def _get_target_assigner(self):
similarity_calc = region_similarity_calculator.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
unmatched_threshold=0.5)
box_coder = mean_stddev_box_coder.MeanStddevBoxCoder(stddev=0.1)
return targetassigner.TargetAssigner(similarity_calc, matcher,
box_coder)
def graph_fn(anchor_means, groundtruth_box_corners):
similarity_calc = region_similarity_calculator.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
unmatched_threshold=0.3)
box_coder = mean_stddev_box_coder.MeanStddevBoxCoder(stddev=0.1)
target_assigner = targetassigner.TargetAssigner(
similarity_calc, matcher, box_coder)
anchors_boxlist = box_list.BoxList(anchor_means)
groundtruth_boxlist = box_list.BoxList(groundtruth_box_corners)
result = target_assigner.assign(anchors_boxlist,
groundtruth_boxlist,
unmatched_class_label=None)
(cls_targets, cls_weights, reg_targets, reg_weights, _) = result
return (cls_targets, cls_weights, reg_targets, reg_weights)
def graph_fn(anchor_means, groundtruth_box_corners,
groundtruth_keypoints):
similarity_calc = region_similarity_calculator.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
unmatched_threshold=0.5)
box_coder = keypoint_box_coder.KeypointBoxCoder(
num_keypoints=6, scale_factors=[10.0, 10.0, 5.0, 5.0])
target_assigner = targetassigner.TargetAssigner(
similarity_calc, matcher, box_coder)
anchors_boxlist = box_list.BoxList(anchor_means)
groundtruth_boxlist = box_list.BoxList(groundtruth_box_corners)
groundtruth_boxlist.add_field(fields.BoxListFields.keypoints,
groundtruth_keypoints)
result = target_assigner.assign(anchors_boxlist,
groundtruth_boxlist,
unmatched_class_label=None)
(cls_targets, cls_weights, reg_targets, reg_weights, _) = result
return (cls_targets, cls_weights, reg_targets, reg_weights)
def graph_fn(anchor_means, groundtruth_box_corners, groundtruth_labels,
groundtruth_weights):
similarity_calc = region_similarity_calculator.IouSimilarity()
matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
unmatched_threshold=0.5)
box_coder = mean_stddev_box_coder.MeanStddevBoxCoder(stddev=0.1)
unmatched_class_label = tf.constant([1, 0, 0, 0, 0, 0, 0],
tf.float32)
target_assigner = targetassigner.TargetAssigner(
similarity_calc, matcher, box_coder)
anchors_boxlist = box_list.BoxList(anchor_means)
groundtruth_boxlist = box_list.BoxList(groundtruth_box_corners)
result = target_assigner.assign(
anchors_boxlist,
groundtruth_boxlist,
groundtruth_labels,
unmatched_class_label=unmatched_class_label,
groundtruth_weights=groundtruth_weights)
(_, cls_weights, _, reg_weights, _) = result
return (cls_weights, reg_weights)
def test_raises_error_on_invalid_groundtruth_labels(self):
similarity_calc = region_similarity_calculator.NegSqDistSimilarity()
matcher = bipartite_matcher.GreedyBipartiteMatcher()
box_coder = mean_stddev_box_coder.MeanStddevBoxCoder(stddev=1.0)
unmatched_class_label = tf.constant([[0, 0], [0, 0], [0, 0]],
tf.float32)
target_assigner = targetassigner.TargetAssigner(
similarity_calc, matcher, box_coder)
prior_means = tf.constant([[0.0, 0.0, 0.5, 0.5]])
priors = box_list.BoxList(prior_means)
box_corners = [[0.0, 0.0, 0.5, 0.5], [0.5, 0.5, 0.9, 0.9],
[.75, 0, .95, .27]]
boxes = box_list.BoxList(tf.constant(box_corners))
groundtruth_labels = tf.constant([[[0, 1], [1, 0]]], tf.float32)
with self.assertRaises(ValueError):
target_assigner.assign(priors,
boxes,
groundtruth_labels,
unmatched_class_label=unmatched_class_label)
def _create_model(self,
model_fn=ssd_meta_arch.SSDMetaArch,
apply_hard_mining=True,
normalize_loc_loss_by_codesize=False,
add_background_class=True,
random_example_sampling=False,
weight_regression_loss_by_score=False,
use_expected_classification_loss_under_sampling=False,
minimum_negative_sampling=1,
desired_negative_sampling_ratio=3,
use_keras=False,
predict_mask=False,
use_static_shapes=False,
