def testAnchorGeneration(self, min_level, max_level, num_scales,
aspect_ratios, anchor_size, expected_boxes):
image_size = [64, 64]
anchors = anchor.Anchor(min_level, max_level, num_scales,
aspect_ratios, anchor_size, image_size)
boxes = anchors.boxes.numpy()
self.assertEqual(expected_boxes, boxes.tolist())
def testAnchorGenerationWithImageSizeAsTensor(self, min_level, max_level,
num_scales, aspect_ratios,
anchor_size, expected_boxes):
image_size = tf.constant([64, 64], tf.int32)
anchors = anchor.Anchor(min_level, max_level, num_scales,
aspect_ratios, anchor_size, image_size)
boxes = anchors.boxes.numpy()
self.assertEqual(expected_boxes, boxes.tolist())
def construct_model_and_anchors(image_size, use_gt_boxes_for_masks):
num_classes = 3
min_level = 3
max_level = 4
num_scales = 3
aspect_ratios = [1.0]
anchor_boxes = anchor.Anchor(
min_level=min_level,
max_level=max_level,
num_scales=num_scales,
aspect_ratios=aspect_ratios,
anchor_size=3,
image_size=image_size).multilevel_boxes
num_anchors_per_location = len(aspect_ratios) * num_scales
input_specs = tf.keras.layers.InputSpec(shape=[None, None, None, 3])
backbone = resnet.ResNet(model_id=50, input_specs=input_specs)
decoder = fpn.FPN(
min_level=min_level,
max_level=max_level,
input_specs=backbone.output_specs)
rpn_head = dense_prediction_heads.RPNHead(
min_level=min_level,
max_level=max_level,
num_anchors_per_location=num_anchors_per_location)
detection_head = instance_heads.DetectionHead(
num_classes=num_classes)
roi_generator_obj = roi_generator.MultilevelROIGenerator()
roi_sampler_obj = roi_sampler.ROISampler()
roi_aligner_obj = roi_aligner.MultilevelROIAligner()
detection_generator_obj = detection_generator.DetectionGenerator()
mask_head = deep_instance_heads.DeepMaskHead(
num_classes=num_classes, upsample_factor=2)
mask_sampler_obj = mask_sampler.MaskSampler(
mask_target_size=28, num_sampled_masks=1)
mask_roi_aligner_obj = roi_aligner.MultilevelROIAligner(crop_size=14)
model = maskrcnn_model.DeepMaskRCNNModel(
backbone,
decoder,
rpn_head,
detection_head,
roi_generator_obj,
roi_sampler_obj,
roi_aligner_obj,
detection_generator_obj,
mask_head,
mask_sampler_obj,
mask_roi_aligner_obj,
use_gt_boxes_for_masks=use_gt_boxes_for_masks)
return model, anchor_boxes
def testEquivalentResult(self, min_level, max_level, aspect_ratios,
num_scales, anchor_size, image_size):
anchor_gen = anchor.build_anchor_generator(min_level=min_level,
max_level=max_level,
num_scales=num_scales,
aspect_ratios=aspect_ratios,
anchor_size=anchor_size)
anchors = anchor_gen(image_size)
expected_anchor_gen = anchor.Anchor(min_level, max_level, num_scales,
aspect_ratios, anchor_size,
image_size)
expected_anchors = expected_anchor_gen.multilevel_boxes
for k in expected_anchors.keys():
self.assertAllClose(expected_anchors[k], anchors[k])
def test_build_model(self, use_separable_conv, build_anchor_boxes,
is_training, has_att_heads):
num_classes = 3
min_level = 3
max_level = 7
num_scales = 3
aspect_ratios = [1.0]
anchor_size = 3
fpn_num_filters = 256
head_num_convs = 4
head_num_filters = 256
num_anchors_per_location = num_scales * len(aspect_ratios)
