class ResnetShapeTest(test_case.TestCase, parameterized.TestCase):
@unittest.skipIf(tf_version.is_tf1(), 'Skipping TF2.X only test.')
@parameterized.parameters(
{
'resnet_type':
'resnet_v1_34',
'output_layer_names': [
'conv2_block3_out', 'conv3_block4_out', 'conv4_block6_out',
'conv5_block3_out'
]
}, {
'resnet_type':
'resnet_v1_18',
'output_layer_names': [
'conv2_block2_out', 'conv3_block2_out', 'conv4_block2_out',
'conv5_block2_out'
]
})
def test_output_shapes(self, resnet_type, output_layer_names):
if resnet_type == 'resnet_v1_34':
model = resnet_v1.resnet_v1_34(input_shape=(64, 64, 3),
weights=None)
else:
model = resnet_v1.resnet_v1_18(input_shape=(64, 64, 3),
weights=None)
outputs = [
model.get_layer(output_layer_name).output
for output_layer_name in output_layer_names
]
resnet_model = tf.keras.models.Model(inputs=model.input,
outputs=outputs)
outputs = resnet_model(np.zeros((2, 64, 64, 3), dtype=np.float32))
# Check the shape of 'conv2_block3_out':
self.assertEqual(outputs[0].shape, [2, 16, 16, 64])
# Check the shape of 'conv3_block4_out':
self.assertEqual(outputs[1].shape, [2, 8, 8, 128])
# Check the shape of 'conv4_block6_out':
self.assertEqual(outputs[2].shape, [2, 4, 4, 256])
# Check the shape of 'conv5_block3_out':
self.assertEqual(outputs[3].shape, [2, 2, 2, 512])
class ResnetShapeTest(test_case.TestCase, parameterized.TestCase):
@unittest.skipIf(tf_version.is_tf1(), 'Skipping TF2.X only test.')
@parameterized.parameters(
{
'resnet_type':
'resnet_v1_34',
'output_layer_names': [
'conv2_block3_out', 'conv3_block4_out', 'conv4_block6_out',
'conv5_block3_out'
]
}, {
'resnet_type':
'resnet_v1_18',
'output_layer_names': [
'conv2_block2_out', 'conv3_block2_out', 'conv4_block2_out',
'conv5_block2_out'
]
})
def test_output_shapes(self, resnet_type, output_layer_names):
if resnet_type == 'resnet_v1_34':
from object_detection.models import center_net_mobilenet_v2_feature_extractor
from object_detection.models import center_net_resnet_feature_extractor
from object_detection.models import center_net_resnet_v1_fpn_feature_extractor
from object_detection.models import faster_rcnn_inception_resnet_v2_keras_feature_extractor as frcnn_inc_res_keras
from object_detection.models import faster_rcnn_resnet_keras_feature_extractor as frcnn_resnet_keras
from object_detection.models import ssd_resnet_v1_fpn_keras_feature_extractor as ssd_resnet_v1_fpn_keras
from object_detection.models import faster_rcnn_resnet_v1_fpn_keras_feature_extractor as frcnn_resnet_fpn_keras
from object_detection.models.ssd_mobilenet_v1_fpn_keras_feature_extractor import SSDMobileNetV1FpnKerasFeatureExtractor
from object_detection.models.ssd_mobilenet_v1_keras_feature_extractor import SSDMobileNetV1KerasFeatureExtractor
from object_detection.models.ssd_mobilenet_v2_fpn_keras_feature_extractor import SSDMobileNetV2FpnKerasFeatureExtractor
from object_detection.models.ssd_mobilenet_v2_keras_feature_extractor import SSDMobileNetV2KerasFeatureExtractor
from object_detection.predictors import rfcn_keras_box_predictor
if sys.version_info[0] >= 3:
from object_detection.models import ssd_efficientnet_bifpn_feature_extractor as ssd_efficientnet_bifpn
if tf_version.is_tf1():
from object_detection.models import faster_rcnn_inception_resnet_v2_feature_extractor as frcnn_inc_res
from object_detection.models import faster_rcnn_inception_v2_feature_extractor as frcnn_inc_v2
