def test_resnet_50layer(self):
feature_extractor_text_proto = """
resnet {
net_type: SINGLE_BRANCH
net_depth: RESNET_50
conv_hyperparams {
op: CONV
regularizer { l2_regularizer { weight: 1e-4 } }
initializer { variance_scaling_initializer { } }
batch_norm { }
}
summarize_activations: false
}
"""
convnet_proto = convnet_pb2.Convnet()
text_format.Merge(feature_extractor_text_proto, convnet_proto)
convnet_object = convnet_builder.build(convnet_proto, True)
self.assertTrue(isinstance(convnet_object, resnet.Resnet50Layer))
test_image_shape = [2, 32, 128, 3]
test_input_image = tf.random_uniform(test_image_shape,
minval=0,
maxval=255.0,
dtype=tf.float32,
seed=1)
feature_maps = convnet_object.extract_features(test_input_image)
self.assertTrue(len(feature_maps) == 1)
print('Outputs of test_resnet_single_branch: {}'.format(feature_maps))
def test_crnn_net_three_branches(self):
feature_extractor_text_proto = """
crnn_net {
net_type: THREE_BRANCHES
conv_hyperparams {
op: CONV
regularizer { l2_regularizer { weight: 1e-4 } }
initializer { variance_scaling_initializer { } }
batch_norm { }
}
summarize_activations: false
}
"""
convnet_proto = convnet_pb2.Convnet()
text_format.Merge(feature_extractor_text_proto, convnet_proto)
convnet_object = convnet_builder.build(convnet_proto, True)
self.assertTrue(isinstance(convnet_object, crnn_net.CrnnNet))
test_image_shape = [2, 32, 128, 3]
test_input_image = tf.random_uniform(test_image_shape,
minval=0,
maxval=255.0,
dtype=tf.float32,
seed=1)
feature_maps = convnet_object.extract_features(test_input_image)
self.assertTrue(len(feature_maps) == 3)
print(
'Outputs of test_crnn_net_three_branches: {}'.format(feature_maps))
def test_build_stn_convnet_tiny(self):
text_proto = """
stn_convnet {
conv_hyperparams {
op: CONV
regularizer { l2_regularizer { weight: 1e-4 } }
initializer { variance_scaling_initializer { } }
batch_norm { decay: 0.99 }
}
tiny: true
}
"""
convnet_proto = convnet_pb2.Convnet()
text_format.Merge(text_proto, convnet_proto)
convnet_object = convnet_builder.build(convnet_proto, True)
self.assertTrue(isinstance(convnet_object, stn_convnet.StnConvnetTiny))
test_image_shape = [2, 64, 128, 3]
test_input_image = tf.random_uniform(test_image_shape,
minval=0,
maxval=255.0,
dtype=tf.float32,
seed=1)
feature_maps = convnet_object.extract_features(test_input_image)
self.assertTrue(len(feature_maps) == 1)
print(
'Outputs of test_build_stn_convnet_tiny: {}'.format(feature_maps))
def build(config, is_training):
if not isinstance(config, feature_extractor_pb2.FeatureExtractor):
raise ValueError(
'config not of type feature_extractor_pb2.FeatureExtractor')
convnet_object = convnet_builder.build(config.convnet, is_training)
brnn_fn_list = [
functools.partial(bidirectional_rnn_builder.build, brnn_config,
is_training)
for brnn_config in config.bidirectional_rnn
]
feature_extractor_object = feature_extractor.FeatureExtractor(
convnet=convnet_object,
brnn_fn_list=brnn_fn_list,
summarize_activations=config.summarize_activations,
is_training=is_training)
return feature_extractor_object
def build(config, is_training):
if not isinstance(config, spatial_transformer_pb2.SpatialTransformer):
raise ValueError(
'config not of type spatial_transformer_pb2.SpatialTransformer')
convnet_object = convnet_builder.build(config.convnet, is_training)
fc_hyperparams_object = hyperparams_builder.build(config.fc_hyperparams,
is_training)
return spatial_transformer.SpatialTransformer(
convnet=convnet_object,
fc_hyperparams=fc_hyperparams_object,
localization_image_size=(config.localization_h, config.localization_w),
output_image_size=(config.output_h, config.output_w),
num_control_points=config.num_control_points,
init_bias_pattern=config.init_bias_pattern,
margins=(config.margin_x, config.margin_y),
activation=config.activation,
summarize_activations=config.summarize_activations)