image_size = 384
images = np.random.rand(2, image_size, image_size, 3)
image_shape = np.array([[image_size, image_size],
[image_size, image_size]])
if build_anchor_boxes:
anchor_boxes = anchor.Anchor(
min_level=min_level,
max_level=max_level,
num_scales=num_scales,
aspect_ratios=aspect_ratios,
anchor_size=anchor_size,
image_size=(image_size, image_size)).multilevel_boxes
for l in anchor_boxes:
anchor_boxes[l] = tf.tile(
tf.expand_dims(anchor_boxes[l], axis=0), [2, 1, 1, 1])
else:
anchor_boxes = None
if has_att_heads:
attribute_heads = [dict(name='depth', type='regression', size=1)]
else:
attribute_heads = None
backbone = resnet.ResNet(model_id=50)
decoder = fpn.FPN(input_specs=backbone.output_specs,
min_level=min_level,
max_level=max_level,
num_filters=fpn_num_filters,
use_separable_conv=use_separable_conv)
head = dense_prediction_heads.RetinaNetHead(
min_level=min_level,
max_level=max_level,
num_classes=num_classes,
attribute_heads=attribute_heads,
num_anchors_per_location=num_anchors_per_location,
use_separable_conv=use_separable_conv,
num_convs=head_num_convs,
num_filters=head_num_filters)
generator = detection_generator.MultilevelDetectionGenerator(
max_num_detections=10)
model = retinanet_model.RetinaNetModel(backbone=backbone,
decoder=decoder,
head=head,
detection_generator=generator,
min_level=min_level,
max_level=max_level,
num_scales=num_scales,
aspect_ratios=aspect_ratios,
anchor_size=anchor_size)
_ = model(images, image_shape, anchor_boxes, training=is_training)
def test_build_model(self,
use_separable_conv,
build_anchor_boxes,
shared_backbone,
shared_decoder,
is_training=True):
num_classes = 3
min_level = 2
max_level = 6
num_scales = 3
aspect_ratios = [1.0]
anchor_size = 3
resnet_model_id = 50
segmentation_resnet_model_id = 50
aspp_dilation_rates = [6, 12, 18]
aspp_decoder_level = 2
fpn_decoder_level = 2
num_anchors_per_location = num_scales * len(aspect_ratios)
image_size = 128
images = tf.random.normal([2, image_size, image_size, 3])
image_info = tf.convert_to_tensor([[[image_size, image_size],
[image_size, image_size], [1, 1],
[0, 0]],
[[image_size, image_size],
[image_size, image_size], [1, 1],
[0, 0]]])
shared_decoder = shared_decoder and shared_backbone
if build_anchor_boxes or not is_training:
anchor_boxes = anchor.Anchor(
min_level=min_level,
max_level=max_level,
num_scales=num_scales,
aspect_ratios=aspect_ratios,
anchor_size=3,
image_size=(image_size, image_size)).multilevel_boxes
for l in anchor_boxes:
anchor_boxes[l] = tf.tile(
tf.expand_dims(anchor_boxes[l], axis=0), [2, 1, 1, 1])
else:
anchor_boxes = None
backbone = resnet.ResNet(model_id=resnet_model_id)
decoder = fpn.FPN(input_specs=backbone.output_specs,
min_level=min_level,
max_level=max_level,
use_separable_conv=use_separable_conv)
rpn_head = dense_prediction_heads.RPNHead(
min_level=min_level,
max_level=max_level,
num_anchors_per_location=num_anchors_per_location,
num_convs=1)
detection_head = instance_heads.DetectionHead(num_classes=num_classes)
roi_generator_obj = roi_generator.MultilevelROIGenerator()
roi_sampler_obj = roi_sampler.ROISampler()
roi_aligner_obj = roi_aligner.MultilevelROIAligner()
