def test_get_expected_feature_map_shapes_with_inception_v3(self):
image_features = {
'Mixed_5d': tf.random_uniform([4, 35, 35, 256], dtype=tf.float32),
'Mixed_6e': tf.random_uniform([4, 17, 17, 576], dtype=tf.float32),
'Mixed_7c': tf.random_uniform([4, 8, 8, 1024], dtype=tf.float32)
}
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=INCEPTION_V3_LAYOUT,
depth_multiplier=1,
min_depth=32,
insert_1x1_conv=True,
image_features=image_features)
expected_feature_map_shapes = {
'Mixed_5d': (4, 35, 35, 256),
'Mixed_6e': (4, 17, 17, 576),
'Mixed_7c': (4, 8, 8, 1024),
'Mixed_7c_2_Conv2d_3_3x3_s2_512': (4, 4, 4, 512),
'Mixed_7c_2_Conv2d_4_3x3_s2_256': (4, 2, 2, 256),
'Mixed_7c_2_Conv2d_5_3x3_s2_128': (4, 1, 1, 128)}
init_op = tf.global_variables_initializer()
with self.test_session() as sess:
sess.run(init_op)
out_feature_maps = sess.run(feature_maps)
out_feature_map_shapes = dict(
(key, value.shape) for key, value in out_feature_maps.items())
self.assertDictEqual(out_feature_map_shapes, expected_feature_map_shapes)
def feature_map_generator(image_features):
return feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=1,
min_depth=32,
insert_1x1_conv=True,
image_features=image_features)
def extract_features(self, preprocessed_inputs, state_saver=None,
state_name='lstm_state', unroll_length=5, scope=None):
with tf.variable_scope('mock_model'):
net_large = self.extract_base_features_large(preprocessed_inputs)
net_small = self.extract_base_features_small(preprocessed_inputs)
net = slim.conv2d(
inputs=tf.concat([net_large, net_small], axis=3),
num_outputs=32,
kernel_size=1,
scope='layer1')
image_features = {'last_layer': net}
self._states_out = {}
feature_map_layout = {
'from_layer': ['last_layer'],
'layer_depth': [-1],
'use_explicit_padding': self._use_explicit_padding,
'use_depthwise': self._use_depthwise,
}
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=(self._depth_multiplier),
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return list(feature_maps.values())
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
'''
feature_map_layout = {
'from_layer': ['Conv2d_11_pointwise', 'Conv2d_13_pointwise', '', '',
'', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
'use_explicit_padding': self._use_explicit_padding,
'use_depthwise': self._use_depthwise,
}
'''
feature_map_layout = {
'from_layer': [
'FeatureExtractor/vgg_16/conv4/conv4_3',
'FeatureExtractor/vgg_16/fc7', '', '', '', ''
],
'layer_depth': [-1, -1, 256, 128, 128, 128],
'use_explicit_padding':
self._use_explicit_padding,
'use_depthwise':
self._use_depthwise,
}
with tf.variable_scope('vgg_16', reuse=self._reuse_weights) as scope:
with slim.arg_scope(vgg.vgg_arg_scope()):
with (slim.arg_scope(self._conv_hyperparams_fn())
if self._override_base_feature_extractor_hyperparams else
context_manager.IdentityContextManager()):
# TODO(skligys): Enable fused batch norm once quantization supports it.
with slim.arg_scope([slim.batch_norm], fused=False):
_, image_features = vgg.vgg_16(
ops.pad_to_multiple(preprocessed_inputs,
self._pad_to_multiple),
num_classes=None,
is_training=self._is_training,
scope=scope)
print(image_features.keys())
print(image_features.values())
with slim.arg_scope(self._conv_hyperparams_fn()):
# TODO(skligys): Enable fused batch norm once quantization supports it.
with slim.arg_scope([slim.batch_norm], fused=False):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def test_get_expected_feature_map_shapes_with_inception_v3(self):
image_features = {
'Mixed_5d': tf.random_uniform([4, 35, 35, 256], dtype=tf.float32),
'Mixed_6e': tf.random_uniform([4, 17, 17, 576], dtype=tf.float32),
'Mixed_7c': tf.random_uniform([4, 8, 8, 1024], dtype=tf.float32)
}
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=INCEPTION_V3_LAYOUT,
depth_multiplier=1,
min_depth=32,
insert_1x1_conv=True,
image_features=image_features)
expected_feature_map_shapes = {
'Mixed_5d': (4, 35, 35, 256),
'Mixed_6e': (4, 17, 17, 576),
'Mixed_7c': (4, 8, 8, 1024),
'Mixed_7c_2_Conv2d_3_3x3_s2_512': (4, 4, 4, 512),
'Mixed_7c_2_Conv2d_4_3x3_s2_256': (4, 2, 2, 256),
'Mixed_7c_2_Conv2d_5_3x3_s2_128': (4, 1, 1, 128)
}
init_op = tf.global_variables_initializer()
with self.test_session() as sess:
sess.run(init_op)
out_feature_maps = sess.run(feature_maps)
out_feature_map_shapes = dict(
(key, value.shape) for key, value in out_feature_maps.items())
self.assertDictEqual(out_feature_map_shapes,
expected_feature_map_shapes)
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
Raises:
ValueError if conv_defs is not provided or from_layer does not meet the
size requirement.
"""
if not self._conv_defs:
raise ValueError('Must provide backbone conv defs.')
if len(self._from_layer) != 2:
raise ValueError('SSD input feature names are not provided.')
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
feature_map_layout = {
'from_layer':
[self._from_layer[0], self._from_layer[1], '', '', '', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
'use_depthwise':
self._use_depthwise,
'use_explicit_padding':
self._use_explicit_padding,
}
with tf.variable_scope(self._scope_name,
reuse=self._reuse_weights) as scope:
with slim.arg_scope(
mobilenet_v3.training_scope(is_training=None, bn_decay=0.9997)), \
slim.arg_scope(
[mobilenet.depth_multiplier], min_depth=self._min_depth):
with (slim.arg_scope(self._conv_hyperparams_fn())
if self._override_base_feature_extractor_hyperparams else
context_manager.IdentityContextManager()):
_, image_features = mobilenet_v3.mobilenet_base(
ops.pad_to_multiple(preprocessed_inputs,
self._pad_to_multiple),
conv_defs=self._conv_defs,
final_endpoint=self._from_layer[1],
depth_multiplier=self._depth_multiplier,
use_explicit_padding=self._use_explicit_padding,
scope=scope)
with slim.arg_scope(self._conv_hyperparams_fn()):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return list(feature_maps.values())
def test_get_expected_feature_map_shapes_with_embedded_ssd_mobilenet_v1(
self):
image_features = {
'Conv2d_11_pointwise':
tf.random_uniform([4, 16, 16, 512], dtype=tf.float32),
'Conv2d_13_pointwise':
tf.random_uniform([4, 8, 8, 1024], dtype=tf.float32),
}
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=EMBEDDED_SSD_MOBILENET_V1_LAYOUT,
depth_multiplier=1,
min_depth=32,
insert_1x1_conv=True,
image_features=image_features)
expected_feature_map_shapes = {
'Conv2d_11_pointwise': (4, 16, 16, 512),
'Conv2d_13_pointwise': (4, 8, 8, 1024),
'Conv2d_13_pointwise_2_Conv2d_2_3x3_s2_512': (4, 4, 4, 512),
'Conv2d_13_pointwise_2_Conv2d_3_3x3_s2_256': (4, 2, 2, 256),
'Conv2d_13_pointwise_2_Conv2d_4_2x2_s2_256': (4, 1, 1, 256)
}
init_op = tf.global_variables_initializer()
with self.test_session() as sess:
sess.run(init_op)
out_feature_maps = sess.run(feature_maps)
out_feature_map_shapes = dict(
(key, value.shape) for key, value in out_feature_maps.items())
self.assertDictEqual(out_feature_map_shapes,
expected_feature_map_shapes)
def extract_features(self, preprocessed_inputs, fpn=0):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
feature_map_layout = {
'from_layer': ['block2/unit_8', 'block3/unit_4', '', '', '', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
}
with slim.arg_scope(self._conv_hyperparams):
with slim.arg_scope([slim.batch_norm], fused=False):
image_features = shufflenet_v1_base(preprocessed_inputs)
for i, j in image_features.items():
print i, j.get_shape()
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs, audio_inputs=None, preprocessed_flows=None):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
preprocessed_flows = shape_utils.check_min_image_dim(
33, preprocessed_flows)
feature_map_layout = {
'from_layer': ['Mixed_5d', 'Mixed_6e', 'Mixed_7c', '', '', ''],
'layer_depth': [-1, -1, -1, 512, 256, 128],
'use_explicit_padding': self._use_explicit_padding,
'use_depthwise': self._use_depthwise,
}
with slim.arg_scope(self._conv_hyperparams_fn()):
with tf.variable_scope('InceptionV3', reuse=self._reuse_weights) as scope:
if (audio_inputs != None):
audio_features = self.extract_audio_features(audio_inputs)
else:
audio_features = None
"""
_, image_features = inception_v3.inception_v3_base(
ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
final_endpoint='Mixed_7c',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
"""
