def _parse_train_data(self, data):
"""Parses data for training.
Args:
data: the decoded tensor dictionary from TfExampleDecoder.
Returns:
image: image tensor that is preproessed to have normalized value and
dimension [output_size[0], output_size[1], 3]
labels: a dictionary of tensors used for training. The following describes
{key: value} pairs in the dictionary.
image_info: a 2D `Tensor` that encodes the information of the image and
the applied preprocessing. It is in the format of
[[original_height, original_width], [scaled_height, scaled_width],
anchor_boxes: ordered dictionary with keys [min_level, min_level+1, ..., max_level].
The values are tensor with shape [height_l, width_l, 4] representing anchor boxes at each level.
rpn_score_targets: ordered dictionary with keys [min_level, min_level+1, ..., max_level].
The values are tensor with shape [height_l, width_l, anchors_per_location]. The height_l and
width_l represent the dimension of class logits at l-th level.
rpn_box_targets: ordered dictionary with keys [min_level, min_level+1, ..., max_level].
The values are tensor with shape [height_l, width_l, anchors_per_location * 4]. The height_l and
width_l represent the dimension of bounding box regression output at
l-th level.
gt_boxes: Groundtruth bounding box annotations. The box is represented in [y1, x1, y2, x2] format.
The coordinates are w.r.t the scaled image that is fed to the network. The tennsor is
padded with -1 to the fixed dimension [self._max_num_instances, 4].
gt_classes: Groundtruth classes annotations. The tennsor is padded with -1 to the fixed
dimension [self._max_num_instances].
gt_masks: groundtrugh masks cropped by the bounding box and resized to a fixed size
determined by mask_crop_size.
"""
classes = data['groundtruth_classes']
boxes = data['groundtruth_boxes']
is_crowds = data['groundtruth_is_crowd']
if self._include_mask:
masks = data['groundtruth_instance_masks']
# Skips annotations with `is_crowd` = True.
if self._skip_crowd_during_training and self._is_training:
num_groundtruths = tf.shape(classes)[0]
with tf.control_dependencies([num_groundtruths, 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_groundtruths), tf.int64))
classes = tf.gather(classes, indices, axis=None)
boxes = tf.gather(boxes, indices, axis=None)
if self._include_mask:
masks = tf.gather(masks, indices, axis=None)
# Gets original image and its size.
image = data['image']
image_shape = tf.shape(image)[0:2]
# Normalizes image with mean and std pixel values.
image = input_utils.normalize_image(image)
# Flips image randomly during training.
if self._aug_rand_hflip:
if self._include_mask:
image, boxes, masks = input_utils.random_horizontal_flip(image, boxes, masks) # pylint: disable=W0632
else:
image, boxes = input_utils.random_horizontal_flip(image, boxes) # pylint: disable=W0632
# Converts boxes from normalized coordinates to pixel coordinates.
# Now the coordinates of boxes are w.r.t. the original image.
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,
padded_size=input_utils.compute_padded_size(self._output_size, 2 ** self._max_level),
aug_scale_min=self._aug_scale_min,
aug_scale_max=self._aug_scale_max)
image_height, image_width, _ = image.get_shape().as_list()
# Resizes and crops boxes.
# Now the coordinates of boxes are w.r.t the scaled image.
image_scale = image_info[2, :]
offset = image_info[3, :]
boxes = input_utils.resize_and_crop_boxes(boxes, image_scale, image_info[1, :], offset)
# Filters out ground truth boxes that are all zeros.
indices = box_utils.get_non_empty_box_indices(boxes)
boxes = tf.gather(boxes, indices, axis=None)
classes = tf.gather(classes, indices, axis=None)
if self._include_mask:
masks = tf.gather(masks, indices, axis=None)
# Transfer boxes to the original image space and do normalization.
cropped_boxes = boxes + tf.tile(tf.expand_dims(offset, axis=0), [1, 2])
cropped_boxes /= tf.tile(tf.expand_dims(image_scale, axis=0), [1, 2])
cropped_boxes = box_utils.normalize_boxes(cropped_boxes, image_shape)
num_masks = tf.shape(masks)[0]
masks = tf.image.crop_and_resize(
tf.expand_dims(masks, axis=-1),
cropped_boxes,
box_indices=tf.range(num_masks, dtype=tf.int32),
crop_size=[self._mask_crop_size, self._mask_crop_size],
method='bilinear')
masks = tf.squeeze(masks, axis=-1)