detection_generator_obj = detection_generator.DetectionGenerator()
panoptic_segmentation_generator_obj = panoptic_segmentation_generator.PanopticSegmentationGenerator(
output_size=[image_size, image_size],
max_num_detections=100,
stuff_classes_offset=90)
mask_head = instance_heads.MaskHead(num_classes=num_classes,
upsample_factor=2)
mask_sampler_obj = mask_sampler.MaskSampler(mask_target_size=28,
num_sampled_masks=1)
mask_roi_aligner_obj = roi_aligner.MultilevelROIAligner(crop_size=14)
if shared_backbone:
segmentation_backbone = None
else:
segmentation_backbone = resnet.ResNet(
model_id=segmentation_resnet_model_id)
if not shared_decoder:
feature_fusion = 'deeplabv3plus'
level = aspp_decoder_level
segmentation_decoder = aspp.ASPP(
level=level, dilation_rates=aspp_dilation_rates)
else:
feature_fusion = 'panoptic_fpn_fusion'
level = fpn_decoder_level
segmentation_decoder = None
segmentation_head = segmentation_heads.SegmentationHead(
num_classes=2, # stuff and common class for things,
level=level,
feature_fusion=feature_fusion,
decoder_min_level=min_level,
decoder_max_level=max_level,
num_convs=2)
model = panoptic_maskrcnn_model.PanopticMaskRCNNModel(
backbone,
decoder,
rpn_head,
detection_head,
roi_generator_obj,
roi_sampler_obj,
roi_aligner_obj,
detection_generator_obj,
panoptic_segmentation_generator_obj,
mask_head,
mask_sampler_obj,
mask_roi_aligner_obj,
segmentation_backbone=segmentation_backbone,
segmentation_decoder=segmentation_decoder,
segmentation_head=segmentation_head,
min_level=min_level,
max_level=max_level,
num_scales=num_scales,
aspect_ratios=aspect_ratios,
anchor_size=anchor_size)
gt_boxes = tf.convert_to_tensor(
[[[10, 10, 15, 15], [2.5, 2.5, 7.5, 7.5], [-1, -1, -1, -1]],
[[100, 100, 150, 150], [-1, -1, -1, -1], [-1, -1, -1, -1]]],
dtype=tf.float32)
gt_classes = tf.convert_to_tensor([[2, 1, -1], [1, -1, -1]],
dtype=tf.int32)
gt_masks = tf.ones((2, 3, 100, 100))
# Results will be checked in test_forward.
_ = model(images,
image_info,
anchor_boxes,
gt_boxes,
gt_classes,
gt_masks,
training=is_training)
def test_forward(self, strategy, training, shared_backbone, shared_decoder,
generate_panoptic_masks):
num_classes = 3
min_level = 2
max_level = 6
num_scales = 3
aspect_ratios = [1.0]
anchor_size = 3
segmentation_resnet_model_id = 101
aspp_dilation_rates = [6, 12, 18]
aspp_decoder_level = 2
fpn_decoder_level = 2
class_agnostic_bbox_pred = False
cascade_class_ensemble = False
image_size = (256, 256)
images = tf.random.normal([2, image_size[0], image_size[1], 3])
image_info = tf.convert_to_tensor([[[224, 100], [224, 100], [1, 1],
[0, 0]],
[[224, 100], [224, 100], [1, 1],
[0, 0]]])
shared_decoder = shared_decoder and shared_backbone
with strategy.scope():
anchor_boxes = anchor.Anchor(
min_level=min_level,
max_level=max_level,
num_scales=num_scales,
aspect_ratios=aspect_ratios,
anchor_size=anchor_size,
image_size=image_size).multilevel_boxes
num_anchors_per_location = len(aspect_ratios) * num_scales
input_specs = tf.keras.layers.InputSpec(
shape=[None, None, None, 3])
backbone = resnet.ResNet(model_id=50, input_specs=input_specs)