_, image_features = inception_v3_m2.inception_v3_base(
ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
ops.pad_to_multiple(preprocessed_flows, self._pad_to_multiple),
final_endpoint='Mixed_7c',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
# A list of multi-resoulition feature maps
# mixed_5d, mixed_6e, ...
#print (len(feature_maps.values()))
return feature_maps.values(), audio_features
def feature_map_generator(image_features):
return feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=1,
min_depth=32,
insert_1x1_conv=True,
image_features=image_features)
def test_get_expected_feature_map_shapes_with_embedded_ssd_mobilenet_v1(
self):
image_features = {
'Conv2d_11_pointwise': tf.random_uniform([4, 16, 16, 512],
dtype=tf.float32),
'Conv2d_13_pointwise': tf.random_uniform([4, 8, 8, 1024],
dtype=tf.float32),
}
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=EMBEDDED_SSD_MOBILENET_V1_LAYOUT,
depth_multiplier=1,
min_depth=32,
insert_1x1_conv=True,
image_features=image_features)
expected_feature_map_shapes = {
'Conv2d_11_pointwise': (4, 16, 16, 512),
'Conv2d_13_pointwise': (4, 8, 8, 1024),
'Conv2d_13_pointwise_2_Conv2d_2_3x3_s2_512': (4, 4, 4, 512),
'Conv2d_13_pointwise_2_Conv2d_3_3x3_s2_256': (4, 2, 2, 256),
'Conv2d_13_pointwise_2_Conv2d_4_2x2_s2_256': (4, 1, 1, 256)}
init_op = tf.global_variables_initializer()
with self.test_session() as sess:
sess.run(init_op)
out_feature_maps = sess.run(feature_maps)
out_feature_map_shapes = dict(
(key, value.shape) for key, value in out_feature_maps.items())
self.assertDictEqual(out_feature_map_shapes, expected_feature_map_shapes)
def _extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
Raises:
ValueError: if image height or width are not 256 pixels.
"""
image_shape = preprocessed_inputs.get_shape()
image_shape.assert_has_rank(4)
image_height = image_shape[1].value
image_width = image_shape[2].value
if image_height is None or image_width is None:
shape_assert = tf.Assert(
tf.logical_and(tf.equal(tf.shape(preprocessed_inputs)[1], 256),
tf.equal(tf.shape(preprocessed_inputs)[2], 256)),
['image size must be 256 in both height and width.'])
with tf.control_dependencies([shape_assert]):
preprocessed_inputs = tf.identity(preprocessed_inputs)
elif image_height != 256 or image_width != 256:
raise ValueError('image size must be = 256 in both height and width;'
' image dim = %d,%d' % (image_height, image_width))
feature_map_layout = {
'from_layer': [
'Conv2d_11_pointwise', 'Conv2d_13_pointwise', '', '', ''
],
'layer_depth': [-1, -1, 512, 256, 256],
'conv_kernel_size': [-1, -1, 3, 3, 2],
'use_explicit_padding': self._use_explicit_padding,
'use_depthwise': self._use_depthwise,
}
with slim.arg_scope(self._conv_hyperparams):
with slim.arg_scope([slim.batch_norm], fused=False):
with tf.variable_scope('MobilenetV1',
reuse=self._reuse_weights) as scope:
_, image_features = mobilenet_v1.mobilenet_v1_base(
ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
final_endpoint='Conv2d_13_pointwise',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
feature_map_layout = {
'from_layer': [
'FeatureExtractor/{}/block3'.format(self._resnet_scope_name),
'FeatureExtractor/{}/block4'.format(self._resnet_scope_name),
'', '', '', ''
],
'layer_depth': [-1, -1, 512, 256, 256, 128],
'use_depthwise':
self._use_depthwise,
'use_explicit_padding':
self._use_explicit_padding,
}
if self._num_layers == 7:
feature_map_layout['from_layer'] += ['']
feature_map_layout['layer_depth'] += [64]
with tf.variable_scope(self._resnet_scope_name,
reuse=self._reuse_weights) as scope:
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
with (slim.arg_scope(self._conv_hyperparams_fn())
if self._override_base_feature_extractor_hyperparams else
context_manager.IdentityContextManager()):
_, image_features = self._resnet_base_fn(
inputs=ops.pad_to_multiple(preprocessed_inputs,
self._pad_to_multiple),
num_classes=None,
is_training=None,
global_pool=False,
output_stride=None,
store_non_strided_activations=True,
min_base_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
with slim.arg_scope(self._conv_hyperparams_fn()):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs,
preprocessed_second_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs.get_shape().assert_has_rank(4)
shape_assert = tf.Assert(
tf.logical_and(
tf.greater_equal(tf.shape(preprocessed_inputs)[1], 33),
tf.greater_equal(tf.shape(preprocessed_inputs)[2], 33)),
['image size must at least be 33 in both height and width.'])
