def _pnasnet(self,
images,
backbone,
num_classes,
is_training=True,
output_stride=16,
final_endpoint=None):
"""Build PNASNet model backbone."""
hparams = tf.contrib.training.HParams(
filter_scaling_rate=2.0,
num_conv_filters=10,
drop_path_keep_prob=1.0,
total_training_steps=200000,
)
if not is_training:
hparams.set_hparam('drop_path_keep_prob', 1.0)
cell = nas_genotypes.PNASCell(hparams.num_conv_filters,
hparams.drop_path_keep_prob,
len(backbone),
hparams.total_training_steps)
with arg_scope([slim.dropout, slim.batch_norm],
is_training=is_training):
return nas_network._build_nas_base(images,
cell=cell,
backbone=backbone,
num_classes=num_classes,
hparams=hparams,
reuse=tf.AUTO_REUSE,
scope='pnasnet_small',
final_endpoint=final_endpoint)
def pnasnet(images,
num_classes,
is_training=True,
global_pool=False,
output_stride=16,
nas_architecture_options=None,
nas_training_hyper_parameters=None,
reuse=None,
scope='pnasnet',
final_endpoint=None,
sync_batch_norm_method='None'):
"""Builds PNASNet model."""
if nas_architecture_options is None:
raise ValueError(
'Using NAS model variants. nas_architecture_options cannot be None.'
)
hparams = config(num_conv_filters=nas_architecture_options[
'nas_stem_output_num_conv_filters'])
if nas_training_hyper_parameters:
hparams.set_hparam(
'drop_path_keep_prob',
nas_training_hyper_parameters['drop_path_keep_prob'])
hparams.set_hparam(
'total_training_steps',
nas_training_hyper_parameters['total_training_steps'])
if not is_training:
tf.logging.info('During inference, setting drop_path_keep_prob = 1.0.')
hparams.set_hparam('drop_path_keep_prob', 1.0)
tf.logging.info(hparams)
if output_stride == 8:
backbone = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
elif output_stride == 16:
backbone = [1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2]
elif output_stride == 32:
backbone = [1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3]
else:
raise ValueError('Unsupported output_stride ', output_stride)
batch_norm = utils.get_batch_norm_fn(sync_batch_norm_method)
cell = nas_genotypes.PNASCell(hparams.num_conv_filters,
hparams.drop_path_keep_prob,
len(backbone),
hparams.total_training_steps,
batch_norm_fn=batch_norm)
with arg_scope([slim.dropout, batch_norm], is_training=is_training):
return _build_nas_base(
images,
cell=cell,
backbone=backbone,
num_classes=num_classes,
hparams=hparams,
global_pool=global_pool,
output_stride=output_stride,
nas_use_classification_head=nas_architecture_options[
'nas_use_classification_head'],
reuse=reuse,
scope=scope,
final_endpoint=final_endpoint,
batch_norm_fn=batch_norm,
nas_remove_os32_stride=nas_architecture_options[
'nas_remove_os32_stride'])
def pnasnet(images,
num_classes,
is_training=True,
global_pool=False,
output_stride=16,
nas_stem_output_num_conv_filters=20,
nas_training_hyper_parameters=None,
reuse=None,
scope='pnasnet',
final_endpoint=None):
"""Builds PNASNet model."""
hparams = config(num_conv_filters=nas_stem_output_num_conv_filters)
if nas_training_hyper_parameters:
hparams.set_hparam('drop_path_keep_prob',
nas_training_hyper_parameters['drop_path_keep_prob'])
hparams.set_hparam('total_training_steps',
nas_training_hyper_parameters['total_training_steps'])
if not is_training:
tf.logging.info('During inference, setting drop_path_keep_prob = 1.0.')
hparams.set_hparam('drop_path_keep_prob', 1.0)
tf.logging.info(hparams)
if output_stride == 8:
backbone = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
elif output_stride == 16:
backbone = [1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2]
elif output_stride == 32:
backbone = [1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3]
else:
raise ValueError('Unsupported output_stride ', output_stride)
cell = nas_genotypes.PNASCell(hparams.num_conv_filters,
hparams.drop_path_keep_prob,
len(backbone),
hparams.total_training_steps)
with arg_scope([slim.dropout, slim.batch_norm], is_training=is_training):
return _build_nas_base(
images,
cell=cell,
backbone=backbone,
num_classes=num_classes,
hparams=hparams,
global_pool=global_pool,
reuse=reuse,
scope=scope,
final_endpoint=final_endpoint)