decoder = fpn.FPN(min_level=min_level,
max_level=max_level,
input_specs=backbone.output_specs)
rpn_head = dense_prediction_heads.RPNHead(
min_level=min_level,
max_level=max_level,
num_anchors_per_location=num_anchors_per_location)
detection_head = instance_heads.DetectionHead(
num_classes=num_classes,
class_agnostic_bbox_pred=class_agnostic_bbox_pred)
roi_generator_obj = roi_generator.MultilevelROIGenerator()
roi_sampler_cascade = []
roi_sampler_obj = roi_sampler.ROISampler()
roi_sampler_cascade.append(roi_sampler_obj)
roi_aligner_obj = roi_aligner.MultilevelROIAligner()
detection_generator_obj = detection_generator.DetectionGenerator()
if generate_panoptic_masks:
panoptic_segmentation_generator_obj = panoptic_segmentation_generator.PanopticSegmentationGenerator(
output_size=list(image_size),
max_num_detections=100,
stuff_classes_offset=90)
else:
panoptic_segmentation_generator_obj = None
mask_head = instance_heads.MaskHead(num_classes=num_classes,
upsample_factor=2)
mask_sampler_obj = mask_sampler.MaskSampler(mask_target_size=28,
num_sampled_masks=1)
mask_roi_aligner_obj = roi_aligner.MultilevelROIAligner(
crop_size=14)
if shared_backbone:
segmentation_backbone = None
else:
segmentation_backbone = resnet.ResNet(
model_id=segmentation_resnet_model_id)
if not shared_decoder:
feature_fusion = 'deeplabv3plus'
level = aspp_decoder_level
segmentation_decoder = aspp.ASPP(
level=level, dilation_rates=aspp_dilation_rates)
else:
feature_fusion = 'panoptic_fpn_fusion'
level = fpn_decoder_level
segmentation_decoder = None
segmentation_head = segmentation_heads.SegmentationHead(
num_classes=2, # stuff and common class for things,
level=level,
feature_fusion=feature_fusion,
decoder_min_level=min_level,
decoder_max_level=max_level,
num_convs=2)
model = panoptic_maskrcnn_model.PanopticMaskRCNNModel(
backbone,
decoder,
rpn_head,
detection_head,
roi_generator_obj,
roi_sampler_obj,
roi_aligner_obj,
detection_generator_obj,
panoptic_segmentation_generator_obj,
mask_head,
mask_sampler_obj,
mask_roi_aligner_obj,
segmentation_backbone=segmentation_backbone,
segmentation_decoder=segmentation_decoder,
segmentation_head=segmentation_head,
class_agnostic_bbox_pred=class_agnostic_bbox_pred,
cascade_class_ensemble=cascade_class_ensemble,
min_level=min_level,
max_level=max_level,
num_scales=num_scales,
aspect_ratios=aspect_ratios,
anchor_size=anchor_size)
gt_boxes = tf.convert_to_tensor(
[[[10, 10, 15, 15], [2.5, 2.5, 7.5, 7.5], [-1, -1, -1, -1]],
[[100, 100, 150, 150], [-1, -1, -1, -1], [-1, -1, -1, -1]]],
dtype=tf.float32)
gt_classes = tf.convert_to_tensor([[2, 1, -1], [1, -1, -1]],
dtype=tf.int32)
gt_masks = tf.ones((2, 3, 100, 100))
results = model(images,
image_info,
anchor_boxes,
gt_boxes,
gt_classes,
gt_masks,
training=training)
self.assertIn('rpn_boxes', results)
self.assertIn('rpn_scores', results)
if training:
self.assertIn('class_targets', results)
self.assertIn('box_targets', results)
self.assertIn('class_outputs', results)
self.assertIn('box_outputs', results)
self.assertIn('mask_outputs', results)
else:
self.assertIn('detection_boxes', results)
self.assertIn('detection_scores', results)