feature_map_layout = {
'from_layer': ['Mixed_5d', 'Mixed_6e', 'Mixed_7c', '', '', ''],
'layer_depth': [-1, -1, -1, 512, 256, 128],
}
#print("preprocessed_second_inputs", preprocessed_second_inputs.get_shape())
with tf.control_dependencies([shape_assert]):
with slim.arg_scope(self._conv_hyperparams):
with tf.variable_scope('InceptionV3',
reuse=self._reuse_weights) as scope:
audio_features = self.extract_audio_features(
preprocessed_second_inputs)
_, image_features = inception_v3.inception_v3_base(
ops.pad_to_multiple(preprocessed_inputs,
self._pad_to_multiple),
# audio_features['fc5'],
# audio_features['flat'],
final_endpoint='Mixed_7c',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
#audio_features = self.extract_audio_features(preprocessed_second_inputs)
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
#audio_features=audio_features)
#return feature_maps.values()
return feature_maps.values(), audio_features['fc4']
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
feature_map_layout = {
'from_layer':
['block4', 'block7', 'block8', 'block9', 'block10', 'block11'],
'layer_depth': [-1, -1, -1, -1, -1, -1],
'use_explicit_padding':
self._use_explicit_padding,
'use_depthwise':
self._use_depthwise,
}
'''
with slim.arg_scope(self._conv_hyperparams_fn()):
with tf.variable_scope('InceptionV2',
reuse=self._reuse_weights) as scope:
_, image_features = inception_v2.inception_v2_base(
ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
final_endpoint='Mixed_5c',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
'''
with slim.arg_scope(vgg.vgg_arg_scope(weight_decay=0.0)):
with tf.variable_scope('vgg16',
reuse=self._reuse_weights) as var_scope:
_, image_features = vgg.vgg_16_ssd(preprocessed_inputs,
num_classes=3,
is_training=True,
dropout_keep_prob=0.9,
spatial_squeeze=False,
scope=var_scope,
fc_conv_padding='VALID',
global_pool=False,
end_point='pool5')
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
'''
feature_map_layout = {
'from_layer': ['layer_15/expansion_output', 'layer_19', '', '', '', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
'use_depthwise': self._use_depthwise,
'use_explicit_padding': self._use_explicit_padding,
}
'''
feature_map_layout = {
'from_layer': ['layer_5/expansion_output', 'layer_6/expansion_output', 'layer_7/expansion_output',
'layer_10/expansion_output', 'layer_15/expansion_output', 'layer_19'],
'layer_depth': [-1, -1, -1, -1, -1, -1],
'use_depthwise': self._use_depthwise,
'use_explicit_padding': self._use_explicit_padding,
}
with tf.variable_scope('MobilenetV2', reuse=self._reuse_weights) as scope:
with slim.arg_scope(resnet_utils.resnet_arg_scope()), \
slim.arg_scope(
[mobilenet.depth_multiplier], min_depth=self._min_depth):
with (slim.arg_scope(self._conv_hyperparams_fn())
if self._override_base_feature_extractor_hyperparams else
context_manager.IdentityContextManager()):
_, image_features = mobilenet_v2.mobilenet_base(
ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
final_endpoint='layer_19',
depth_multiplier=self._depth_multiplier,
use_explicit_padding=self._use_explicit_padding,
scope=scope)
with slim.arg_scope(self._conv_hyperparams_fn()):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def _extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
feature_map_layout = {
'from_layer':
['layer_15/expansion_output', 'layer_19', '', '', '', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
'use_depthwise': self._use_depthwise,
'use_explicit_padding': self._use_explicit_padding,
}
with tf.variable_scope('MobilenetV2',
reuse=self._reuse_weights) as scope:
with slim.arg_scope(
mobilenet_v2.training_scope(
is_training=(self._is_training and self._batch_norm_trainable),
bn_decay=0.9997)), \
slim.arg_scope(
[mobilenet.depth_multiplier], min_depth=self._min_depth):
# TODO(b/68150321): Enable fused batch norm once quantization
# supports it.
with slim.arg_scope([slim.batch_norm], fused=False):
_, image_features = mobilenet_v2.mobilenet_base(
ops.pad_to_multiple(preprocessed_inputs,
self._pad_to_multiple),
final_endpoint='layer_19',
depth_multiplier=self._depth_multiplier,
use_explicit_padding=self._use_explicit_padding,
scope=scope)
with slim.arg_scope(self._conv_hyperparams):
# TODO(b/68150321): Enable fused batch norm once quantization
# supports it.
with slim.arg_scope([slim.batch_norm], fused=False):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs.get_shape().assert_has_rank(4)
shape_assert = tf.Assert(
tf.logical_and(tf.greater_equal(tf.shape(preprocessed_inputs)[1], 33),
tf.greater_equal(tf.shape(preprocessed_inputs)[2], 33)),
['image size must at least be 33 in both height and width.'])
feature_map_layout = {
'from_layer': ['Cell_10', 'Cell_11', '', '',
'', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
}
with tf.control_dependencies([shape_assert]):
with slim.arg_scope(self._conv_hyperparams):
with slim.arg_scope([slim.batch_norm], fused=False):
with tf.variable_scope('NasNetLarge',
reuse=self._reuse_weights) as scope:
preprocessed_and_padded_inputs = ops.pad_to_multiple(preprocessed_inputs,
self._pad_to_multiple)
_, image_features = nasnet.build_nasnet_large(
preprocessed_and_padded_inputs,
num_classes=None,
final_endpoint='Cell_11')
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
batch_size = preprocessed_and_padded_inputs.get_shape().as_list()[0]
for feature_map_name in feature_maps:
print(feature_maps[feature_map_name])
feature_maps[feature_map_name].set_shape([batch_size] + feature_maps[feature_map_name].get_shape().as_list()[1:])
print(batch_size)
print("Inputs: ", preprocessed_and_padded_inputs)
print("Image features: ", image_features)
print("Feature maps:", feature_maps)
raw_input()
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
feature_map_layout = {
'from_layer':
['Conv2d_11_pointwise', 'Conv2d_13_pointwise', '', '', '',
''][:self._num_layers],
'layer_depth': [-1, -1, 512, 256, 256, 128][:self._num_layers],
'use_explicit_padding':
self._use_explicit_padding,
'use_depthwise':
self._use_depthwise,
}
with tf.variable_scope('MobilenetV1',
reuse=self._reuse_weights) as scope:
with slim.arg_scope(
mobilenet_v1.mobilenet_v1_arg_scope(
is_training=None, regularize_depthwise=True)):
with (slim.arg_scope(self._conv_hyperparams_fn())
if self._override_base_feature_extractor_hyperparams else
context_manager.IdentityContextManager()):
_, image_features = mobilenet_v1.mobilenet_v1_base(
ops.pad_to_multiple(preprocessed_inputs,
self._pad_to_multiple),
final_endpoint='Conv2d_13_pointwise',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
use_explicit_padding=self._use_explicit_padding,
scope=scope)
with slim.arg_scope(self._conv_hyperparams_fn()):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
feature_map_layout = {
'from_layer': ['layer_15/expansion_output', 'layer_19', '', '', '', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
'use_depthwise': self._use_depthwise,
'use_explicit_padding': self._use_explicit_padding,
}
with tf.variable_scope('MobilenetV2', reuse=self._reuse_weights) as scope:
with slim.arg_scope(
mobilenet_v2.training_scope(is_training=None, bn_decay=0.9997)), \
slim.arg_scope(
[mobilenet.depth_multiplier], min_depth=self._min_depth):
with (slim.arg_scope(self._conv_hyperparams_fn())
if self._override_base_feature_extractor_hyperparams else
context_manager.IdentityContextManager()):
# TODO(b/68150321): Enable fused batch norm once quantization
# supports it.
with slim.arg_scope([slim.batch_norm], fused=False):
_, image_features = mobilenet_v2.mobilenet_base(
ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
final_endpoint='layer_19',
depth_multiplier=self._depth_multiplier,
use_explicit_padding=self._use_explicit_padding,
scope=scope)
with slim.arg_scope(self._conv_hyperparams_fn()):
# TODO(b/68150321): Enable fused batch norm once quantization
# supports it.
with slim.arg_scope([slim.batch_norm], fused=False):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def _extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
feature_map_layout = {
'from_layer': ['Conv2d_11_pointwise', 'Conv2d_13_pointwise', '', '',
'', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
'use_explicit_padding': self._use_explicit_padding,
'use_depthwise': self._use_depthwise,
}
with tf.variable_scope('MobilenetV1',
reuse=self._reuse_weights) as scope:
with slim.arg_scope(
mobilenet_v1.mobilenet_v1_arg_scope(
is_training=(self._batch_norm_trainable and self._is_training))):
