def build_outputs(self, features, labels, mode):
is_training = (mode == mode_keys.TRAIN)
backbone_features = self._backbone_fn(features,
is_training=is_training)
fpn_features = self._fpn_fn(backbone_features, is_training=is_training)
cls_outputs, box_outputs = self._retinanet_head_fn(
fpn_features, is_training=is_training)
# Shapemask mask prediction.
if is_training:
boxes = labels['mask_boxes']
outer_boxes = labels['mask_outer_boxes']
classes = labels['mask_classes']
else:
detection_results = self._generate_detections_fn(
box_outputs, cls_outputs, labels['anchor_boxes'],
labels['image_info'][:, 1:2, :])
boxes = detection_results['detection_boxes']
scores = detection_results['detection_scores']
classes = detection_results['detection_classes']
valid_detections = detection_results['num_detections']
# Use list as input to avoide segmentation fault on TPU.
feature_size = features.get_shape().as_list()[1:3]
outer_boxes = box_utils.compute_outer_boxes(
tf.reshape(boxes, [-1, 4]),
feature_size,
scale=self._outer_box_scale)
outer_boxes = tf.reshape(outer_boxes, tf.shape(boxes))
classes = tf.cast(classes, tf.int32)
instance_features, prior_masks = self._shape_prior_head_fn(
fpn_features, boxes, outer_boxes, classes, is_training)
coarse_mask_logits = self._coarse_mask_fn(instance_features,
prior_masks, classes,
is_training)
fine_mask_logits = self._fine_mask_fn(instance_features,
coarse_mask_logits, classes,
is_training)
model_outputs = {
'cls_outputs': cls_outputs,
'box_outputs': box_outputs,
'fine_mask_logits': fine_mask_logits,
'coarse_mask_logits': coarse_mask_logits,
'prior_masks': prior_masks,
}
self._log_model_statistics(features)
if not is_training:
model_outputs.update({
'num_detections': valid_detections,
'detection_boxes': boxes,
'detection_outer_boxes': outer_boxes,
'detection_masks': fine_mask_logits,
'detection_classes': classes,
'detection_scores': scores,
})
return model_outputs
def _build_outputs(self, images, labels, mode):
is_training = (mode == mode_keys.TRAIN)
if 'anchor_boxes' in labels:
anchor_boxes = labels['anchor_boxes']
else:
anchor_boxes = anchor.Anchor(
self._params.architecture.min_level,
self._params.architecture.max_level,
self._params.anchor.num_scales,
self._params.anchor.aspect_ratios,
self._params.anchor.anchor_size,
images.get_shape().as_list()[1:3]).multilevel_boxes
batch_size = tf.shape(images)[0]
for level in anchor_boxes:
anchor_boxes[level] = tf.tile(
tf.expand_dims(anchor_boxes[level], 0), [batch_size, 1, 1])
backbone_features = self._backbone_fn(images, is_training=is_training)
fpn_features = self._fpn_fn(backbone_features, is_training=is_training)
cls_outputs, box_outputs = self._retinanet_head_fn(
fpn_features, is_training=is_training)
# Shapemask mask prediction.
if is_training:
boxes = labels['mask_boxes']
outer_boxes = labels['mask_outer_boxes']
classes = labels['mask_classes']
else:
detection_results = self._generate_detections_fn(
box_outputs, cls_outputs, anchor_boxes,
labels['image_info'][:, 1:2, :])
boxes = detection_results['detection_boxes']
scores = detection_results['detection_scores']
classes = detection_results['detection_classes']
valid_detections = detection_results['num_detections']
# Use list as input to avoide segmentation fault on TPU.
image_size = images.get_shape().as_list()[1:3]
outer_boxes = box_utils.compute_outer_boxes(
tf.reshape(boxes, [-1, 4]),
image_size,
scale=self._outer_box_scale)
outer_boxes = tf.reshape(outer_boxes, tf.shape(boxes))
classes = tf.cast(classes, tf.int32)
instance_features, prior_masks = self._shape_prior_head_fn(
fpn_features, boxes, outer_boxes, classes, is_training)
coarse_mask_logits = self._coarse_mask_fn(instance_features,
prior_masks, classes,
is_training)
fine_mask_logits = self._fine_mask_fn(instance_features,
coarse_mask_logits, classes,
is_training)
model_outputs = {
'cls_outputs': cls_outputs,
'box_outputs': box_outputs,
'fine_mask_logits': fine_mask_logits,
'coarse_mask_logits': coarse_mask_logits,
'prior_masks': prior_masks,
'fpn_features': fpn_features,
}
if not is_training:
model_outputs.update({
'num_detections':
valid_detections,
'detection_boxes':
boxes,
'detection_outer_boxes':
outer_boxes,
'detection_masks':
tf.sigmoid(fine_mask_logits),
'detection_classes':
tf.cast(classes, dtype=tf.int32),
'detection_scores':
scores,
})
return model_outputs
def parse_train_data(self, data):
"""Parse data for ShapeMask training."""