self.assertIn('detection_classes', results)
self.assertIn('num_detections', results)
self.assertIn('detection_masks', results)
self.assertIn('segmentation_outputs', results)
self.assertAllEqual([
2, image_size[0] // (2**level), image_size[1] // (2**level), 2
], results['segmentation_outputs'].numpy().shape)
if generate_panoptic_masks:
self.assertIn('panoptic_outputs', results)
self.assertIn('category_mask', results['panoptic_outputs'])
self.assertIn('instance_mask', results['panoptic_outputs'])
self.assertAllEqual(
[2, image_size[0], image_size[1]],
results['panoptic_outputs']['category_mask'].numpy().shape)
self.assertAllEqual(
[2, image_size[0], image_size[1]],
results['panoptic_outputs']['instance_mask'].numpy().shape)
else:
self.assertNotIn('panoptic_outputs', results)
def test_build_model(self, include_mask, use_separable_conv,
build_anchor_boxes, is_training):
num_classes = 3
min_level = 3
max_level = 7
num_scales = 3
aspect_ratios = [1.0]
anchor_size = 3
resnet_model_id = 50
num_anchors_per_location = num_scales * len(aspect_ratios)
image_size = 384
images = np.random.rand(2, image_size, image_size, 3)
image_shape = np.array([[image_size, image_size],
[image_size, image_size]])
if build_anchor_boxes:
anchor_boxes = anchor.Anchor(
min_level=min_level,
max_level=max_level,
num_scales=num_scales,
aspect_ratios=aspect_ratios,
anchor_size=3,
image_size=(image_size, image_size)).multilevel_boxes
for l in anchor_boxes:
anchor_boxes[l] = tf.tile(
tf.expand_dims(anchor_boxes[l], axis=0), [2, 1, 1, 1])
else:
anchor_boxes = None
backbone = resnet.ResNet(model_id=resnet_model_id)
decoder = fpn.FPN(input_specs=backbone.output_specs,
min_level=min_level,
max_level=max_level,
use_separable_conv=use_separable_conv)
rpn_head = dense_prediction_heads.RPNHead(
min_level=min_level,
max_level=max_level,
num_anchors_per_location=num_anchors_per_location,
num_convs=1)
detection_head = instance_heads.DetectionHead(num_classes=num_classes)
roi_generator_obj = roi_generator.MultilevelROIGenerator()
roi_sampler_obj = roi_sampler.ROISampler()
roi_aligner_obj = roi_aligner.MultilevelROIAligner()
detection_generator_obj = detection_generator.DetectionGenerator()
if include_mask:
mask_head = instance_heads.MaskHead(num_classes=num_classes,
upsample_factor=2)
mask_sampler_obj = mask_sampler.MaskSampler(mask_target_size=28,
num_sampled_masks=1)
mask_roi_aligner_obj = roi_aligner.MultilevelROIAligner(
crop_size=14)
else:
mask_head = None
mask_sampler_obj = None
mask_roi_aligner_obj = None
model = maskrcnn_model.MaskRCNNModel(backbone,
decoder,
rpn_head,
detection_head,
roi_generator_obj,
roi_sampler_obj,
roi_aligner_obj,
detection_generator_obj,
mask_head,
mask_sampler_obj,
mask_roi_aligner_obj,
min_level=min_level,
max_level=max_level,
num_scales=num_scales,
aspect_ratios=aspect_ratios,
anchor_size=anchor_size)
gt_boxes = np.array(
[[[10, 10, 15, 15], [2.5, 2.5, 7.5, 7.5], [-1, -1, -1, -1]],
[[100, 100, 150, 150], [-1, -1, -1, -1], [-1, -1, -1, -1]]],
dtype=np.float32)
gt_classes = np.array([[2, 1, -1], [1, -1, -1]], dtype=np.int32)
if include_mask:
gt_masks = np.ones((2, 3, 100, 100))
else:
gt_masks = None
# Results will be checked in test_forward.
_ = model(images,
image_shape,
anchor_boxes,
gt_boxes,
gt_classes,
gt_masks,
training=is_training)
def test_forward(self, strategy, include_mask, build_anchor_boxes,
training, use_cascade_heads):
num_classes = 3
min_level = 3
max_level = 4
num_scales = 3
aspect_ratios = [1.0]
anchor_size = 3
if use_cascade_heads:
cascade_iou_thresholds = [0.6]
class_agnostic_bbox_pred = True
cascade_class_ensemble = True
else:
cascade_iou_thresholds = None
class_agnostic_bbox_pred = False
cascade_class_ensemble = False
image_size = (256, 256)
images = np.random.rand(2, image_size[0], image_size[1], 3)
image_shape = np.array([[224, 100], [100, 224]])
with strategy.scope():
if build_anchor_boxes:
anchor_boxes = anchor.Anchor(
min_level=min_level,
max_level=max_level,
num_scales=num_scales,
aspect_ratios=aspect_ratios,
anchor_size=anchor_size,
image_size=image_size).multilevel_boxes
else:
anchor_boxes = None
num_anchors_per_location = len(aspect_ratios) * num_scales
input_specs = tf.keras.layers.InputSpec(
shape=[None, None, None, 3])
backbone = resnet.ResNet(model_id=50, input_specs=input_specs)
decoder = fpn.FPN(min_level=min_level,
max_level=max_level,
input_specs=backbone.output_specs)
rpn_head = dense_prediction_heads.RPNHead(
min_level=min_level,
max_level=max_level,
num_anchors_per_location=num_anchors_per_location)
detection_head = instance_heads.DetectionHead(
num_classes=num_classes,
class_agnostic_bbox_pred=class_agnostic_bbox_pred)
roi_generator_obj = roi_generator.MultilevelROIGenerator()
roi_sampler_cascade = []
roi_sampler_obj = roi_sampler.ROISampler()
roi_sampler_cascade.append(roi_sampler_obj)
if cascade_iou_thresholds:
for iou in cascade_iou_thresholds:
roi_sampler_obj = roi_sampler.ROISampler(
mix_gt_boxes=False,
foreground_iou_threshold=iou,
background_iou_high_threshold=iou,
background_iou_low_threshold=0.0,
skip_subsampling=True)
roi_sampler_cascade.append(roi_sampler_obj)
roi_aligner_obj = roi_aligner.MultilevelROIAligner()
detection_generator_obj = detection_generator.DetectionGenerator()
if include_mask:
mask_head = instance_heads.MaskHead(num_classes=num_classes,
upsample_factor=2)
mask_sampler_obj = mask_sampler.MaskSampler(
mask_target_size=28, num_sampled_masks=1)
mask_roi_aligner_obj = roi_aligner.MultilevelROIAligner(
crop_size=14)
else:
mask_head = None
mask_sampler_obj = None
mask_roi_aligner_obj = None
model = maskrcnn_model.MaskRCNNModel(
backbone,
decoder,
rpn_head,
detection_head,
roi_generator_obj,
roi_sampler_obj,
roi_aligner_obj,
detection_generator_obj,
mask_head,
mask_sampler_obj,
mask_roi_aligner_obj,
class_agnostic_bbox_pred=class_agnostic_bbox_pred,
cascade_class_ensemble=cascade_class_ensemble,
min_level=min_level,
max_level=max_level,
num_scales=num_scales,
aspect_ratios=aspect_ratios,
anchor_size=anchor_size)
gt_boxes = np.array(
[[[10, 10, 15, 15], [2.5, 2.5, 7.5, 7.5], [-1, -1, -1, -1]],
[[100, 100, 150, 150], [-1, -1, -1, -1], [-1, -1, -1, -1]]],
dtype=np.float32)
gt_classes = np.array([[2, 1, -1], [1, -1, -1]], dtype=np.int32)
if include_mask:
gt_masks = np.ones((2, 3, 100, 100))
else:
gt_masks = None
results = model(images,
image_shape,
anchor_boxes,
gt_boxes,
gt_classes,
gt_masks,
training=training)
self.assertIn('rpn_boxes', results)
self.assertIn('rpn_scores', results)
if training:
self.assertIn('class_targets', results)
self.assertIn('box_targets', results)
self.assertIn('class_outputs', results)
self.assertIn('box_outputs', results)
if include_mask:
self.assertIn('mask_outputs', results)
else:
self.assertIn('detection_boxes', results)
self.assertIn('detection_scores', results)
self.assertIn('detection_classes', results)
self.assertIn('num_detections', results)
if include_mask:
self.assertIn('detection_masks', results)
def _call_box_outputs(
self, images: tf.Tensor,
image_shape: tf.Tensor,
anchor_boxes: Optional[Mapping[str, tf.Tensor]] = None,
gt_boxes: Optional[tf.Tensor] = None,
gt_classes: Optional[tf.Tensor] = None,
training: Optional[bool] = None) -> Tuple[
Mapping[str, tf.Tensor], Mapping[str, tf.Tensor]]:
"""Implementation of the Faster-RCNN logic for boxes."""