from object_detection.models import faster_rcnn_nas_feature_extractor as frcnn_nas
from object_detection.models import faster_rcnn_pnas_feature_extractor as frcnn_pnas
from object_detection.models import faster_rcnn_resnet_v1_feature_extractor as frcnn_resnet_v1
from object_detection.models import ssd_resnet_v1_fpn_feature_extractor as ssd_resnet_v1_fpn
from object_detection.models import ssd_resnet_v1_ppn_feature_extractor as ssd_resnet_v1_ppn
from object_detection.models.embedded_ssd_mobilenet_v1_feature_extractor import EmbeddedSSDMobileNetV1FeatureExtractor
from object_detection.models.ssd_inception_v2_feature_extractor import SSDInceptionV2FeatureExtractor
from object_detection.models.ssd_mobilenet_v2_fpn_feature_extractor import SSDMobileNetV2FpnFeatureExtractor
from object_detection.models.ssd_mobilenet_v2_mnasfpn_feature_extractor import SSDMobileNetV2MnasFPNFeatureExtractor
from object_detection.models.ssd_inception_v3_feature_extractor import SSDInceptionV3FeatureExtractor
from object_detection.models.ssd_mobilenet_edgetpu_feature_extractor import SSDMobileNetEdgeTPUFeatureExtractor
from object_detection.models.ssd_mobilenet_v1_feature_extractor import SSDMobileNetV1FeatureExtractor
from object_detection.models.ssd_mobilenet_v1_fpn_feature_extractor import SSDMobileNetV1FpnFeatureExtractor
}
"""
conv_hyperparams_proto = hyperparams_pb2.Hyperparams()
text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams_proto)
scope_fn = hyperparams_builder.build(conv_hyperparams_proto,
is_training=True)
scope = scope_fn()
conv_scope_arguments = scope[_get_scope_key(slim.conv2d)]
initializer = conv_scope_arguments['weights_initializer']
self._assert_variance_in_range(initializer,
shape=[100, 40],
variance=0.64,
tol=1e-1)
@unittest.skipIf(tf_version.is_tf1(), 'Skipping TF2.X only tests.')
class KerasHyperparamsBuilderTest(tf.test.TestCase):
def _assert_variance_in_range(self,
initializer,
shape,
variance,
tol=1e-2):
var = tf.Variable(initializer(shape=shape, dtype=tf.float32))
self.assertAllClose(np.var(var.numpy()), variance, tol, tol)
def test_return_l1_regularized_weights_keras(self):
conv_hyperparams_text_proto = """
regularizer {
l1_regularizer {
weight: 0.5
}
def __init__(self,
is_training,
num_classes,
image_resizer_fn,
feature_extractor,
number_of_stages,
first_stage_anchor_generator,
first_stage_target_assigner,
first_stage_atrous_rate,
first_stage_box_predictor_arg_scope_fn,
first_stage_box_predictor_kernel_size,
first_stage_box_predictor_depth,
first_stage_minibatch_size,
first_stage_sampler,
first_stage_non_max_suppression_fn,
first_stage_max_proposals,
first_stage_localization_loss_weight,
first_stage_objectness_loss_weight,
crop_and_resize_fn,
initial_crop_size,
maxpool_kernel_size,
maxpool_stride,
second_stage_target_assigner,
second_stage_mask_rcnn_box_predictor,
second_stage_batch_size,
second_stage_sampler,
second_stage_non_max_suppression_fn,
second_stage_score_conversion_fn,
second_stage_localization_loss_weight,
second_stage_classification_loss_weight,
second_stage_classification_loss,
second_stage_mask_prediction_loss_weight=1.0,
hard_example_miner=None,
parallel_iterations=16,
add_summaries=True,
clip_anchors_to_image=False,
use_static_shapes=False,
resize_masks=True,
freeze_batchnorm=False,
return_raw_detections_during_predict=False,
output_final_box_features=False,
attention_bottleneck_dimension=None,
attention_temperature=None):
"""ContextRCNNMetaArch Constructor.
