def _create_feature_extractor(self,
depth_multiplier,
pad_to_multiple,
use_explicit_padding=False,
num_layers=6,
is_training=False,
use_keras=False):
"""Constructs a new feature extractor.
Args:
depth_multiplier: float depth multiplier for feature extractor
pad_to_multiple: the nearest multiple to zero pad the input height and
width dimensions to.
use_explicit_padding: Use 'VALID' padding for convolutions, but prepad
inputs so that the output dimensions are the same as if 'SAME' padding
were used.
num_layers: number of SSD layers.
is_training: whether the network is in training mode.
use_keras: if True builds a keras-based feature extractor, if False builds
a slim-based one.
Returns:
an ssd_meta_arch.SSDFeatureExtractor object.
"""
min_depth = 32
del use_keras
return ssd_mobilenet_v1_feature_extractor.SSDMobileNetV1FeatureExtractor(
is_training,
depth_multiplier,
min_depth,
pad_to_multiple,
self.conv_hyperparams_fn,
use_explicit_padding=use_explicit_padding,
num_layers=num_layers)
def _create_feature_extractor(self,
depth_multiplier,
pad_to_multiple,
use_explicit_padding=False,
num_layers=6,
is_training=False,
use_keras=False):
min_depth = 32
if use_keras:
return (ssd_mobilenet_v1_keras_feature_extractor.
SSDMobileNetV1KerasFeatureExtractor(
is_training=is_training,
depth_multiplier=depth_multiplier,
min_depth=min_depth,
pad_to_multiple=pad_to_multiple,
conv_hyperparams=self._build_conv_hyperparams(
add_batch_norm=False),
freeze_batchnorm=False,
inplace_batchnorm_update=False,
use_explicit_padding=use_explicit_padding,
num_layers=num_layers,
name='MobilenetV1'))
else:
return ssd_mobilenet_v1_feature_extractor.SSDMobileNetV1FeatureExtractor(
is_training,
depth_multiplier,
min_depth,
pad_to_multiple,
self.conv_hyperparams_fn,
use_explicit_padding=use_explicit_padding,
num_layers=num_layers)
def _create_feature_extractor(self,
depth_multiplier,
pad_to_multiple,
is_training=True,
batch_norm_trainable=True,
use_explicit_padding=False):
"""Constructs a new feature extractor.
Args:
depth_multiplier: float depth multiplier for feature extractor
pad_to_multiple: the nearest multiple to zero pad the input height and
width dimensions to.
is_training: whether the network is in training mode.
batch_norm_trainable: Whether to update batch norm parameters during
training or not.
use_explicit_padding: Use 'VALID' padding for convolutions, but prepad
inputs so that the output dimensions are the same as if 'SAME' padding
were used.
Returns:
an ssd_meta_arch.SSDFeatureExtractor object.
"""
min_depth = 32
with slim.arg_scope([slim.conv2d],
normalizer_fn=slim.batch_norm) as sc:
conv_hyperparams = sc
return ssd_mobilenet_v1_feature_extractor.SSDMobileNetV1FeatureExtractor(
is_training,
depth_multiplier,
min_depth,
pad_to_multiple,
conv_hyperparams,
batch_norm_trainable=batch_norm_trainable,
use_explicit_padding=use_explicit_padding)
def _create_feature_extractor(self, depth_multiplier):
"""Constructs a new feature extractor.
Args:
depth_multiplier: float depth multiplier for feature extractor
Returns:
an ssd_meta_arch.SSDFeatureExtractor object.
"""
min_depth = 32
conv_hyperparams = {}
return ssd_mobilenet_v1_feature_extractor.SSDMobileNetV1FeatureExtractor(
depth_multiplier, min_depth, conv_hyperparams)
def _create_feature_extractor(self,
depth_multiplier,
pad_to_multiple,
use_explicit_padding=False,
is_training=False,
use_keras=False):
"""Constructs a new feature extractor.
Args:
depth_multiplier: float depth multiplier for feature extractor
pad_to_multiple: the nearest multiple to zero pad the input height and
width dimensions to.
use_explicit_padding: Use 'VALID' padding for convolutions, but prepad
inputs so that the output dimensions are the same as if 'SAME' padding
were used.
is_training: whether the network is in training mode.
use_keras: if True builds a keras-based feature extractor, if False builds
a slim-based one.
Returns:
an ssd_meta_arch.SSDFeatureExtractor object.
"""
min_depth = 32
if use_keras:
return (ssd_mobilenet_v1_keras_feature_extractor.
SSDMobileNetV1KerasFeatureExtractor(
is_training=is_training,
depth_multiplier=depth_multiplier,
min_depth=min_depth,
pad_to_multiple=pad_to_multiple,
conv_hyperparams=self._build_conv_hyperparams(
add_batch_norm=False),
freeze_batchnorm=False,
inplace_batchnorm_update=False,
use_explicit_padding=use_explicit_padding,
name='MobilenetV1'))
else:
return ssd_mobilenet_v1_feature_extractor.SSDMobileNetV1FeatureExtractor(
is_training,
depth_multiplier,
min_depth,
pad_to_multiple,
self.conv_hyperparams_fn,
use_explicit_padding=use_explicit_padding)
def _create_feature_extractor(self, depth_multiplier, pad_to_multiple,
is_training=True, batch_norm_trainable=True):
"""Constructs a new feature extractor.
Args:
depth_multiplier: float depth multiplier for feature extractor
pad_to_multiple: the nearest multiple to zero pad the input height and
width dimensions to.
is_training: whether the network is in training mode.
batch_norm_trainable: Whether to update batch norm parameters during
training or not.
Returns:
an ssd_meta_arch.SSDFeatureExtractor object.
"""
min_depth = 32
with slim.arg_scope([slim.conv2d], normalizer_fn=slim.batch_norm) as sc:
conv_hyperparams = sc
return ssd_mobilenet_v1_feature_extractor.SSDMobileNetV1FeatureExtractor(
is_training, depth_multiplier, min_depth, pad_to_multiple,
conv_hyperparams, batch_norm_trainable)