# TODO(skligys): Enable fused batch norm once quantization supports it.
with slim.arg_scope([slim.batch_norm], fused=False):
_, image_features = mobilenet_v1.mobilenet_v1_base(
ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
final_endpoint='Conv2d_13_pointwise',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
use_explicit_padding=self._use_explicit_padding,
scope=scope)
with slim.arg_scope(self._conv_hyperparams):
# TODO(skligys): Enable fused batch norm once quantization supports it.
with slim.arg_scope([slim.batch_norm], fused=False):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
feature_map_layout = {
'from_layer': ['Conv2d_11_pointwise', 'Conv2d_13_pointwise', '', '',
'', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
'use_explicit_padding': self._use_explicit_padding,
'use_depthwise': self._use_depthwise,
}
with tf.variable_scope('MobilenetV1',
reuse=self._reuse_weights) as scope:
with slim.arg_scope(
mobilenet_v1.mobilenet_v1_arg_scope(
is_training=None, regularize_depthwise=True)):
with (slim.arg_scope(self._conv_hyperparams_fn())
if self._override_base_feature_extractor_hyperparams
else context_manager.IdentityContextManager()):
_, image_features = mobilenet_v1.mobilenet_v1_base(
ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
final_endpoint='Conv2d_13_pointwise',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
use_explicit_padding=self._use_explicit_padding,
scope=scope)
with slim.arg_scope(self._conv_hyperparams_fn()):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
feature_map_layout = {
'from_layer': ['Cell_7', 'Cell_11', '', '', '',
''][:self._num_layers],
'layer_depth': [-1, -1, 512, 256, 256, 128][:self._num_layers],
'use_explicit_padding': self._use_explicit_padding,
'use_depthwise': self._use_depthwise,
}
with slim.arg_scope(
pnasnet_large_arg_scope_for_detection(
is_batch_norm_training=self._is_training)):
with slim.arg_scope(
[slim.conv2d, slim.batch_norm, slim.separable_conv2d],
reuse=self._reuse_weights):
with (slim.arg_scope(self._conv_hyperparams_fn())
if self._override_base_feature_extractor_hyperparams else
context_manager.IdentityContextManager()):
_, image_features = pnasnet.build_pnasnet_large(
ops.pad_to_multiple(preprocessed_inputs,
self._pad_to_multiple),
num_classes=None,
is_training=self._is_training,
final_endpoint='Cell_11')
with tf.compat.v1.variable_scope('SSD_feature_maps',
reuse=self._reuse_weights):
with slim.arg_scope(self._conv_hyperparams_fn()):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
padded_inputs = ops.pad_to_multiple(preprocessed_inputs,
self._pad_to_multiple)
feature_map_layout = {
'from_layer': ['C4', 'C5', '', '', '', ''],
# Do not specify the layer depths (number of filters) for C4 and C5, as
# their values are determined based on the backbone.
'layer_depth': [-1, -1, 512, 256, 256, 128],
'use_depthwise': self._use_depthwise,
'use_explicit_padding': self._use_explicit_padding,
}
with tf.variable_scope(self._scope_name, reuse=self._reuse_weights):
with slim.arg_scope([slim.batch_norm],
is_training=self._is_training,
epsilon=0.01,
decay=0.99,
center=True,
scale=True):
endpoints = self._backbone_fn(
padded_inputs, multiplier=self._depth_multiplier)
image_features = {'C4': endpoints['C4'], 'C5': endpoints['C5']}
with slim.arg_scope(self._conv_hyperparams_fn()):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return list(feature_maps.values())
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs.get_shape().assert_has_rank(4)
shape_assert = tf.Assert(
tf.logical_and(
tf.greater_equal(tf.shape(preprocessed_inputs)[1], 33),
tf.greater_equal(tf.shape(preprocessed_inputs)[2], 33)),
['image size must at least be 33 in both height and width.'])
feature_map_layout = {
'from_layer':
['Conv2d_11_pointwise', 'Conv2d_13_pointwise', '', '', '', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
#'use_depthwise':True
}
with tf.control_dependencies([shape_assert]):
with slim.arg_scope(self._conv_hyperparams):
with slim.arg_scope([slim.batch_norm], fused=False):
with tf.variable_scope('MobilenetV1',
reuse=self._reuse_weights) as scope:
_, image_features = mobilenet_v1.mobilenet_v1_base(
ops.pad_to_multiple(preprocessed_inputs,
self._pad_to_multiple),
final_endpoint='Conv2d_13_pointwise',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
feature_map_layout = {
'from_layer':
['Conv2d_11_pointwise', 'Conv2d_13_pointwise', '', '', '', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
'use_explicit_padding':
self._use_explicit_padding,
'use_depthwise':
self._use_depthwise,
}
with slim.arg_scope(self._conv_hyperparams):
# TODO: Enable fused batch norm once quantization supports it.
with slim.arg_scope([slim.batch_norm], fused=False):
with tf.variable_scope('MobilenetV1',
reuse=self._reuse_weights) as scope:
_, image_features = mobilenet_v1.mobilenet_v1_base(
ops.pad_to_multiple(preprocessed_inputs,
self._pad_to_multiple),
final_endpoint='Conv2d_13_pointwise',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs.get_shape().assert_has_rank(4)
shape_assert = tf.Assert(
tf.logical_and(
tf.equal(tf.shape(preprocessed_inputs)[1], 256),
tf.equal(tf.shape(preprocessed_inputs)[2], 256)),
['image size must be 256 in both height and width.'])
feature_map_layout = {
'from_layer': [
'Conv2d_11_pointwise', 'Conv2d_13_pointwise', '', '', ''
],
'layer_depth': [-1, -1, 512, 256, 256],
'conv_kernel_size': [-1, -1, 3, 3, 2],
}
with tf.control_dependencies([shape_assert]):
with slim.arg_scope(self._conv_hyperparams):
with tf.variable_scope('MobilenetV1',
reuse=self._reuse_weights) as scope:
_, image_features = mobilenet_v1.mobilenet_v1_base(
ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
final_endpoint='Conv2d_13_pointwise',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
feature_map_layout = {
'from_layer': ['Cell_7', 'Cell_11', '', '', '', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
'use_explicit_padding': self._use_explicit_padding,
'use_depthwise': self._use_depthwise,
}
with slim.arg_scope(
pnasnet_large_arg_scope_for_detection(
is_batch_norm_training=self._is_training)):
with slim.arg_scope([slim.conv2d, slim.batch_norm, slim.separable_conv2d],
reuse=self._reuse_weights):
with (slim.arg_scope(self._conv_hyperparams_fn())
if self._override_base_feature_extractor_hyperparams else
context_manager.IdentityContextManager()):
_, image_features = pnasnet.build_pnasnet_large(
ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
num_classes=None,
is_training=self._is_training,
final_endpoint='Cell_11')
with tf.variable_scope('SSD_feature_maps', reuse=self._reuse_weights):
with slim.arg_scope(self._conv_hyperparams_fn()):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
feature_map_layout = {
'from_layer': ['Mixed_4c', 'Mixed_5c', '', '', '',
''][:self._num_layers],
'layer_depth': [-1, -1, 512, 256, 256, 128][:self._num_layers],
'use_explicit_padding':
self._use_explicit_padding,
'use_depthwise':
self._use_depthwise,
}
with slim.arg_scope(self._conv_hyperparams_fn()):
with tf.variable_scope('InceptionV2',
reuse=self._reuse_weights) as scope:
_, image_features = inception_v2.inception_v2_base(
ops.pad_to_multiple(preprocessed_inputs,
self._pad_to_multiple),
final_endpoint='Mixed_5c',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs.get_shape().assert_has_rank(4)
shape_assert = tf.Assert(
tf.logical_and(tf.greater_equal(tf.shape(preprocessed_inputs)[1], 33),
tf.greater_equal(tf.shape(preprocessed_inputs)[2], 33)),
['image size must at least be 33 in both height and width.'])