classes = data['groundtruth_classes']
boxes = data['groundtruth_boxes']
masks = data['groundtruth_instance_masks']
is_crowds = data['groundtruth_is_crowd']
# Skips annotations with `is_crowd` = True.
if self._skip_crowd_during_training and self._is_training:
num_groundtrtuhs = tf.shape(classes)[0]
with tf.control_dependencies([num_groundtrtuhs, is_crowds]):
indices = tf.cond(
tf.greater(tf.size(is_crowds), 0),
lambda: tf.where(tf.logical_not(is_crowds))[:, 0],
lambda: tf.cast(tf.range(num_groundtrtuhs), tf.int64))
classes = tf.gather(classes, indices)
boxes = tf.gather(boxes, indices)
masks = tf.gather(masks, indices)
# If not using category, makes all categories with id = 0.
if not self._use_category:
classes = tf.cast(tf.greater(classes, 0), dtype=tf.float32)
image = self.get_normalized_image(data)
# Flips image randomly during training.
if self._aug_rand_hflip:
image, boxes, masks = input_utils.random_horizontal_flip(
image, boxes, masks)
# Converts boxes from normalized coordinates to pixel coordinates.
image_shape = tf.shape(image)[0:2]
boxes = box_utils.denormalize_boxes(boxes, image_shape)
# Resizes and crops image.
image, image_info = input_utils.resize_and_crop_image(
image,
self._output_size,
self._output_size,
aug_scale_min=self._aug_scale_min,
aug_scale_max=self._aug_scale_max)
self._train_image_scale = image_info[2, :]
self._train_offset = image_info[3, :]
# Resizes and crops boxes and masks.
boxes = input_utils.resize_and_crop_boxes(boxes,
self._train_image_scale,
image_info[1, :],
self._train_offset)
# Filters out ground truth boxes that are all zeros.
indices = box_utils.get_non_empty_box_indices(boxes)
boxes = tf.gather(boxes, indices)
classes = tf.gather(classes, indices)
masks = tf.gather(masks, indices)
# Assigns anchors.
input_anchor = anchor.Anchor(self._min_level, self._max_level,
self._num_scales, self._aspect_ratios,
self._anchor_size, self._output_size)
anchor_labeler = anchor.AnchorLabeler(input_anchor,
self._match_threshold,
self._unmatched_threshold)
(cls_targets, box_targets,
num_positives) = anchor_labeler.label_anchors(
boxes, tf.cast(tf.expand_dims(classes, axis=1), tf.float32))
# Sample groundtruth masks/boxes/classes for mask branch.
num_masks = tf.shape(masks)[0]
mask_shape = tf.shape(masks)[1:3]
# Pad sampled boxes/masks/classes to a constant batch size.
padded_boxes = input_utils.pad_to_fixed_size(boxes,
self._num_sampled_masks)
padded_classes = input_utils.pad_to_fixed_size(classes,
self._num_sampled_masks)
padded_masks = input_utils.pad_to_fixed_size(masks,
self._num_sampled_masks)
# Randomly sample groundtruth masks for mask branch training. For the image
# without groundtruth masks, it will sample the dummy padded tensors.
rand_indices = tf.random.shuffle(
tf.range(tf.maximum(num_masks, self._num_sampled_masks)))
rand_indices = tf.mod(rand_indices, tf.maximum(num_masks, 1))
rand_indices = rand_indices[0:self._num_sampled_masks]
rand_indices = tf.reshape(rand_indices, [self._num_sampled_masks])
sampled_boxes = tf.gather(padded_boxes, rand_indices)
sampled_classes = tf.gather(padded_classes, rand_indices)
sampled_masks = tf.gather(padded_masks, rand_indices)
# Jitter the sampled boxes to mimic the noisy detections.
sampled_boxes = box_utils.jitter_boxes(
sampled_boxes, noise_scale=self._box_jitter_scale)
sampled_boxes = box_utils.clip_boxes(sampled_boxes, self._output_size)