nms_max_size_per_class=5):
is_training = False
num_classes = 1
mock_anchor_generator = MockAnchorGenerator2x2()
if use_keras:
mock_box_predictor = test_utils.MockKerasBoxPredictor(
is_training, num_classes, predict_mask=predict_mask)
else:
mock_box_predictor = test_utils.MockBoxPredictor(
is_training, num_classes, predict_mask=predict_mask)
mock_box_coder = test_utils.MockBoxCoder()
if use_keras:
fake_feature_extractor = FakeSSDKerasFeatureExtractor()
else:
fake_feature_extractor = FakeSSDFeatureExtractor()
mock_matcher = test_utils.MockMatcher()
region_similarity_calculator = sim_calc.IouSimilarity()
encode_background_as_zeros = False
def image_resizer_fn(image):
return [tf.identity(image), tf.shape(image)]
classification_loss = losses.WeightedSigmoidClassificationLoss()
localization_loss = losses.WeightedSmoothL1LocalizationLoss()
non_max_suppression_fn = functools.partial(
post_processing.batch_multiclass_non_max_suppression,
score_thresh=-20.0,
iou_thresh=1.0,
max_size_per_class=nms_max_size_per_class,
max_total_size=nms_max_size_per_class,
use_static_shapes=use_static_shapes)
classification_loss_weight = 1.0
localization_loss_weight = 1.0
negative_class_weight = 1.0
normalize_loss_by_num_matches = False
hard_example_miner = None
if apply_hard_mining:
# This hard example miner is expected to be a no-op.
hard_example_miner = losses.HardExampleMiner(
num_hard_examples=None, iou_threshold=1.0)
random_example_sampler = None
if random_example_sampling:
random_example_sampler = sampler.BalancedPositiveNegativeSampler(
positive_fraction=0.5)
target_assigner_instance = target_assigner.TargetAssigner(
region_similarity_calculator,
mock_matcher,
mock_box_coder,
negative_class_weight=negative_class_weight,
weight_regression_loss_by_score=weight_regression_loss_by_score)
expected_classification_loss_under_sampling = None
if use_expected_classification_loss_under_sampling:
expected_classification_loss_under_sampling = functools.partial(
ops.expected_classification_loss_under_sampling,
minimum_negative_sampling=minimum_negative_sampling,
desired_negative_sampling_ratio=desired_negative_sampling_ratio
)
code_size = 4
model = model_fn(
is_training=is_training,
anchor_generator=mock_anchor_generator,
box_predictor=mock_box_predictor,
box_coder=mock_box_coder,
feature_extractor=fake_feature_extractor,
encode_background_as_zeros=encode_background_as_zeros,
image_resizer_fn=image_resizer_fn,
non_max_suppression_fn=non_max_suppression_fn,
score_conversion_fn=tf.identity,
classification_loss=classification_loss,
localization_loss=localization_loss,
classification_loss_weight=classification_loss_weight,
localization_loss_weight=localization_loss_weight,
normalize_loss_by_num_matches=normalize_loss_by_num_matches,
hard_example_miner=hard_example_miner,
target_assigner_instance=target_assigner_instance,
add_summaries=False,
normalize_loc_loss_by_codesize=normalize_loc_loss_by_codesize,
freeze_batchnorm=False,
inplace_batchnorm_update=False,
add_background_class=add_background_class,
random_example_sampler=random_example_sampler,
expected_classification_loss_under_sampling=
expected_classification_loss_under_sampling)
return model, num_classes, mock_anchor_generator.num_anchors(
), code_size
def __init__(self,
is_training,
anchor_generator,
box_predictor,
box_coder,
feature_extractor,
matcher,
region_similarity_calculator,
encode_background_as_zeros,
negative_class_weight,
image_resizer_fn,
non_max_suppression_fn,
score_conversion_fn,
classification_loss,
localization_loss,
classification_loss_weight,
localization_loss_weight,
normalize_loss_by_num_matches,
hard_example_miner,
add_summaries=True,
normalize_loc_loss_by_codesize=False,
freeze_batchnorm=False,
inplace_batchnorm_update=False,
add_background_class=True,
random_example_sampler=None):
"""SSDMetaArch Constructor.