Args:
is_training: A boolean indicating whether the training version of the
computation graph should be constructed.
num_classes: Number of classes. Note that num_classes *does not*
include the background category, so if groundtruth labels take values
in {0, 1, .., K-1}, num_classes=K (and not K+1, even though the
assigned classification targets can range from {0,... K}).
image_resizer_fn: A callable for image resizing. This callable
takes a rank-3 image tensor of shape [height, width, channels]
(corresponding to a single image), an optional rank-3 instance mask
tensor of shape [num_masks, height, width] and returns a resized rank-3
image tensor, a resized mask tensor if one was provided in the input. In
addition this callable must also return a 1-D tensor of the form
[height, width, channels] containing the size of the true image, as the
image resizer can perform zero padding. See protos/image_resizer.proto.
feature_extractor: A FasterRCNNFeatureExtractor object.
number_of_stages: An integer values taking values in {1, 2, 3}. If
1, the function will construct only the Region Proposal Network (RPN)
part of the model. If 2, the function will perform box refinement and
other auxiliary predictions all in the second stage. If 3, it will
extract features from refined boxes and perform the auxiliary
predictions on the non-maximum suppressed refined boxes.
If is_training is true and the value of number_of_stages is 3, it is
reduced to 2 since all the model heads are trained in parallel in second
stage during training.
first_stage_anchor_generator: An anchor_generator.AnchorGenerator object
(note that currently we only support
grid_anchor_generator.GridAnchorGenerator objects)
first_stage_target_assigner: Target assigner to use for first stage of
Faster R-CNN (RPN).
first_stage_atrous_rate: A single integer indicating the atrous rate for
the single convolution op which is applied to the `rpn_features_to_crop`
tensor to obtain a tensor to be used for box prediction. Some feature
extractors optionally allow for producing feature maps computed at
denser resolutions. The atrous rate is used to compensate for the
denser feature maps by using an effectively larger receptive field.
(This should typically be set to 1).
first_stage_box_predictor_arg_scope_fn: Either a
Keras layer hyperparams object or a function to construct tf-slim
arg_scope for conv2d, separable_conv2d and fully_connected ops. Used
for the RPN box predictor. If it is a keras hyperparams object the
RPN box predictor will be a Keras model. If it is a function to
construct an arg scope it will be a tf-slim box predictor.
first_stage_box_predictor_kernel_size: Kernel size to use for the
convolution op just prior to RPN box predictions.
first_stage_box_predictor_depth: Output depth for the convolution op
just prior to RPN box predictions.
first_stage_minibatch_size: The "batch size" to use for computing the
objectness and location loss of the region proposal network. This
"batch size" refers to the number of anchors selected as contributing
to the loss function for any given image within the image batch and is
only called "batch_size" due to terminology from the Faster R-CNN paper.
first_stage_sampler: Sampler to use for first stage loss (RPN loss).
first_stage_non_max_suppression_fn: batch_multiclass_non_max_suppression
callable that takes `boxes`, `scores` and optional `clip_window`(with
all other inputs already set) and returns a dictionary containing
tensors with keys: `detection_boxes`, `detection_scores`,
`detection_classes`, `num_detections`. This is used to perform non max
suppression on the boxes predicted by the Region Proposal Network
(RPN).
