_gt = sample_fast_rcnn_targets(
proposal_boxes, gt_boxes, gt_labels)
else:
# The boxes to be used to crop RoIs.
rcnn_boxes = proposal_boxes
roi_feature_fastrcnn = multilevel_roi_align(p23456[:4], rcnn_boxes, 7)
fastrcnn_head_func = getattr(model, cfg.FPN.FRCNN_HEAD_FUNC)
fastrcnn_label_logits, fastrcnn_box_logits = fastrcnn_head_func(
'fastrcnn', roi_feature_fastrcnn, cfg.DATA.NUM_CLASS)
if is_training:
# rpn loss is already defined above
with tf.name_scope('rpn_losses'):
rpn_total_label_loss = tf.add_n(rpn_loss_collection[::2],
name='label_loss')
rpn_total_box_loss = tf.add_n(rpn_loss_collection[1::2],
name='box_loss')
add_moving_summary(rpn_total_box_loss, rpn_total_label_loss)
# fastrcnn loss:
matched_gt_boxes = tf.gather(gt_boxes, fg_inds_wrt_gt)
fg_inds_wrt_sample = tf.reshape(tf.where(rcnn_labels > 0),
[-1]) # fg inds w.r.t all samples
fg_sampled_boxes = tf.gather(rcnn_boxes, fg_inds_wrt_sample)
fg_fastrcnn_box_logits = tf.gather(fastrcnn_box_logits,
fg_inds_wrt_sample)
fastrcnn_label_loss, fastrcnn_box_loss = self.fastrcnn_training(
image, rcnn_labels, fg_sampled_boxes, matched_gt_boxes,
fastrcnn_label_logits, fg_fastrcnn_box_logits)
if cfg.MODE_MASK:
# maskrcnn loss
fg_labels = tf.gather(rcnn_labels, fg_inds_wrt_sample)
roi_feature_maskrcnn = multilevel_roi_align(
p23456[:4], fg_sampled_boxes, 14)
mask_logits = maskrcnn_upXconv_head('maskrcnn',
roi_feature_maskrcnn,
cfg.DATA.NUM_CATEGORY,
4) # #fg x #cat x 28 x 28
target_masks_for_fg = crop_and_resize(
tf.expand_dims(gt_masks, 1),
fg_sampled_boxes,
fg_inds_wrt_gt,
28,
pad_border=False) # fg x 1x28x28
target_masks_for_fg = tf.squeeze(target_masks_for_fg, 1,
'sampled_fg_mask_targets')
mrcnn_loss = maskrcnn_loss(mask_logits, fg_labels,
target_masks_for_fg)
else:
mrcnn_loss = 0.0
wd_cost = regularize_cost(
'(?:group1|group2|group3|rpn|fpn|fastrcnn|maskrcnn)/.*W',
l2_regularizer(cfg.TRAIN.WEIGHT_DECAY),
name='wd_cost')
total_cost = tf.add_n(
rpn_loss_collection +
[fastrcnn_label_loss, fastrcnn_box_loss, mrcnn_loss, wd_cost],
'total_cost')
add_moving_summary(total_cost, wd_cost)
return total_cost
else:
final_boxes, final_labels = self.fastrcnn_inference(
image_shape2d, rcnn_boxes, fastrcnn_label_logits,
fastrcnn_box_logits)
if cfg.MODE_MASK:
# Cascade inference needs roi transform with refined boxes.
roi_feature_maskrcnn = multilevel_roi_align(
p23456[:4], final_boxes, 14)
mask_logits = maskrcnn_upXconv_head('maskrcnn',
roi_feature_maskrcnn,
cfg.DATA.NUM_CATEGORY,
4) # #fg x #cat x 28 x 28
indices = tf.stack([
tf.range(tf.size(final_labels)),
tf.to_int32(final_labels) - 1
],
axis=1)
final_mask_logits = tf.gather_nd(mask_logits,
indices) # #resultx28x28
tf.sigmoid(final_mask_logits, name='final_masks')
def build_graph(self, *inputs):
inputs = dict(zip(self.input_names, inputs))
is_training = get_current_tower_context().is_training
image = self.preprocess(inputs['image']) # 1CHW
featuremap = resnet_c4_backbone(image, cfg.BACKBONE.RESNET_NUM_BLOCK[:3])
rpn_label_logits, rpn_box_logits = rpn_head('rpn', featuremap, cfg.RPN.HEAD_DIM, cfg.RPN.NUM_ANCHOR)
anchors = RPNAnchors(get_all_anchors(), inputs['anchor_labels'], inputs['anchor_boxes'])
anchors = anchors.narrow_to(featuremap)
image_shape2d = tf.shape(image)[2:] # h,w
pred_boxes_decoded = anchors.decode_logits(rpn_box_logits) # fHxfWxNAx4, floatbox
proposal_boxes, proposal_scores = generate_rpn_proposals(
tf.reshape(pred_boxes_decoded, [-1, 4]),
tf.reshape(rpn_label_logits, [-1]),
image_shape2d,
cfg.RPN.TRAIN_PRE_NMS_TOPK if is_training else cfg.RPN.TEST_PRE_NMS_TOPK,
cfg.RPN.TRAIN_POST_NMS_TOPK if is_training else cfg.RPN.TEST_POST_NMS_TOPK)
gt_boxes, gt_labels = inputs['gt_boxes'], inputs['gt_labels']
if is_training:
# sample proposal boxes in training
proposals = sample_fast_rcnn_targets(proposal_boxes, gt_boxes, gt_labels)
else:
# The boxes to be used to crop RoIs.
# Use all proposal boxes in inference
proposals = BoxProposals(proposal_boxes)
boxes_on_featuremap = proposals.boxes * (1.0 / cfg.RPN.ANCHOR_STRIDE)
roi_resized = roi_align(featuremap, boxes_on_featuremap, 14)
feature_fastrcnn = resnet_conv5(roi_resized, cfg.BACKBONE.RESNET_NUM_BLOCK[-1]) # nxcx7x7
# Keep C5 feature to be shared with mask branch
feature_gap = GlobalAvgPooling('gap', feature_fastrcnn, data_format='channels_first')
fastrcnn_label_logits, fastrcnn_box_logits = fastrcnn_outputs('fastrcnn', feature_gap, cfg.DATA.NUM_CLASS)
fastrcnn_head = FastRCNNHead(proposals, fastrcnn_box_logits, fastrcnn_label_logits,
tf.constant(cfg.FRCNN.BBOX_REG_WEIGHTS, dtype=tf.float32))
if is_training:
all_losses = []
# rpn loss
all_losses.extend(rpn_losses(
anchors.gt_labels, anchors.encoded_gt_boxes(), rpn_label_logits, rpn_box_logits))
# fastrcnn loss
all_losses.extend(fastrcnn_head.losses())
if cfg.MODE_MASK:
# maskrcnn loss
# In training, mask branch shares the same C5 feature.
fg_feature = tf.gather(feature_fastrcnn, proposals.fg_inds())
mask_logits = maskrcnn_upXconv_head(
'maskrcnn', fg_feature, cfg.DATA.NUM_CATEGORY, num_convs=0) # #fg x #cat x 14x14
target_masks_for_fg = crop_and_resize(
tf.expand_dims(inputs['gt_masks'], 1),
proposals.fg_boxes(),
proposals.fg_inds_wrt_gt, 14,
pad_border=False) # nfg x 1x14x14
target_masks_for_fg = tf.squeeze(target_masks_for_fg, 1, 'sampled_fg_mask_targets')
all_losses.append(maskrcnn_loss(mask_logits, proposals.fg_labels(), target_masks_for_fg))
wd_cost = regularize_cost(
'.*/W', l2_regularizer(cfg.TRAIN.WEIGHT_DECAY), name='wd_cost')
all_losses.append(wd_cost)
total_cost = tf.add_n(all_losses, 'total_cost')
add_moving_summary(total_cost, wd_cost)
return total_cost
else:
decoded_boxes = fastrcnn_head.decoded_output_boxes()
decoded_boxes = clip_boxes(decoded_boxes, image_shape2d, name='fastrcnn_all_boxes')
label_scores = fastrcnn_head.output_scores(name='fastrcnn_all_scores')
final_boxes, final_scores, final_labels = fastrcnn_predictions(
decoded_boxes, label_scores, name_scope='output')
if cfg.MODE_MASK:
roi_resized = roi_align(featuremap, final_boxes * (1.0 / cfg.RPN.ANCHOR_STRIDE), 14)
feature_maskrcnn = resnet_conv5(roi_resized, cfg.BACKBONE.RESNET_NUM_BLOCK[-1])
mask_logits = maskrcnn_upXconv_head(
'maskrcnn', feature_maskrcnn, cfg.DATA.NUM_CATEGORY, 0) # #result x #cat x 14x14
indices = tf.stack([tf.range(tf.size(final_labels)), tf.to_int32(final_labels) - 1], axis=1)
final_mask_logits = tf.gather_nd(mask_logits, indices) # #resultx14x14
tf.sigmoid(final_mask_logits, name='output/masks')