# Assigns anchor targets.
# Note that after the target assignment, box targets are absolute pixel
# offsets w.r.t. the scaled image.
input_anchor = anchor.Anchor(self._min_level, self._max_level,
self._num_scales, self._aspect_ratios,
self._anchor_size, (image_height, image_width))
anchor_labeler = anchor.RpnAnchorLabeler(input_anchor,
self._rpn_match_threshold,
self._rpn_unmatched_threshold,
self._rpn_batch_size_per_im,
self._rpn_fg_fraction)
rpn_score_targets, rpn_box_targets = anchor_labeler.label_anchors(
boxes, tf.cast(tf.expand_dims(classes, axis=-1), tf.float32))
inputs = {
'image': image,
'image_info': image_info,
}
# Packs labels for model_fn outputs.
labels = {
'anchor_boxes': input_anchor.multilevel_boxes,
'image_info': image_info,
'rpn_score_targets': rpn_score_targets,
'rpn_box_targets': rpn_box_targets,
}
inputs['gt_boxes'] = input_utils.pad_to_fixed_size(boxes, self._max_num_instances, -1)
inputs['gt_classes'] = input_utils.pad_to_fixed_size(classes, self._max_num_instances, -1)
if self._include_mask:
inputs['gt_masks'] = input_utils.pad_to_fixed_size(masks, self._max_num_instances, -1)
return inputs, labels
def _parse_train_data(self, data):
"""Parses data for training and evaluation."""
classes = data['groundtruth_classes']
boxes = data['groundtruth_boxes']
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(input=classes)[0]
with tf.control_dependencies([num_groundtrtuhs, is_crowds]):
indices = tf.cond(
pred=tf.greater(tf.size(input=is_crowds), 0),
true_fn=lambda: tf.where(tf.logical_not(is_crowds))[:, 0],
false_fn=lambda: tf.cast(tf.range(num_groundtrtuhs), tf.
int64))
classes = tf.gather(classes, indices, axis=None)
boxes = tf.gather(boxes, indices, axis=None)
# Gets original image and its size.
image = data['image']
image_shape = tf.shape(input=image)[0:2]
# 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 = input_utils.random_horizontal_flip(image, boxes) # pylint: disable=W0632
# 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,
padded_size=input_utils.compute_padded_size(
self._output_size, 2**self._max_level),
aug_scale_min=self._aug_scale_min,
aug_scale_max=self._aug_scale_max)
image_height, image_width, _ = image.get_shape().as_list()
# Resizes and crops boxes.
image_scale = image_info[2, :]
offset = image_info[3, :]
boxes = input_utils.resize_and_crop_boxes(boxes, image_scale,
image_info[1, :], offset)
# Filters out ground truth boxes that are all zeros.
indices = box_utils.get_non_empty_box_indices(boxes)
boxes = tf.gather(boxes, indices, axis=None)
classes = tf.gather(classes, indices, axis=None)
# Assigns anchors.
input_anchor = anchor.Anchor(self._min_level, self._max_level,
self._num_scales, self._aspect_ratios,
self._anchor_size,
(image_height, image_width))
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))
# 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,
}
return image, labels
def _parse_predict_data(self, data):
"""Parses data for prediction."""
# Gets original image and its size.
image = data['image']
image_shape = tf.shape(input=image)[0:2]
# Normalizes image with mean and std pixel values.
image = input_utils.normalize_image(image)
# Resizes and crops image.
image, image_info = input_utils.resize_and_crop_image(
image,
self._output_size,
padded_size=input_utils.compute_padded_size(
self._output_size, 2**self._max_level),
aug_scale_min=1.0,
aug_scale_max=1.0)
image_height, image_width, _ = image.get_shape().as_list()
# Compute Anchor boxes.
input_anchor = anchor.Anchor(self._min_level, self._max_level,
self._num_scales, self._aspect_ratios,
self._anchor_size,
(image_height, image_width))
labels = {
'anchor_boxes': input_anchor.multilevel_boxes,
'image_info': image_info,
}
# Converts boxes from normalized coordinates to pixel coordinates.
boxes = box_utils.denormalize_boxes(data['groundtruth_boxes'],
image_shape)
groundtruths = {
'source_id': data['source_id'],
'num_detections':
tf.squeeze(tf.shape(data['groundtruth_classes'])),
'boxes': boxes,
'classes': data['groundtruth_classes'],
'areas': data['groundtruth_area'],
'is_crowds': tf.cast(data['groundtruth_is_crowd'], tf.int32),
}
groundtruths['source_id'] = dataloader_utils.process_source_id(
groundtruths['source_id'])
groundtruths = dataloader_utils.pad_groundtruths_to_fixed_size(
groundtruths, self._max_num_instances)
labels.update(groundtruths)
# Computes training objective for evaluation loss.
classes = data['groundtruth_classes']
image_scale = image_info[2, :]
offset = image_info[3, :]
boxes = input_utils.resize_and_crop_boxes(boxes, image_scale,
image_info[1, :], offset)
# Filters out ground truth boxes that are all zeros.
indices = box_utils.get_non_empty_box_indices(boxes)
boxes = tf.gather(boxes, indices, axis=None)