feature_map_layout = {
'from_layer': ['Mixed_4c', 'Mixed_5c', '', '', '', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
}
with tf.control_dependencies([shape_assert]):
with slim.arg_scope(self._conv_hyperparams):
with tf.variable_scope('InceptionV2',
reuse=self._reuse_weights) as scope:
_, image_features = inception_v2.inception_v2_base(
preprocessed_inputs,
final_endpoint='Mixed_5c',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs.get_shape().assert_has_rank(4)
shape_assert = tf.Assert(
tf.logical_and(
tf.greater_equal(tf.shape(preprocessed_inputs)[1], 33),
tf.greater_equal(tf.shape(preprocessed_inputs)[2], 33)),
['image size must at least be 33 in both height and width.'])
feature_map_layout = {
'from_layer': ['conv4', '', '', '', '', '', ''],
'layer_depth': [-1, 1024, 1024, 512, 256, 256, 256],
}
with tf.control_dependencies([shape_assert]):
with slim.arg_scope(self._conv_hyperparams):
with tf.variable_scope('vgg_16',
reuse=self._reuse_weights) as scope:
net, image_features = vgg.vgg_16_base(
preprocessed_inputs,
final_endpoint='pool5',
trainable=False,
scope=scope)
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
feature_map_layout = {
'from_layer': ['Mixed_4c', 'Mixed_5c', '', '', '', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
'use_explicit_padding': self._use_explicit_padding,
'use_depthwise': self._use_depthwise,
}
with slim.arg_scope(self._conv_hyperparams_fn()):
with tf.variable_scope('InceptionV2',
reuse=self._reuse_weights) as scope:
_, image_features = inception_v2.inception_v2_base(
ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
final_endpoint='Mixed_5c',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs, state_saver=None,
state_name='lstm_state', unroll_length=10, scope=None):
"""Extract features from preprocessed inputs.
The features include the base network features, lstm features and SSD
features, organized in the following name scope:
<scope>/MobilenetV2_1/...
<scope>/MobilenetV2_2/...
<scope>/LSTM/...
<scope>/FeatureMap/...
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of consecutive frames from video clips.
state_saver: A state saver object with methods `state` and `save_state`.
state_name: Python string, the name to use with the state_saver.
unroll_length: number of steps to unroll the lstm.
scope: Scope for the base network of the feature extractor.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
Raises:
ValueError: if interleave_method not recognized or large and small base
network output feature maps of different sizes.
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
preprocessed_inputs = ops.pad_to_multiple(
preprocessed_inputs, self._pad_to_multiple)
batch_size = preprocessed_inputs.shape[0].value / unroll_length
batch_axis = 0
nets = []
# Batch processing of mobilenet features.
with slim.arg_scope(mobilenet_v2.training_scope(
is_training=self._is_training,
bn_decay=0.9997)), \
slim.arg_scope([mobilenet.depth_multiplier],
min_depth=self._min_depth, divisible_by=8):
# Big model.
net, _ = self.extract_base_features_large(preprocessed_inputs)
nets.append(net)
large_base_feature_shape = net.shape
# Small models
net, _ = self.extract_base_features_small(preprocessed_inputs)
nets.append(net)
small_base_feature_shape = net.shape
if not (large_base_feature_shape[1] == small_base_feature_shape[1] and
large_base_feature_shape[2] == small_base_feature_shape[2]):
raise ValueError('Large and Small base network feature map dimension '
'not equal!')
with slim.arg_scope(self._conv_hyperparams_fn()):
with tf.variable_scope('LSTM', reuse=self._reuse_weights):
output_size = (large_base_feature_shape[1], large_base_feature_shape[2])
lstm_cell, init_state, step = self.create_lstm_cell(
batch_size, output_size, state_saver, state_name)
nets_seq = [
tf.split(net, unroll_length, axis=batch_axis) for net in nets
]
net_seq, states_out = rnn_decoder.multi_input_rnn_decoder(
nets_seq,
init_state,
lstm_cell,
step,
selection_strategy=self._interleave_method,
is_training=self._is_training,
is_quantized=self._is_quantized,
pre_bottleneck=self._pre_bottleneck,
flatten_state=self._flatten_state,
scope=None)
self._states_out = states_out
batcher_ops = None
if state_saver is not None:
self._step = state_saver.state(state_name + '_step')
batcher_ops = [
state_saver.save_state(state_name + '_c', states_out[-1][0]),
state_saver.save_state(state_name + '_h', states_out[-1][1]),
state_saver.save_state(state_name + '_step', self._step + 1)]
image_features = {}
with tf_ops.control_dependencies(batcher_ops):
image_features['layer_19'] = tf.concat(net_seq, 0)
# SSD layers.
with tf.variable_scope('FeatureMap'):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=self._feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features,
pool_residual=True)
return feature_maps.values()
def extract_features(self,
preprocessed_inputs,
state_saver=None,
state_name='lstm_state',
unroll_length=5,
scope=None):
"""Extracts features from preprocessed inputs.
The features include the base network features, lstm features and SSD
features, organized in the following name scope:
<parent scope>/MobilenetV1/...
<parent scope>/LSTM/...
<parent scope>/FeatureMaps/...
Args:
preprocessed_inputs: A [batch, height, width, channels] float tensor
representing a batch of consecutive frames from video clips.
state_saver: A state saver object with methods `state` and `save_state`.
state_name: A python string for the name to use with the state_saver.
unroll_length: The number of steps to unroll the lstm.
scope: The scope for the base network of the feature extractor.
Returns:
A list of tensors where the ith tensor has shape [batch, height_i,
width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
with slim.arg_scope(
mobilenet_v1.mobilenet_v1_arg_scope(
is_training=self._is_training)):
with (slim.arg_scope(self._conv_hyperparams_fn())
if self._override_base_feature_extractor_hyperparams else
context_manager.IdentityContextManager()):
with slim.arg_scope([slim.batch_norm], fused=False):
# Base network.
with tf.variable_scope(scope,
self._base_network_scope,
reuse=self._reuse_weights) as scope:
net, image_features = mobilenet_v1.mobilenet_v1_base(
ops.pad_to_multiple(preprocessed_inputs,
self._pad_to_multiple),
final_endpoint='Conv2d_13_pointwise',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
with slim.arg_scope(self._conv_hyperparams_fn()):
with slim.arg_scope([slim.batch_norm],
fused=False,
is_training=self._is_training):
# ConvLSTM layers.
with tf.variable_scope(
'LSTM', reuse=self._reuse_weights) as lstm_scope:
lstm_cell = lstm_cells.BottleneckConvLSTMCell(
filter_size=(3, 3),
output_size=(net.shape[1].value, net.shape[2].value),
num_units=max(self._min_depth, self._lstm_state_depth),
activation=tf.nn.relu6,
visualize_gates=True)
net_seq = list(tf.split(net, unroll_length))
if state_saver is None:
init_state = lstm_cell.init_state(
state_name, net.shape[0].value / unroll_length,
tf.float32)
else:
c = state_saver.state('%s_c' % state_name)
h = state_saver.state('%s_h' % state_name)
init_state = (c, h)