TODO(rathodv,jonathanhuang): group NMS parameters + score converter into
a class and loss parameters into a class and write config protos for
postprocessing and losses.
Args:
is_training: A boolean indicating whether the training version of the
computation graph should be constructed.
anchor_generator: an anchor_generator.AnchorGenerator object.
box_predictor: a box_predictor.BoxPredictor object.
box_coder: a box_coder.BoxCoder object.
feature_extractor: a SSDFeatureExtractor object.
matcher: a matcher.Matcher object.
region_similarity_calculator: a
region_similarity_calculator.RegionSimilarityCalculator object.
encode_background_as_zeros: boolean determining whether background
targets are to be encoded as an all zeros vector or a one-hot
vector (where background is the 0th class).
negative_class_weight: Weight for confidence loss of negative anchors.
image_resizer_fn: a callable for image resizing. This callable always
takes a rank-3 image tensor (corresponding to a single image) and
returns a rank-3 image tensor, possibly with new spatial dimensions and
a 1-D tensor of shape [3] indicating shape of true image within
the resized image tensor as the resized image tensor could be padded.
See builders/image_resizer_builder.py.
non_max_suppression_fn: batch_multiclass_non_max_suppression
callable that takes `boxes`, `scores` and optional `clip_window`
inputs (with all other inputs already set) and returns a dictionary
hold tensors with keys: `detection_boxes`, `detection_scores`,
`detection_classes` and `num_detections`. See `post_processing.
batch_multiclass_non_max_suppression` for the type and shape of these
tensors.
score_conversion_fn: callable elementwise nonlinearity (that takes tensors
as inputs and returns tensors). This is usually used to convert logits
to probabilities.
classification_loss: an object_detection.core.losses.Loss object.
localization_loss: a object_detection.core.losses.Loss object.
classification_loss_weight: float
localization_loss_weight: float
normalize_loss_by_num_matches: boolean
hard_example_miner: a losses.HardExampleMiner object (can be None)
add_summaries: boolean (default: True) controlling whether summary ops
should be added to tensorflow graph.
normalize_loc_loss_by_codesize: whether to normalize localization loss
by code size of the box encoder.
freeze_batchnorm: Whether to freeze batch norm parameters during
training or not. When training with a small batch size (e.g. 1), it is
desirable to freeze batch norm update and use pretrained batch norm
params.
inplace_batchnorm_update: Whether to update batch norm moving average
values inplace. When this is false train op must add a control
dependency on tf.graphkeys.UPDATE_OPS collection in order to update
batch norm statistics.
add_background_class: Whether to add an implicit background class to
one-hot encodings of groundtruth labels. Set to false if using
groundtruth labels with an explicit background class or using multiclass
scores instead of truth in the case of distillation.
random_example_sampler: a BalancedPositiveNegativeSampler object that can
perform random example sampling when computing loss. If None, random
sampling process is skipped. Note that random example sampler and hard
example miner can both be applied to the model. In that case, random
sampler will take effect first and hard example miner can only process
the random sampled examples.
"""
super(SSDMetaArch,
self).__init__(num_classes=box_predictor.num_classes)
self._is_training = is_training
self._freeze_batchnorm = freeze_batchnorm
self._inplace_batchnorm_update = inplace_batchnorm_update
# Needed for fine-tuning from classification checkpoints whose
# variables do not have the feature extractor scope.
self._extract_features_scope = 'FeatureExtractor'
self._anchor_generator = anchor_generator
self._box_predictor = box_predictor
self._box_coder = box_coder
self._feature_extractor = feature_extractor
self._matcher = matcher
self._region_similarity_calculator = region_similarity_calculator
self._add_background_class = add_background_class
# TODO(jonathanhuang): handle agnostic mode
# weights
unmatched_cls_target = None
unmatched_cls_target = tf.constant([1] + self.num_classes * [0],
tf.float32)
if encode_background_as_zeros:
unmatched_cls_target = tf.constant((self.num_classes + 1) * [0],
tf.float32)
self._target_assigner = target_assigner.TargetAssigner(
self._region_similarity_calculator,
self._matcher,
self._box_coder,
negative_class_weight=negative_class_weight,
unmatched_cls_target=unmatched_cls_target)
self._classification_loss = classification_loss
self._localization_loss = localization_loss
self._classification_loss_weight = classification_loss_weight
self._localization_loss_weight = localization_loss_weight
self._normalize_loss_by_num_matches = normalize_loss_by_num_matches
self._normalize_loc_loss_by_codesize = normalize_loc_loss_by_codesize
self._hard_example_miner = hard_example_miner
self._random_example_sampler = random_example_sampler
self._parallel_iterations = 16
self._image_resizer_fn = image_resizer_fn
self._non_max_suppression_fn = non_max_suppression_fn
self._score_conversion_fn = score_conversion_fn
self._anchors = None
self._add_summaries = add_summaries
self._batched_prediction_tensor_names = []
def _build_ssd_model(ssd_config, is_training, add_summaries,
add_background_class=True):
"""Builds an SSD detection model based on the model config.
