def pad_groundtruths_to_fixed_size(gt, n): """Pads the first dimension of groundtruths labels to the fixed size.""" gt['boxes'] = input_utils.clip_or_pad_to_fixed_size(gt['boxes'], n, -1) gt['is_crowds'] = input_utils.clip_or_pad_to_fixed_size( gt['is_crowds'], n, 0) gt['areas'] = input_utils.clip_or_pad_to_fixed_size(gt['areas'], n, -1) gt['classes'] = input_utils.clip_or_pad_to_fixed_size(gt['classes'], n, -1) return gt
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'] if self._include_mask: 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) if self._include_mask: masks = tf.gather(masks, indices) # 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) else: image, boxes = input_utils.random_horizontal_flip(image, boxes) # 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) classes = tf.gather(classes, indices) if self._include_mask: masks = tf.gather(masks, indices) # 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), dtype=tf.float32)) # 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 = { 'anchor_boxes': input_anchor.multilevel_boxes, 'image_info': image_info, 'rpn_score_targets': rpn_score_targets, 'rpn_box_targets': rpn_box_targets, } labels['gt_boxes'] = input_utils.clip_or_pad_to_fixed_size( boxes, self._max_num_instances, -1) labels['gt_classes'] = input_utils.clip_or_pad_to_fixed_size( classes, self._max_num_instances, -1) if self._include_mask: labels['gt_masks'] = input_utils.clip_or_pad_to_fixed_size( masks, self._max_num_instances, -1) 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) # If not using category, makes all categories with id = 1. if not self._use_category: classes = tf.cast(tf.greater(classes, 0), dtype=tf.int32) 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] # Randomly shuffle groundtruth masks for mask branch training. rand_indices = tf.random.shuffle(tf.range(num_masks)) shuffled_boxes = tf.gather(boxes, rand_indices) shuffled_classes = tf.gather(classes, rand_indices) shuffled_masks = tf.gather(masks, rand_indices) # Pad sampled boxes/masks/classes to a constant batch size. If the image # has more masks than `num_sampled_masks`, the tensor will be clipped. padded_boxes = input_utils.clip_or_pad_to_fixed_size( shuffled_boxes, self._num_sampled_masks) padded_classes = input_utils.clip_or_pad_to_fixed_size( shuffled_classes, self._num_sampled_masks) padded_masks = input_utils.clip_or_pad_to_fixed_size( shuffled_masks, self._num_sampled_masks) # Jitter the sampled boxes to mimic the noisy detections. padded_boxes = box_utils.jitter_boxes( padded_boxes, noise_scale=self._box_jitter_scale) padded_boxes = box_utils.clip_boxes(padded_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( padded_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]. padded_masks = tf.cast(tf.expand_dims(padded_masks, axis=-1), tf.float32) mask_targets = tf.image.crop_and_resize( padded_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( padded_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(padded_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), padded_classes.dtype) match = tf.reduce_any( tf.equal(tf.expand_dims(padded_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': padded_boxes, 'mask_outer_boxes': mask_outer_boxes, 'mask_targets': mask_targets, 'fine_mask_targets': fine_mask_targets, 'mask_classes': padded_classes, 'mask_is_valid': mask_is_valid, } return image, labels