# Compute mask targets in feature crop. A feature crop fully contains a
# sampled box.
mask_outer_boxes = box_utils.compute_outer_boxes(
sampled_boxes, tf.shape(image)[0:2], scale=self._outer_box_scale)
mask_outer_boxes = box_utils.clip_boxes(mask_outer_boxes,
self._output_size)
# Compensate the offset of mask_outer_boxes to map it back to original image
# scale.
mask_outer_boxes_ori = mask_outer_boxes
mask_outer_boxes_ori += tf.tile(
tf.expand_dims(self._train_offset, axis=0), [1, 2])
mask_outer_boxes_ori /= tf.tile(
tf.expand_dims(self._train_image_scale, axis=0), [1, 2])
norm_mask_outer_boxes_ori = box_utils.normalize_boxes(
mask_outer_boxes_ori, mask_shape)
# Set sampled_masks shape to [batch_size, height, width, 1].
sampled_masks = tf.cast(tf.expand_dims(sampled_masks, axis=-1),
tf.float32)
mask_targets = tf.image.crop_and_resize(
sampled_masks,
norm_mask_outer_boxes_ori,
box_ind=tf.range(self._num_sampled_masks),
crop_size=[self._mask_crop_size, self._mask_crop_size],
method='bilinear',
extrapolation_value=0,
name='train_mask_targets')
mask_targets = tf.where(tf.greater_equal(mask_targets, 0.5),
tf.ones_like(mask_targets),
tf.zeros_like(mask_targets))
mask_targets = tf.squeeze(mask_targets, axis=-1)
if self._up_sample_factor > 1:
fine_mask_targets = tf.image.crop_and_resize(
sampled_masks,
norm_mask_outer_boxes_ori,
box_ind=tf.range(self._num_sampled_masks),
crop_size=[
self._mask_crop_size * self._up_sample_factor,
self._mask_crop_size * self._up_sample_factor
],
method='bilinear',
extrapolation_value=0,
name='train_mask_targets')
fine_mask_targets = tf.where(
tf.greater_equal(fine_mask_targets, 0.5),
tf.ones_like(fine_mask_targets),
tf.zeros_like(fine_mask_targets))
fine_mask_targets = tf.squeeze(fine_mask_targets, axis=-1)
else:
fine_mask_targets = mask_targets
# If bfloat16 is used, casts input image to tf.bfloat16.
if self._use_bfloat16:
image = tf.cast(image, dtype=tf.bfloat16)
valid_image = tf.cast(tf.not_equal(num_masks, 0), tf.int32)
if self._mask_train_class == 'all':
mask_is_valid = valid_image * tf.ones_like(sampled_classes,
tf.int32)
else:
# Get the intersection of sampled classes with training splits.
mask_valid_classes = tf.cast(
tf.expand_dims(
class_utils.coco_split_class_ids(self._mask_train_class),
1), sampled_classes.dtype)
match = tf.reduce_any(
tf.equal(tf.expand_dims(sampled_classes, 0),
mask_valid_classes), 0)
mask_is_valid = valid_image * tf.cast(match, tf.int32)
# Packs labels for model_fn outputs.
labels = {
'cls_targets': cls_targets,
'box_targets': box_targets,
'anchor_boxes': input_anchor.multilevel_boxes,
'num_positives': num_positives,
'image_info': image_info,
# For ShapeMask.
'mask_boxes': sampled_boxes,
'mask_outer_boxes': mask_outer_boxes,
'mask_targets': mask_targets,
'fine_mask_targets': fine_mask_targets,
'mask_classes': sampled_classes,
'mask_is_valid': mask_is_valid,
}
return image, labels
def _parse_train_data(self, data):
"""Parse data for ShapeMask training."""
classes = data['groundtruth_classes']
boxes = data['groundtruth_boxes']
masks = data['groundtruth_instance_masks']
is_crowds = data['groundtruth_is_crowd']
# Skips annotations with `is_crowd` = True.
if self._skip_crowd_during_training and self._is_training:
num_groundtrtuhs = tf.shape(classes)[0]
with tf.control_dependencies([num_groundtrtuhs, is_crowds]):
indices = tf.cond(
tf.greater(tf.size(is_crowds), 0),
lambda: tf.where(tf.logical_not(is_crowds))[:, 0],
lambda: tf.cast(tf.range(num_groundtrtuhs), tf.int64))
classes = tf.gather(classes, indices)
boxes = tf.gather(boxes, indices)
masks = tf.gather(masks, indices)
# Gets original image and its size.
image = data['image']
image_shape = tf.shape(image)[0:2]
