def _create_application_with_layer_outputs(
self,
layer_names,
batchnorm_training,
output_stride=16,
align_feature_maps=False,
batchnorm_scale=False,
weight_decay=0.00004,
default_batchnorm_momentum=0.9997,
default_batchnorm_epsilon=0.001,
):
"""Constructs Keras inception_resnet_v2 that extracts layer outputs."""
# Have to clear the Keras backend to ensure isolation in layer naming
tf.keras.backend.clear_session()
if not layer_names:
layer_names = _KERAS_LAYERS_TO_CHECK
full_model = inception_resnet_v2.inception_resnet_v2(
batchnorm_training=batchnorm_training,
output_stride=output_stride,
align_feature_maps=align_feature_maps,
weights=None,
batchnorm_scale=batchnorm_scale,
weight_decay=weight_decay,
default_batchnorm_momentum=default_batchnorm_momentum,
default_batchnorm_epsilon=default_batchnorm_epsilon,
include_top=False)
layer_outputs = [
full_model.get_layer(name=layer).output for layer in layer_names
]
return tf.keras.Model(inputs=full_model.inputs, outputs=layer_outputs)
def test_hyperparam_override(self):
model = inception_resnet_v2.inception_resnet_v2(
batchnorm_training=True,
default_batchnorm_momentum=0.2,
default_batchnorm_epsilon=0.1,
weights=None,
include_top=False)
bn_layer = model.get_layer(name='freezable_batch_norm')
self.assertAllClose(bn_layer.momentum, 0.2)
self.assertAllClose(bn_layer.epsilon, 0.1)
def get_box_classifier_feature_extractor_model(self, name=None):
"""Returns a model that extracts second stage box classifier features.
This function reconstructs the "second half" of the Inception ResNet v2
network after the part defined in `get_proposal_feature_extractor_model`.
Args:
name: A scope name to construct all variables within.
Returns:
A Keras model that takes proposal_feature_maps:
A 4-D float tensor with shape
[batch_size * self.max_num_proposals, crop_height, crop_width, depth]
representing the feature map cropped to each proposal.
And returns proposal_classifier_features:
A 4-D float tensor with shape
[batch_size * self.max_num_proposals, height, width, depth]
representing box classifier features for each proposal.
"""
if not self.classification_backbone:
self.classification_backbone = inception_resnet_v2.inception_resnet_v2(
self._train_batch_norm,
output_stride=self._first_stage_features_stride,
align_feature_maps=True,
weight_decay=self._weight_decay,
weights=None,
include_top=False)
with tf.name_scope(name):
with tf.name_scope('InceptionResnetV2'):
proposal_feature_maps = self.classification_backbone.get_layer(
name='block17_20_ac').output
proposal_classifier_features = self.classification_backbone.get_layer(
name='conv_7b_ac').output
keras_model = model_util.extract_submodel(
model=self.classification_backbone,
inputs=proposal_feature_maps,
outputs=proposal_classifier_features)
for variable in keras_model.variables:
self._variable_dict[variable.name[:-2]] = variable
return keras_model
def get_proposal_feature_extractor_model(self, name=None):
"""Returns a model that extracts first stage RPN features.
Extracts features using the first half of the Inception Resnet v2 network.
We construct the network in `align_feature_maps=True` mode, which means
that all VALID paddings in the network are changed to SAME padding so that
the feature maps are aligned.
Args:
name: A scope name to construct all variables within.
Returns:
A Keras model that takes preprocessed_inputs:
A [batch, height, width, channels] float32 tensor
representing a batch of images.
And returns rpn_feature_map:
A tensor with shape [batch, height, width, depth]
"""
if not self.classification_backbone:
self.classification_backbone = inception_resnet_v2.inception_resnet_v2(
self._train_batch_norm,
output_stride=self._first_stage_features_stride,
align_feature_maps=True,
weight_decay=self._weight_decay,
weights=None,
include_top=False)
with tf.name_scope(name):
with tf.name_scope('InceptionResnetV2'):
proposal_features = self.classification_backbone.get_layer(
name='block17_20_ac').output
keras_model = tf.keras.Model(
inputs=self.classification_backbone.inputs,
outputs=proposal_features)
for variable in keras_model.variables:
self._variable_dict[variable.name[:-2]] = variable
return keras_model