model_outputs = {}
# Feature extraction.
(backbone_features,
decoder_features) = self._get_backbone_and_decoder_features(images)
# Region proposal network.
rpn_scores, rpn_boxes = self.rpn_head(decoder_features)
model_outputs.update({
'backbone_features': backbone_features,
'decoder_features': decoder_features,
'rpn_boxes': rpn_boxes,
'rpn_scores': rpn_scores
})
# Generate anchor boxes for this batch if not provided.
if anchor_boxes is None:
_, image_height, image_width, _ = images.get_shape().as_list()
anchor_boxes = anchor.Anchor(
min_level=self._config_dict['min_level'],
max_level=self._config_dict['max_level'],
num_scales=self._config_dict['num_scales'],
aspect_ratios=self._config_dict['aspect_ratios'],
anchor_size=self._config_dict['anchor_size'],
image_size=(image_height, image_width)).multilevel_boxes
for l in anchor_boxes:
anchor_boxes[l] = tf.tile(
tf.expand_dims(anchor_boxes[l], axis=0),
[tf.shape(images)[0], 1, 1, 1])
# Generate RoIs.
current_rois, _ = self.roi_generator(rpn_boxes, rpn_scores, anchor_boxes,
image_shape, training)
next_rois = current_rois
all_class_outputs = []
for cascade_num in range(len(self.roi_sampler)):
# In cascade RCNN we want the higher layers to have different regression
# weights as the predicted deltas become smaller and smaller.
regression_weights = self._cascade_layer_to_weights[cascade_num]
current_rois = next_rois
(class_outputs, box_outputs, model_outputs, matched_gt_boxes,
matched_gt_classes, matched_gt_indices,
current_rois) = self._run_frcnn_head(
features=decoder_features,
rois=current_rois,
gt_boxes=gt_boxes,
gt_classes=gt_classes,
training=training,
model_outputs=model_outputs,
cascade_num=cascade_num,
regression_weights=regression_weights)
all_class_outputs.append(class_outputs)
# Generate ROIs for the next cascade head if there is any.
if cascade_num < len(self.roi_sampler) - 1:
next_rois = box_ops.decode_boxes(
tf.cast(box_outputs, tf.float32),
current_rois,
weights=regression_weights)
next_rois = box_ops.clip_boxes(next_rois,
tf.expand_dims(image_shape, axis=1))
if not training:
if self._config_dict['cascade_class_ensemble']:
class_outputs = tf.add_n(all_class_outputs) / len(all_class_outputs)
detections = self.detection_generator(
box_outputs,
class_outputs,
current_rois,
image_shape,
regression_weights,
bbox_per_class=(not self._config_dict['class_agnostic_bbox_pred']))
model_outputs.update({
'cls_outputs': class_outputs,
'box_outputs': box_outputs,
})
if self.detection_generator.get_config()['apply_nms']:
model_outputs.update({
'detection_boxes': detections['detection_boxes'],
'detection_scores': detections['detection_scores'],
'detection_classes': detections['detection_classes'],
'num_detections': detections['num_detections']
})
else:
model_outputs.update({
'decoded_boxes': detections['decoded_boxes'],
'decoded_box_scores': detections['decoded_box_scores']
})
intermediate_outputs = {
'matched_gt_boxes': matched_gt_boxes,
'matched_gt_indices': matched_gt_indices,
'matched_gt_classes': matched_gt_classes,
'current_rois': current_rois,
}
return (model_outputs, intermediate_outputs)
def call(self,
images: Union[tf.Tensor, Sequence[tf.Tensor]],
image_shape: Optional[tf.Tensor] = None,
anchor_boxes: Optional[Mapping[str, tf.Tensor]] = None,
output_intermediate_features: bool = False,
training: bool = None) -> Mapping[str, tf.Tensor]:
"""Forward pass of the RetinaNet model.