# Identities added for inputing state tensors externally.
c_ident = tf.identity(init_state[0],
name='lstm_state_in_c')
h_ident = tf.identity(init_state[1],
name='lstm_state_in_h')
init_state = (c_ident, h_ident)
net_seq, states_out = rnn_decoder.rnn_decoder(
net_seq, init_state, lstm_cell, scope=lstm_scope)
batcher_ops = None
self._states_out = states_out
if state_saver is not None:
self._step = state_saver.state('%s_step' % state_name)
batcher_ops = [
state_saver.save_state('%s_c' % state_name,
states_out[-1][0]),
state_saver.save_state('%s_h' % state_name,
states_out[-1][1]),
state_saver.save_state('%s_step' % state_name,
self._step - 1)
]
with tf_ops.control_dependencies(batcher_ops):
image_features['Conv2d_13_pointwise_lstm'] = tf.concat(
net_seq, 0)
# Identities added for reading output states, to be reused externally.
tf.identity(states_out[-1][0], name='lstm_state_out_c')
tf.identity(states_out[-1][1], name='lstm_state_out_h')
# SSD layers.
with tf.variable_scope('FeatureMaps',
reuse=self._reuse_weights):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=self._feature_map_layout,
depth_multiplier=(self._depth_multiplier),
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
Raises:
ValueError: if image height or width are not 256 pixels.
"""
image_shape = preprocessed_inputs.get_shape()
image_shape.assert_has_rank(4)
image_height = image_shape[1].value
image_width = image_shape[2].value
if image_height is None or image_width is None:
shape_assert = tf.Assert(
tf.logical_and(tf.equal(tf.shape(preprocessed_inputs)[1], 256),
tf.equal(tf.shape(preprocessed_inputs)[2],
256)),
['image size must be 256 in both height and width.'])
with tf.control_dependencies([shape_assert]):
preprocessed_inputs = tf.identity(preprocessed_inputs)
elif image_height != 256 or image_width != 256:
raise ValueError(
'image size must be = 256 in both height and width;'
' image dim = %d,%d' % (image_height, image_width))
feature_map_layout = {
'from_layer':
['Conv2d_11_pointwise', 'Conv2d_13_pointwise', '', '', ''],
'layer_depth': [-1, -1, 512, 256, 256],
'conv_kernel_size': [-1, -1, 3, 3, 2],
'use_explicit_padding':
self._use_explicit_padding,
'use_depthwise':
self._use_depthwise,
}
with tf.variable_scope('MobilenetV1',
reuse=self._reuse_weights) as scope:
with slim.arg_scope(
mobilenet_v1.mobilenet_v1_arg_scope(is_training=None)):
with (slim.arg_scope(self._conv_hyperparams_fn())
if self._override_base_feature_extractor_hyperparams else
context_manager.IdentityContextManager()):
_, image_features = mobilenet_v1.mobilenet_v1_base(
ops.pad_to_multiple(preprocessed_inputs,
self._pad_to_multiple),
final_endpoint='Conv2d_13_pointwise',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
use_explicit_padding=self._use_explicit_padding,
scope=scope)
with slim.arg_scope(self._conv_hyperparams_fn()):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self,
preprocessed_inputs,
state_saver=None,
state_name='lstm_state',
unroll_length=5,
scope=None):
"""Extracts features from preprocessed inputs.
The features include the base network features, lstm features and SSD
features, organized in the following name scope:
<parent scope>/MobilenetV1/...
<parent scope>/LSTM/...
<parent scope>/FeatureMaps/...
Args:
preprocessed_inputs: A [batch, height, width, channels] float tensor
representing a batch of consecutive frames from video clips.
state_saver: A state saver object with methods `state` and `save_state`.
state_name: A python string for the name to use with the state_saver.
unroll_length: The number of steps to unroll the lstm.
scope: The scope for the base network of the feature extractor.
Returns:
A list of tensors where the ith tensor has shape [batch, height_i,
width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
with slim.arg_scope(
mobilenet_v1.mobilenet_v1_arg_scope(is_training=self._is_training)):
with (slim.arg_scope(self._conv_hyperparams_fn())
if self._override_base_feature_extractor_hyperparams else
context_manager.IdentityContextManager()):
with slim.arg_scope([slim.batch_norm], fused=False):
# Base network.
with tf.variable_scope(
scope, self._base_network_scope,
reuse=self._reuse_weights) as scope:
net, image_features = mobilenet_v1.mobilenet_v1_base(
ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
final_endpoint='Conv2d_13_pointwise',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
with slim.arg_scope(self._conv_hyperparams_fn()):
with slim.arg_scope(
[slim.batch_norm], fused=False, is_training=self._is_training):
# ConvLSTM layers.
with tf.variable_scope('LSTM', reuse=self._reuse_weights) as lstm_scope:
lstm_cell = lstm_cells.BottleneckConvLSTMCell(
filter_size=(3, 3),
output_size=(net.shape[1].value, net.shape[2].value),
num_units=max(self._min_depth, self._lstm_state_depth),
activation=tf.nn.relu6,
visualize_gates=True)
net_seq = list(tf.split(net, unroll_length))
if state_saver is None:
init_state = lstm_cell.init_state(
state_name, net.shape[0].value / unroll_length, tf.float32)
else:
c = state_saver.state('%s_c' % state_name)
h = state_saver.state('%s_h' % state_name)
init_state = (c, h)
# Identities added for inputing state tensors externally.
c_ident = tf.identity(init_state[0], name='lstm_state_in_c')
h_ident = tf.identity(init_state[1], name='lstm_state_in_h')
init_state = (c_ident, h_ident)
net_seq, states_out = rnn_decoder.rnn_decoder(
net_seq, init_state, lstm_cell, scope=lstm_scope)
batcher_ops = None
self._states_out = states_out
if state_saver is not None:
self._step = state_saver.state('%s_step' % state_name)
batcher_ops = [
state_saver.save_state('%s_c' % state_name, states_out[-1][0]),
state_saver.save_state('%s_h' % state_name, states_out[-1][1]),
state_saver.save_state('%s_step' % state_name, self._step - 1)
]
with tf_ops.control_dependencies(batcher_ops):
image_features['Conv2d_13_pointwise_lstm'] = tf.concat(net_seq, 0)
# Identities added for reading output states, to be reused externally.
tf.identity(states_out[-1][0], name='lstm_state_out_c')
tf.identity(states_out[-1][1], name='lstm_state_out_h')