Args:
ssd_config: A ssd.proto object containing the config for the desired
SSDMetaArch.
is_training: True if this model is being built for training purposes.
add_summaries: Whether to add tf summaries in the model.
add_background_class: Whether to add an implicit background class to one-hot
encodings of groundtruth labels. Set to false if using groundtruth labels
with an explicit background class or using multiclass scores instead of
truth in the case of distillation.
Returns:
SSDMetaArch based on the config.
Raises:
ValueError: If ssd_config.type is not recognized (i.e. not registered in
model_class_map).
"""
num_classes = ssd_config.num_classes
# Feature extractor
feature_extractor = _build_ssd_feature_extractor(
feature_extractor_config=ssd_config.feature_extractor,
is_training=is_training)
box_coder = box_coder_builder.build(ssd_config.box_coder)
matcher = matcher_builder.build(ssd_config.matcher)
region_similarity_calculator = sim_calc.build(
ssd_config.similarity_calculator)
encode_background_as_zeros = ssd_config.encode_background_as_zeros
negative_class_weight = ssd_config.negative_class_weight
ssd_box_predictor = box_predictor_builder.build(hyperparams_builder.build,
ssd_config.box_predictor,
is_training, num_classes)
anchor_generator = anchor_generator_builder.build(
ssd_config.anchor_generator)
image_resizer_fn = image_resizer_builder.build(ssd_config.image_resizer)
non_max_suppression_fn, score_conversion_fn = post_processing_builder.build(
ssd_config.post_processing)
(classification_loss, localization_loss, classification_weight,
localization_weight, hard_example_miner,
random_example_sampler) = losses_builder.build(ssd_config.loss)
normalize_loss_by_num_matches = ssd_config.normalize_loss_by_num_matches
normalize_loc_loss_by_codesize = ssd_config.normalize_loc_loss_by_codesize
weight_regression_loss_by_score = ssd_config.weight_regression_loss_by_score
target_assigner_instance = target_assigner.TargetAssigner(
region_similarity_calculator,
matcher,
box_coder,
negative_class_weight=negative_class_weight,
weight_regression_loss_by_score=weight_regression_loss_by_score)
expected_classification_loss_under_sampling = None
if ssd_config.use_expected_classification_loss_under_sampling:
expected_classification_loss_under_sampling = functools.partial(
ops.expected_classification_loss_under_sampling,
minimum_negative_sampling=ssd_config.minimum_negative_sampling,
desired_negative_sampling_ratio=ssd_config.desired_negative_sampling_ratio)
ssd_meta_arch_fn = ssd_meta_arch.SSDMetaArch
return ssd_meta_arch_fn(
is_training=is_training,
anchor_generator=anchor_generator,
box_predictor=ssd_box_predictor,
box_coder=box_coder,
feature_extractor=feature_extractor,
encode_background_as_zeros=encode_background_as_zeros,
image_resizer_fn=image_resizer_fn,
non_max_suppression_fn=non_max_suppression_fn,
score_conversion_fn=score_conversion_fn,
classification_loss=classification_loss,
localization_loss=localization_loss,
classification_loss_weight=classification_weight,
localization_loss_weight=localization_weight,
normalize_loss_by_num_matches=normalize_loss_by_num_matches,
hard_example_miner=hard_example_miner,
target_assigner_instance=target_assigner_instance,
add_summaries=add_summaries,
normalize_loc_loss_by_codesize=normalize_loc_loss_by_codesize,
freeze_batchnorm=ssd_config.freeze_batchnorm,
inplace_batchnorm_update=ssd_config.inplace_batchnorm_update,
add_background_class=add_background_class,
random_example_sampler=random_example_sampler,
expected_classification_loss_under_sampling=expected_classification_loss_under_sampling)