Args:
images: `Tensor` or a sequence of `Tensor`, the input batched images to
the backbone network, whose shape(s) is [batch, height, width, 3]. If it
is a sequence of `Tensor`, we will assume the anchors are generated
based on the shape of the first image(s).
image_shape: `Tensor`, the actual shape of the input images, whose shape
is [batch, 2] where the last dimension is [height, width]. Note that
this is the actual image shape excluding paddings. For example, images
in the batch may be resized into different shapes before padding to the
fixed size.
anchor_boxes: a dict of tensors which includes multilevel anchors.
- key: `str`, the level of the multilevel predictions.
- values: `Tensor`, the anchor coordinates of a particular feature
level, whose shape is [height_l, width_l, num_anchors_per_location].
output_intermediate_features: `bool` indicating whether to return the
intermediate feature maps generated by backbone and decoder.
training: `bool`, indicating whether it is in training mode.
Returns:
scores: a dict of tensors which includes scores of the predictions.
- key: `str`, the level of the multilevel predictions.
- values: `Tensor`, the box scores predicted from a particular feature
level, whose shape is
[batch, height_l, width_l, num_classes * num_anchors_per_location].
boxes: a dict of tensors which includes coordinates of the predictions.
- key: `str`, the level of the multilevel predictions.
- values: `Tensor`, the box coordinates predicted from a particular
feature level, whose shape is
[batch, height_l, width_l, 4 * num_anchors_per_location].
attributes: a dict of (attribute_name, attribute_predictions). Each
attribute prediction is a dict that includes:
- key: `str`, the level of the multilevel predictions.
- values: `Tensor`, the attribute predictions from a particular
feature level, whose shape is
[batch, height_l, width_l, att_size * num_anchors_per_location].
"""
outputs = {}
# Feature extraction.
features = self.backbone(images)
if output_intermediate_features:
outputs.update(
{'backbone_{}'.format(k): v
for k, v in features.items()})
if self.decoder:
features = self.decoder(features)
if output_intermediate_features:
outputs.update(
{'decoder_{}'.format(k): v
for k, v in features.items()})
# Dense prediction. `raw_attributes` can be empty.
raw_scores, raw_boxes, raw_attributes = self.head(features)
if training:
outputs.update({
'cls_outputs': raw_scores,
'box_outputs': raw_boxes,
})
if raw_attributes:
outputs.update({'attribute_outputs': raw_attributes})
return outputs
else:
# Generate anchor boxes for this batch if not provided.
if anchor_boxes is None:
if isinstance(images, Sequence):
primary_images = images[0]
elif isinstance(images, tf.Tensor):
primary_images = images
else:
raise ValueError(
'Input should be a tf.Tensor or a sequence of tf.Tensor, not {}.'
.format(type(images)))
_, image_height, image_width, _ = primary_images.get_shape(
).as_list()
anchor_boxes = anchor.Anchor(
min_level=self._config_dict['min_level'],
max_level=self._config_dict['max_level'],
num_scales=self._config_dict['num_scales'],
aspect_ratios=self._config_dict['aspect_ratios'],
anchor_size=self._config_dict['anchor_size'],
image_size=(image_height, image_width)).multilevel_boxes
for l in anchor_boxes:
anchor_boxes[l] = tf.tile(
tf.expand_dims(anchor_boxes[l], axis=0),
[tf.shape(primary_images)[0], 1, 1, 1])
# Post-processing.
final_results = self.detection_generator(raw_boxes, raw_scores,
anchor_boxes, image_shape,
raw_attributes)
outputs.update({
'cls_outputs': raw_scores,
'box_outputs': raw_boxes,
})
if self.detection_generator.get_config()['apply_nms']:
outputs.update({
'detection_boxes':
final_results['detection_boxes'],
'detection_scores':
final_results['detection_scores'],
'detection_classes':
final_results['detection_classes'],
'num_detections':
final_results['num_detections']
})
else:
outputs.update({
'decoded_boxes':
final_results['decoded_boxes'],
'decoded_box_scores':
final_results['decoded_box_scores']
})
if raw_attributes:
outputs.update({
'attribute_outputs':
raw_attributes,
'detection_attributes':
final_results['detection_attributes'],
})
return outputs