# SSD layers.
with tf.variable_scope('FeatureMaps', reuse=self._reuse_weights):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=self._feature_map_layout,
depth_multiplier=(self._depth_multiplier),
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
feature_map_layout = {
'from_layer': ['vgg_16/conv4_3_norm', 'vgg_16/fc7', 'vgg_16/conv6_2', 'vgg_16/conv7_2', 'vgg_16/conv8_2', 'vgg_16/conv9_2'],
'layer_depth': [-1, -1, -1, -1, -1, -1],
'use_explicit_padding': self._use_explicit_padding,
'use_depthwise': self._use_depthwise,
}
net,image_features = vgg.vgg_16(ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
final_endpoint='pool5',spatial_squeeze=False)
#double check scale filler
image_features['vgg_16/conv4_3_norm']=custom_layers.l2_normalization(image_features['vgg_16/conv4/conv4_3'],scaling=True,scope='vgg_16/conv4_3_norm')
with slim.arg_scope(self._conv_hyperparams):
with tf.variable_scope('vgg_16', reuse=self._reuse_weights) as scope:
# In [5]: net
# Out[5]: <tf.Tensor 'vgg_16/pool5/MaxPool:0' shape=(32, 18, 18, 512) dtype=float32>
# In [6]: end_points
# Out[6]:
# OrderedDict([('vgg_16/conv1/conv1_1',
# <tf.Tensor 'vgg_16/conv1/conv1_1/Relu:0' shape=(32, 300, 300, 64) dtype=float32>),
# ('vgg_16/conv1/conv1_2',
# <tf.Tensor 'vgg_16/conv1/conv1_2/Relu:0' shape=(32, 300, 300, 64) dtype=float32>),
# ('vgg_16/pool1',
# <tf.Tensor 'vgg_16/pool1/MaxPool:0' shape=(32, 150, 150, 64) dtype=float32>),
# ('vgg_16/conv2/conv2_1',
# <tf.Tensor 'vgg_16/conv2/conv2_1/Relu:0' shape=(32, 150, 150, 128) dtype=float32>),
# ('vgg_16/conv2/conv2_2',
# <tf.Tensor 'vgg_16/conv2/conv2_2/Relu:0' shape=(32, 150, 150, 128) dtype=float32>),
# ('vgg_16/pool2',
# <tf.Tensor 'vgg_16/pool2/MaxPool:0' shape=(32, 75, 75, 128) dtype=float32>),
# ('vgg_16/conv3/conv3_1',
# <tf.Tensor 'vgg_16/conv3/conv3_1/Relu:0' shape=(32, 75, 75, 256) dtype=float32>),
# ('vgg_16/conv3/conv3_2',
# <tf.Tensor 'vgg_16/conv3/conv3_2/Relu:0' shape=(32, 75, 75, 256) dtype=float32>),
# ('vgg_16/conv3/conv3_3',
# <tf.Tensor 'vgg_16/conv3/conv3_3/Relu:0' shape=(32, 75, 75, 256) dtype=float32>),
# ('vgg_16/pool3',
# <tf.Tensor 'vgg_16/pool3/MaxPool:0' shape=(32, 37, 37, 256) dtype=float32>),
# ('vgg_16/conv4/conv4_1',
# <tf.Tensor 'vgg_16/conv4/conv4_1/Relu:0' shape=(32, 37, 37, 512) dtype=float32>),
# ('vgg_16/conv4/conv4_2',
# <tf.Tensor 'vgg_16/conv4/conv4_2/Relu:0' shape=(32, 37, 37, 512) dtype=float32>),
# ('vgg_16/conv4/conv4_3',
# <tf.Tensor 'vgg_16/conv4/conv4_3/Relu:0' shape=(32, 37, 37, 512) dtype=float32>),
# ('vgg_16/pool4',
# <tf.Tensor 'vgg_16/pool4/MaxPool:0' shape=(32, 18, 18, 512) dtype=float32>),
# ('vgg_16/conv5/conv5_1',
# <tf.Tensor 'vgg_16/conv5/conv5_1/Relu:0' shape=(32, 18, 18, 512) dtype=float32>),
# ('vgg_16/conv5/conv5_2',
# <tf.Tensor 'vgg_16/conv5/conv5_2/Relu:0' shape=(32, 18, 18, 512) dtype=float32>),
# ('vgg_16/conv5/conv5_3',
# <tf.Tensor 'vgg_16/conv5/conv5_3/Relu:0' shape=(32, 18, 18, 512) dtype=float32>),
# ('vgg_16/pool5',
# <tf.Tensor 'vgg_16/pool5/MaxPool:0' shape=(32, 18, 18, 512) dtype=float32>)])
end_points_collection = scope.original_name_scope + '_end_points'
with slim.arg_scope([slim.conv2d, slim.fully_connected, slim.max_pool2d],
outputs_collections=end_points_collection):
net = slim.convolution(net, 1024, [3, 3], padding='SAME', rate=6, scope='fc6')
# def convolution(inputs,
# num_outputs,
# kernel_size,
# stride=1,
# padding='SAME',
# data_format=None,
# rate=1,
# activation_fn=nn.relu,
# normalizer_fn=None,
# normalizer_params=None,
# weights_initializer=initializers.xavier_initializer(),
# weights_regularizer=None,
# biases_initializer=init_ops.zeros_initializer(),
# biases_regularizer=None,
# reuse=None,
# variables_collections=None,
# outputs_collections=None,
# trainable=True,
# scope=None):
#fc6 is dilated conv
# layer {
# name: "fc6"
# type: "Convolution"
# bottom: "pool5"
# top: "fc6"
# param {
# lr_mult: 1.0
# decay_mult: 1.0
# }
# param {
# lr_mult: 2.0
# decay_mult: 0.0
# }
# convolution_param {
# num_output: 1024
# pad: 6
# kernel_size: 3
# weight_filler {
# type: "xavier"
# }
# bias_filler {
# type: "constant"
# value: 0.0
# }
# dilation: 6
# }
# }
# layer {
# name: "relu6"
# type: "ReLU"
# bottom: "fc6"
# top: "fc6"
# }
#fc7 is 1*1 conv
# layer {
# name: "fc7"
# type: "Convolution"
# bottom: "fc6"
# top: "fc7"
# param {
# lr_mult: 1.0
# decay_mult: 1.0
# }
# param {
# lr_mult: 2.0
# decay_mult: 0.0
# }
# convolution_param {
# num_output: 1024
# kernel_size: 1
# weight_filler {
# type: "xavier"
# }
# bias_filler {
# type: "constant"
# value: 0.0
# }
# }
# }
# layer {
# name: "relu7"
# type: "ReLU"
# bottom: "fc7"
# top: "fc7"
# }
net = slim.conv2d(net,1024,[1,1],padding='SAME',scope='fc7')
net = slim.conv2d(net,256, [1,1],padding='SAME',scope='conv6_1')
net = slim.conv2d(net,512, [3,3],padding='SAME',stride=2,scope='conv6_2')
net = slim.conv2d(net,128, [1,1],padding='SAME',scope='conv7_1')
net = slim.conv2d(net,256, [3,3],padding='SAME',stride=2,scope='conv7_2')
net = slim.conv2d(net,128, [1,1],padding='VALID',scope='conv8_1')
net = slim.conv2d(net,256, [3,3],padding='VALID',stride=1,scope='conv8_2')
net = slim.conv2d(net,128, [1,1],padding='VALID',scope='conv9_1')
net = slim.conv2d(net,256, [3,3],padding='VALID',stride=1,scope='conv9_2')
_image_features_new = slim.utils.convert_collection_to_dict(end_points_collection)
for k,v in _image_features_new.iteritems():
image_features[k] = v
# _, image_features = inception_v3.inception_v3_base(
# ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
# final_endpoint='Mixed_7c',
# min_depth=self._min_depth,
# depth_multiplier=self._depth_multiplier,
# scope=scope)
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
depth = lambda d: max(int(d * self._depth_multiplier), self._min_depth)
trunc_normal = lambda stddev: tf.truncated_normal_initializer(
0.0, stddev)
#add convolution autoencoder
encoder_1_conv = slim.conv2d(preprocessed_inputs,
depth(64), [3, 3],
weights_initializer=trunc_normal(0.09),
scope='encoder_c1_conv')
#encoder_1_pool = slim.max_pool2d(encoder_1_conv, [2, 2], stride=2,
# scope='encoder_c1_pool')
#encoder_1_dropout1 = slim.dropout(encoder_1_pool, 0.7, scope='encoder_c1_dropout1')
encoder_2_conv = slim.conv2d(encoder_1_conv,
depth(128), [5, 5],
weights_initializer=trunc_normal(0.09),
scope='encoder_c2_conv')
encoder_3_conv = slim.conv2d(encoder_2_conv,
depth(128), [5, 5],
weights_initializer=trunc_normal(0.09),
scope='encoder_c3_conv')
#decoder
decoder_3_deconv = slim.conv2d_transpose(
encoder_3_conv,
depth(128), [5, 5],
weights_initializer=trunc_normal(0.09),
scope='decoder_c3_deconv')
decoder_2_deconv = slim.conv2d_transpose(
decoder_3_deconv,
depth(128), [5, 5],
weights_initializer=trunc_normal(0.09),
scope='decoder_c2_deconv')
decoder_1_deconv = slim.conv2d_transpose(
decoder_2_deconv,
depth(1), [3, 3],
weights_initializer=trunc_normal(0.09),
scope='decoder_c1_deconv')
feature_map_layout = {
'from_layer': ['Mixed_4c', 'Mixed_5c', '', '', '', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
'use_explicit_padding': self._use_explicit_padding,
'use_depthwise': self._use_depthwise,
}
with slim.arg_scope(self._conv_hyperparams_fn()):
with tf.variable_scope('InceptionV2',
reuse=self._reuse_weights) as scope:
_, image_features = inception_v2.inception_v2_base(
ops.pad_to_multiple(decoder_1_deconv,
self._pad_to_multiple),
final_endpoint='Mixed_5c',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
def extract_features(
self, preprocessed_inputs
): #this will extract features from iamge w.r.t mobilenet archtecture
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs.get_shape().assert_has_rank(4)
shape_assert = tf.Assert(
tf.logical_and(
tf.greater_equal(tf.shape(preprocessed_inputs)[1], 33),
tf.greater_equal(tf.shape(preprocessed_inputs)[2], 33)),
['image size must at least be 33 in both height and width.'])