See `post_processing.batch_multiclass_non_max_suppression` for the type
and shape of these tensors.
first_stage_max_proposals: Maximum number of boxes to retain after
performing Non-Max Suppression (NMS) on the boxes predicted by the
Region Proposal Network (RPN).
first_stage_localization_loss_weight: A float
first_stage_objectness_loss_weight: A float
crop_and_resize_fn: A differentiable resampler to use for cropping RPN
proposal features.
initial_crop_size: A single integer indicating the output size
(width and height are set to be the same) of the initial bilinear
interpolation based cropping during ROI pooling.
maxpool_kernel_size: A single integer indicating the kernel size of the
max pool op on the cropped feature map during ROI pooling.
maxpool_stride: A single integer indicating the stride of the max pool
op on the cropped feature map during ROI pooling.
second_stage_target_assigner: Target assigner to use for second stage of
Faster R-CNN. If the model is configured with multiple prediction heads,
this target assigner is used to generate targets for all heads (with the
correct `unmatched_class_label`).
second_stage_mask_rcnn_box_predictor: Mask R-CNN box predictor to use for
the second stage.
second_stage_batch_size: The batch size used for computing the
classification and refined location loss of the box classifier. This
"batch size" refers to the number of proposals selected as contributing
to the loss function for any given image within the image batch and is
only called "batch_size" due to terminology from the Faster R-CNN paper.
second_stage_sampler: Sampler to use for second stage loss (box
classifier loss).
second_stage_non_max_suppression_fn: batch_multiclass_non_max_suppression
callable that takes `boxes`, `scores`, optional `clip_window` and
optional (kwarg) `mask` inputs (with all other inputs already set)
and returns a dictionary containing tensors with keys:
`detection_boxes`, `detection_scores`, `detection_classes`,
`num_detections`, and (optionally) `detection_masks`. See
`post_processing.batch_multiclass_non_max_suppression` for the type and
shape of these tensors.
second_stage_score_conversion_fn: Callable elementwise nonlinearity
(that takes tensors as inputs and returns tensors). This is usually
used to convert logits to probabilities.
second_stage_localization_loss_weight: A float indicating the scale factor
for second stage localization loss.
second_stage_classification_loss_weight: A float indicating the scale
factor for second stage classification loss.
second_stage_classification_loss: Classification loss used by the second
stage classifier. Either losses.WeightedSigmoidClassificationLoss or
losses.WeightedSoftmaxClassificationLoss.
second_stage_mask_prediction_loss_weight: A float indicating the scale
factor for second stage mask prediction loss. This is applicable only if
second stage box predictor is configured to predict masks.
hard_example_miner: A losses.HardExampleMiner object (can be None).
parallel_iterations: (Optional) The number of iterations allowed to run
in parallel for calls to tf.map_fn.
add_summaries: boolean (default: True) controlling whether summary ops
should be added to tensorflow graph.
clip_anchors_to_image: Normally, anchors generated for a given image size
are pruned during training if they lie outside the image window. This
option clips the anchors to be within the image instead of pruning.
use_static_shapes: If True, uses implementation of ops with static shape
guarantees.
resize_masks: Indicates whether the masks presend in the groundtruth
should be resized in the model with `image_resizer_fn`
freeze_batchnorm: Whether to freeze batch norm parameters in the first
stage box predictor 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.
return_raw_detections_during_predict: Whether to return raw detection
boxes in the predict() method. These are decoded boxes that have not
been through postprocessing (i.e. NMS). Default False.
output_final_box_features: Whether to output final box features. If true,
it crops the feauture map based on the final box prediction and returns
in the dict as detection_features.
attention_bottleneck_dimension: A single integer. The bottleneck feature
dimension of the attention block.
attention_temperature: A single float. The attention temperature.
Raises:
ValueError: If `second_stage_batch_size` > `first_stage_max_proposals` at
training time.
ValueError: If first_stage_anchor_generator is not of type
grid_anchor_generator.GridAnchorGenerator.