# If not using category, makes all categories with id = 0.
if not self._use_category:
classes = tf.cast(tf.greater(classes, 0), dtype=tf.float32)
# Normalizes image with mean and std pixel values.
image = input_utils.normalize_image(image)
# Flips image randomly during training.
if self._aug_rand_hflip:
image, boxes, masks = input_utils.random_horizontal_flip(
image, boxes, masks)
# Converts boxes from normalized coordinates to pixel coordinates.
boxes = box_utils.denormalize_boxes(boxes, image_shape)
# Resizes and crops image.
image, image_info = input_utils.resize_and_crop_image(
image,
self._output_size,
self._output_size,
aug_scale_min=self._aug_scale_min,
aug_scale_max=self._aug_scale_max)
image_scale = image_info[2, :]
offset = image_info[3, :]
# Resizes and crops boxes and masks.
boxes = input_utils.resize_and_crop_boxes(boxes, image_scale,
self._output_size, offset)
masks = input_utils.resize_and_crop_masks(
tf.expand_dims(masks, axis=-1), image_scale, self._output_size,
offset)
masks = tf.squeeze(masks, axis=-1)
# Filters out ground truth boxes that are all zeros.
indices = input_utils.get_non_empty_box_indices(boxes)
boxes = tf.gather(boxes, indices)
classes = tf.gather(classes, indices)
masks = tf.gather(masks, indices)
# Assigns anchors.
input_anchor = anchor.Anchor(self._min_level, self._max_level,
self._num_scales, self._aspect_ratios,
self._anchor_size, self._output_size)
anchor_labeler = anchor.AnchorLabeler(input_anchor,
self._match_threshold,
self._unmatched_threshold)
(cls_targets, box_targets,
num_positives) = anchor_labeler.label_anchors(
boxes, tf.cast(tf.expand_dims(classes, axis=1), tf.float32))
# Sample groundtruth masks/boxes/classes for mask branch.
num_masks = tf.shape(masks)[0]
mask_shape = tf.shape(masks)[1:3]
# Pad sampled boxes/masks/classes to a constant batch size.
padded_boxes = input_utils.pad_to_fixed_size(boxes,
self._num_sampled_masks)
padded_classes = input_utils.pad_to_fixed_size(classes,
self._num_sampled_masks)
padded_masks = input_utils.pad_to_fixed_size(masks,
self._num_sampled_masks)
# Randomly sample groundtruth masks for mask branch training. For the image
# without groundtruth masks, it will sample the dummy padded tensors.
rand_indices = tf.random.uniform([self._num_sampled_masks],
minval=0,
maxval=tf.maximum(num_masks, 1),
dtype=tf.dtypes.int32)
sampled_boxes = tf.gather(padded_boxes, rand_indices)
sampled_classes = tf.gather(padded_classes, rand_indices)
sampled_masks = tf.gather(padded_masks, rand_indices)
# Compute mask targets in feature crop. A feature crop fully contains a
# sampled box.
mask_outer_boxes = box_utils.compute_outer_boxes(
sampled_boxes, mask_shape, scale=self._outer_box_scale)
norm_mask_outer_boxes = box_utils.normalize_boxes(
mask_outer_boxes, mask_shape)
# Set sampled_masks shape to [batch_size, height, width, 1].
sampled_masks = tf.expand_dims(sampled_masks, axis=-1)
mask_targets = tf.image.crop_and_resize(
sampled_masks,
norm_mask_outer_boxes,
box_ind=tf.range(self._num_sampled_masks),
crop_size=[self._mask_crop_size, self._mask_crop_size],
method='bilinear',
extrapolation_value=0,
name='train_mask_targets')
mask_targets = tf.where(tf.greater_equal(mask_targets, 0.5),
tf.ones_like(mask_targets),
tf.zeros_like(mask_targets))
mask_targets = tf.squeeze(mask_targets, axis=-1)
# If bfloat16 is used, casts input image to tf.bfloat16.
if self._use_bfloat16:
image = tf.cast(image, dtype=tf.bfloat16)
# Packs labels for model_fn outputs.
labels = {
'cls_targets': cls_targets,
'box_targets': box_targets,
'anchor_boxes': input_anchor.multilevel_boxes,
'num_positives': num_positives,
'image_info': image_info,
# For ShapeMask.
'mask_boxes': sampled_boxes,
'mask_outer_boxes': mask_outer_boxes,
'mask_targets': mask_targets,
'mask_classes': sampled_classes,
'mask_is_valid': tf.cast(tf.not_equal(num_masks, 0), tf.int32)
}
return image, labels