# Assigns anchors.
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))
labels['cls_targets'] = cls_targets
labels['box_targets'] = box_targets
labels['num_positives'] = num_positives
return image, labels
def build_outputs(self, inputs, is_training):
model_outputs = {}
image = inputs['image']
_, image_height, image_width, _ = image.get_shape().as_list()
backbone_features = self._backbone_fn(image, is_training)
fpn_features = self._fpn_fn(backbone_features, is_training)
rpn_score_outputs, rpn_box_outputs = self._rpn_head_fn(
fpn_features, is_training)
model_outputs.update({
'rpn_score_outputs':
tf.nest.map_structure(lambda x: tf.cast(x, tf.float32),
rpn_score_outputs),
'rpn_box_outputs':
tf.nest.map_structure(lambda x: tf.cast(x, tf.float32),
rpn_box_outputs),
})
input_anchor = anchor.Anchor(
self._params.model_params.architecture.min_level,
self._params.model_params.architecture.max_level,
self._params.anchor.num_scales, self._params.anchor.aspect_ratios,
self._params.anchor.anchor_size, (image_height, image_width))
rpn_rois, _ = self._generate_rois_fn(rpn_box_outputs,
rpn_score_outputs,
input_anchor.multilevel_boxes,
inputs['image_info'][:, 1, :],
is_training)
if is_training:
rpn_rois = tf.stop_gradient(rpn_rois)
# Sample proposals.
rpn_rois, matched_gt_boxes, matched_gt_classes, matched_gt_indices = (
self._sample_rois_fn(rpn_rois, inputs['gt_boxes'],
inputs['gt_classes']))
# Create bounding box training targets.
box_targets = box_utils.encode_boxes(
matched_gt_boxes, rpn_rois, weights=[10.0, 10.0, 5.0, 5.0])
# If the target is background, the box target is set to all 0s.
box_targets = tf.where(
tf.tile(
tf.expand_dims(tf.equal(matched_gt_classes, 0), axis=-1),
[1, 1, 4]), tf.zeros_like(box_targets), box_targets)
model_outputs.update({
'class_targets': matched_gt_classes,
'box_targets': box_targets,
})
roi_features = spatial_transform_ops.multilevel_crop_and_resize(
fpn_features, rpn_rois, output_size=7)
class_outputs, box_outputs = self._frcnn_head_fn(
roi_features, is_training)
model_outputs.update({
'class_outputs':
tf.nest.map_structure(lambda x: tf.cast(x, tf.float32),
class_outputs),
'box_outputs':
tf.nest.map_structure(lambda x: tf.cast(x, tf.float32),
box_outputs),
})
# Add this output to train to make the checkpoint loadable in predict mode.
# If we skip it in train mode, the heads will be out-of-order and checkpoint
# loading will fail.
boxes, scores, classes, valid_detections = self._generate_detections_fn(
box_outputs, class_outputs, rpn_rois, inputs['image_info'][:,
1:2, :])
model_outputs.update({
'num_detections': valid_detections,
'detection_boxes': boxes,
'detection_classes': classes,
'detection_scores': scores,
})
if not self._include_mask:
return model_outputs
if is_training:
rpn_rois, classes, mask_targets = self._sample_masks_fn(
rpn_rois, matched_gt_boxes, matched_gt_classes,
matched_gt_indices, inputs['gt_masks'])
mask_targets = tf.stop_gradient(mask_targets)
classes = tf.cast(classes, tf.int32)
model_outputs.update({
'mask_targets': mask_targets,
'sampled_class_targets': classes,
})
else:
rpn_rois = boxes
classes = tf.cast(classes, tf.int32)
mask_roi_features = spatial_transform_ops.multilevel_crop_and_resize(
fpn_features, rpn_rois, output_size=14)
mask_outputs = self._mrcnn_head_fn(mask_roi_features, classes,
is_training)
if is_training:
model_outputs.update({
'mask_outputs':
tf.nest.map_structure(lambda x: tf.cast(x, tf.float32),
mask_outputs),
})
else:
model_outputs.update(
{'detection_masks': tf.nn.sigmoid(mask_outputs)})
return model_outputs