"""
super(ContextRCNNMetaArch, self).__init__(
is_training,
num_classes,
image_resizer_fn,
feature_extractor,
number_of_stages,
first_stage_anchor_generator,
first_stage_target_assigner,
first_stage_atrous_rate,
first_stage_box_predictor_arg_scope_fn,
first_stage_box_predictor_kernel_size,
first_stage_box_predictor_depth,
first_stage_minibatch_size,
first_stage_sampler,
first_stage_non_max_suppression_fn,
first_stage_max_proposals,
first_stage_localization_loss_weight,
first_stage_objectness_loss_weight,
crop_and_resize_fn,
initial_crop_size,
maxpool_kernel_size,
maxpool_stride,
second_stage_target_assigner,
second_stage_mask_rcnn_box_predictor,
second_stage_batch_size,
second_stage_sampler,
second_stage_non_max_suppression_fn,
second_stage_score_conversion_fn,
second_stage_localization_loss_weight,
second_stage_classification_loss_weight,
second_stage_classification_loss,
second_stage_mask_prediction_loss_weight=(
second_stage_mask_prediction_loss_weight),
hard_example_miner=hard_example_miner,
parallel_iterations=parallel_iterations,
add_summaries=add_summaries,
clip_anchors_to_image=clip_anchors_to_image,
use_static_shapes=use_static_shapes,
resize_masks=resize_masks,
freeze_batchnorm=freeze_batchnorm,
return_raw_detections_during_predict=(
return_raw_detections_during_predict),
output_final_box_features=output_final_box_features)
if tf_version.is_tf1():
self._context_feature_extract_fn = functools.partial(
context_rcnn_lib.compute_box_context_attention,
bottleneck_dimension=attention_bottleneck_dimension,
attention_temperature=attention_temperature,
is_training=is_training)
else:
self._context_feature_extract_fn = context_rcnn_lib_tf2.AttentionBlock(
bottleneck_dimension=attention_bottleneck_dimension,
attention_temperature=attention_temperature,
is_training=is_training)
np.array([0, 0, 0]),
fields.InputDataFields.groundtruth_not_exhaustive_classes:
np.array([0, 0, 0])
})
lvis_evaluator.add_single_detected_image_info(
image_id='image3',
detections_dict={
fields.DetectionResultFields.detection_masks: masks3,
fields.DetectionResultFields.detection_scores: np.array([.8]),
fields.DetectionResultFields.detection_classes: np.array([1])
})
metrics = lvis_evaluator.evaluate()
self.assertAlmostEqual(metrics['DetectionMasks_AP'], 1.0)
@unittest.skipIf(tf_version.is_tf1(), 'Only Supported in TF2.X')
class LVISMaskEvaluationPyFuncTest(tf.test.TestCase):
def testAddEvalDict(self):
lvis_evaluator = lvis_evaluation.LVISMaskEvaluator(
_get_categories_list())
image_id = tf.constant('image1', dtype=tf.string)
groundtruth_boxes = tf.constant(np.array([[100., 100., 200., 200.],
[50., 50., 100., 100.]]),
dtype=tf.float32)
groundtruth_classes = tf.constant(np.array([1, 2]), dtype=tf.float32)
groundtruth_masks = tf.constant(np.stack([
np.pad(np.ones([100, 100], dtype=np.uint8), ((10, 10), (10, 10)),
mode='constant'),
np.pad(np.ones([50, 50], dtype=np.uint8), ((0, 70), (0, 70)),
mode='constant')
]),
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Provides functions to prefetch tensors to feed into models."""
import tensorflow as tf
from object_detection.utils import tf_version
if not tf_version.is_tf1():
raise ValueError('`prefetcher.py` is only supported in Tensorflow 1.X')
def prefetch(tensor_dict, capacity):
"""Creates a prefetch queue for tensors.
Creates a FIFO queue to asynchronously enqueue tensor_dicts and returns a
dequeue op that evaluates to a tensor_dict. This function is useful in
prefetching preprocessed tensors so that the data is readily available for
consumers.
Example input pipeline when you don't need batching:
----------------------------------------------------
key, string_tensor = slim.parallel_reader.parallel_read(...)
tensor_dict = decoder.decode(string_tensor)