feature_map_layout = {
'from_layer': [
'Conv2d_11_pointwise',
'Conv2d_13_pointwise',
'',
'', #we first extract 2 layers from mobilenet
'',
''
],
'layer_depth': [
-1, -1, 512, 256, 256, 128
], #for first two things it's -1 means we directly take the depth as in the feature maps
}
with tf.control_dependencies([shape_assert]):
#with following we apply all the hyperparams in the scrip by keeping arg scope free
with slim.arg_scope(
self._conv_hyperparams
): #arg score - Here the convolutional hyper params are for feature extractor we create ot
with tf.variable_scope('MobilenetV1',
reuse=self._reuse_weights) as scope:
_, image_features = mobilenet_v1.mobilenet_v1_base( #getting the feature extracted from mobilnet in the slim
preprocessed_inputs,
final_endpoint=
'Conv2d_13_pointwise', #this is extracting the features
min_depth=self.
_min_depth, #our min deph is 16 , It's like our depth of the feature extator
depth_multiplier=self.
_depth_multiplier, #there is 1 we take all the layers in depth demension
scope=scope
) #this is a dicrionalt with names of the feature maps and feature maps
#the following function can extract the features from above feature maps , also it can create new one's too acording to the output stride thing which we are not using Alos we give a featue map lay_out what should be there , and this also can create addicitonal feature maps
feature_maps = feature_map_generators.multi_resolution_feature_maps( #This is for generating feature maps
feature_map_layout=
feature_map_layout, #wanted feature maps extracted from above model maps and create new maps for empty things
depth_multiplier=self.
_depth_multiplier, #depth multi-plier
min_depth=self._min_depth, #this is 16
insert_1x1_conv=True, #
image_features=image_features) #feature dictionary
return feature_maps.values() #list of 6 feature maps for the ssd
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
preprocessed_inputs.get_shape().assert_has_rank(4)
shape_assert = tf.Assert(
tf.logical_and(tf.greater_equal(tf.shape(preprocessed_inputs)[1], 33),
tf.greater_equal(tf.shape(preprocessed_inputs)[2], 33)),
['image size must at least be 33 in both height and width.'])
feature_map_layout = {
'from_layer': ['east_conv1_3x3'],
'layer_depth': [-1],
}
with tf.control_dependencies([shape_assert]):
with slim.arg_scope(self._conv_hyperparams):
with tf.variable_scope('MobilenetV1',
reuse=self._reuse_weights) as scope:
_, image_features = mobilenet_v1.mobilenet_v1_base(
preprocessed_inputs,
final_endpoint='Conv2d_13_pointwise',
min_depth=self._min_depth,
depth_multiplier=self._depth_multiplier,
scope=scope)
"""
by chenx
"""
east_conv_1 = image_features['Conv2d_3_pointwise']
east_conv_2 = image_features['Conv2d_5_pointwise']
east_conv_3 = image_features['Conv2d_11_pointwise']
east_conv_4 = image_features['Conv2d_13_pointwise']
east_deconv4 = slim.conv2d_transpose(east_conv_4, 512, [4, 4], 2, \
padding='SAME', scope='east_deconv4')
east_conv4_concat = tf.concat([east_conv_4, east_deconv4], axis=3)
east_conv4_1x1 = slim.conv2d(east_conv4_concat, 256, [1,1],
stride=1,
normalizer_fn=slim.batch_norm,
scope='east_conv4_1x1')
east_conv4_3x3 = slim.conv2d(east_conv4_1x1, 256, [3,3],
stride=1,
normalizer_fn=slim.batch_norm,
scope='east_conv4_3x3')
image_features['east_conv4_3x3'] = east_conv4_3x3
east_deconv3 = slim.conv2d_transpose(east_conv4_3x3, 256, [4, 4], 2, \
padding='SAME', scope='east_deconv3')
east_conv3_concat = tf.concat([east_conv_3, east_deconv3], axis=3)
east_conv3_1x1 = slim.conv2d(east_conv4_concat, 128, [1,1],
stride=1,
normalizer_fn=slim.batch_norm,
scope='east_conv3_1x1')
east_conv3_3x3 = slim.conv2d(east_conv4_1x1, 128, [3,3],
stride=1,
normalizer_fn=slim.batch_norm,
scope='east_conv3_3x3')
image_features['east_conv3_3x3'] = east_conv3_3x3
east_deconv2 = slim.conv2d_transpose(east_conv3_3x3, 128, [4, 4], 2, \
padding='SAME', scope='east_deconv2')
east_conv2_concat = tf.concat([east_conv_2, east_deconv3], axis=3)
east_conv2_1x1 = slim.conv2d(east_conv2_concat, 64, [1,1],
stride=1,
normalizer_fn=slim.batch_norm,
scope='east_conv2_1x1')
east_conv2_3x3 = slim.conv2d(east_conv2_1x1, 64, [3,3],
stride=1,
normalizer_fn=slim.batch_norm,
scope='east_conv2_3x3')
image_features['east_conv2_3x3'] = east_conv2_3x3
east_deconv1 = slim.conv2d_transpose(east_conv2_3x3, 64, [4, 4], 2, \
padding='SAME', scope='east_deconv1')
east_conv1_concat = tf.concat([east_conv_1, east_deconv1], axis=3)
east_conv1_1x1 = slim.conv2d(east_conv1_concat, 32, [1,1],
stride=1,
normalizer_fn=slim.batch_norm,
scope='east_conv1_1x1')
east_conv1_3x3 = slim.conv2d(east_conv1_1x1, 32, [3,3],
stride=1,
normalizer_fn=slim.batch_norm,
scope='east_conv1_3x3')
image_features['east_conv1_3x3'] = east_conv1_3x3
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()