def testOverrideHParamsLargeModel(self):
batch_size = 5
height, width = 331, 331
num_classes = 1000
inputs = tf.random_uniform((batch_size, height, width, 3))
tf.train.create_global_step()
config = pnasnet.large_imagenet_config()
config.set_hparam('data_format', 'NCHW')
with slim.arg_scope(pnasnet.pnasnet_large_arg_scope()):
_, end_points = pnasnet.build_pnasnet_large(
inputs, num_classes, config=config)
self.assertListEqual(
end_points['Stem'].shape.as_list(), [batch_size, 540, 42, 42])
def testOverrideHParamsLargeModel(self):
batch_size = 5
height, width = 331, 331
num_classes = 1000
inputs = tf.random_uniform((batch_size, height, width, 3))
tf.train.create_global_step()
config = pnasnet.large_imagenet_config()
config.set_hparam('data_format', 'NCHW')
with slim.arg_scope(pnasnet.pnasnet_large_arg_scope()):
_, end_points = pnasnet.build_pnasnet_large(
inputs, num_classes, config=config)
self.assertListEqual(
end_points['Stem'].shape.as_list(), [batch_size, 540, 42, 42])
def testNoAuxHeadLargeModel(self):
batch_size = 5
height, width = 331, 331
num_classes = 1000
for use_aux_head in (True, False):
tf.reset_default_graph()
inputs = tf.random_uniform((batch_size, height, width, 3))
tf.train.create_global_step()
config = pnasnet.large_imagenet_config()
config.set_hparam('use_aux_head', int(use_aux_head))
with slim.arg_scope(pnasnet.pnasnet_large_arg_scope()):
_, end_points = pnasnet.build_pnasnet_large(inputs, num_classes,
config=config)
self.assertEqual('AuxLogits' in end_points, use_aux_head)
def testNoAuxHeadLargeModel(self):
batch_size = 5
height, width = 331, 331
num_classes = 1000
for use_aux_head in (True, False):
tf.reset_default_graph()
inputs = tf.random_uniform((batch_size, height, width, 3))
tf.train.create_global_step()
config = pnasnet.large_imagenet_config()
config.set_hparam('use_aux_head', int(use_aux_head))
with slim.arg_scope(pnasnet.pnasnet_large_arg_scope()):
_, end_points = pnasnet.build_pnasnet_large(inputs, num_classes,
config=config)
self.assertEqual('AuxLogits' in end_points, use_aux_head)
def pnasnet_large(inputs, is_training, opts):
with slim.arg_scope(pnasnet.pnasnet_large_arg_scope(
weight_decay=opts.weight_decay,
batch_norm_decay=opts.batch_norm_decay,
batch_norm_epsilon=opts.batch_norm_epsilon)):
config = pnasnet.large_imagenet_config()
config.set_hparam('dense_dropout_keep_prob', opts.dropout_keep_prob)
config.set_hparam('use_aux_head', int(opts.create_aux_logits))
return pnasnet.build_pnasnet_large(
inputs,
num_classes=opts.num_classes,
is_training=is_training,
config=config)
def _extract_box_classifier_features(self, proposal_feature_maps, scope):
"""Extracts second stage box classifier features.
This function reconstructs the "second half" of the PNASNet
network after the part defined in `_extract_proposal_features`.
Args:
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.
scope: A scope name.
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.
"""
del scope
# Number of used stem cells.
num_stem_cells = 2
# Note that we always feed into 2 layers of equal depth
# where the first N channels corresponds to previous hidden layer
# and the second N channels correspond to the final hidden layer.
hidden_previous, hidden = tf.split(proposal_feature_maps, 2, axis=3)
# Note that what follows is largely a copy of build_pnasnet_large() within
# pnasnet.py. We are copying to minimize code pollution in slim.
# TODO(shlens,skornblith): Determine the appropriate drop path schedule.
# For now the schedule is the default (1.0->0.7 over 250,000 train steps).
hparams = pnasnet.large_imagenet_config()
if not self._is_training:
hparams.set_hparam('drop_path_keep_prob', 1.0)
# Calculate the total number of cells in the network
total_num_cells = hparams.num_cells + num_stem_cells
normal_cell = pnasnet.PNasNetNormalCell(hparams.num_conv_filters,
hparams.drop_path_keep_prob,
total_num_cells,
hparams.total_training_steps)
with arg_scope([slim.dropout, nasnet_utils.drop_path],
is_training=self._is_training):
with arg_scope([slim.batch_norm],
is_training=self._train_batch_norm):
with arg_scope([
slim.avg_pool2d, slim.max_pool2d, slim.conv2d,
slim.batch_norm, slim.separable_conv2d,
nasnet_utils.factorized_reduction,
nasnet_utils.global_avg_pool,
nasnet_utils.get_channel_index,
nasnet_utils.get_channel_dim
],
data_format=hparams.data_format):
# This corresponds to the cell number just past 'Cell_7' used by
# _extract_proposal_features().
start_cell_num = 8
true_cell_num = start_cell_num + num_stem_cells
with slim.arg_scope(pnasnet.pnasnet_large_arg_scope()):
net = _build_pnasnet_base(
hidden_previous,
hidden,
normal_cell=normal_cell,
hparams=hparams,
true_cell_num=true_cell_num,
start_cell_num=start_cell_num)
proposal_classifier_features = net
return proposal_classifier_features
def _extract_box_classifier_features(self, proposal_feature_maps, scope):
"""Extracts second stage box classifier features.
This function reconstructs the "second half" of the PNASNet
network after the part defined in `_extract_proposal_features`.
Args:
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.
scope: A scope name.
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.
"""
del scope
# Number of used stem cells.
num_stem_cells = 2
# Note that we always feed into 2 layers of equal depth
# where the first N channels corresponds to previous hidden layer
# and the second N channels correspond to the final hidden layer.
hidden_previous, hidden = tf.split(proposal_feature_maps, 2, axis=3)
# Note that what follows is largely a copy of build_pnasnet_large() within
# pnasnet.py. We are copying to minimize code pollution in slim.
# TODO(shlens,skornblith): Determine the appropriate drop path schedule.
# For now the schedule is the default (1.0->0.7 over 250,000 train steps).
hparams = pnasnet.large_imagenet_config()
if not self._is_training:
hparams.set_hparam('drop_path_keep_prob', 1.0)
# Calculate the total number of cells in the network
total_num_cells = hparams.num_cells + num_stem_cells
normal_cell = pnasnet.PNasNetNormalCell(
hparams.num_conv_filters, hparams.drop_path_keep_prob,
total_num_cells, hparams.total_training_steps)
with arg_scope([slim.dropout, nasnet_utils.drop_path],
is_training=self._is_training):
with arg_scope([slim.batch_norm], is_training=self._train_batch_norm):
with arg_scope([slim.avg_pool2d,
slim.max_pool2d,
slim.conv2d,
slim.batch_norm,
slim.separable_conv2d,
nasnet_utils.factorized_reduction,
nasnet_utils.global_avg_pool,
nasnet_utils.get_channel_index,
nasnet_utils.get_channel_dim],
data_format=hparams.data_format):
# This corresponds to the cell number just past 'Cell_7' used by
# _extract_proposal_features().
start_cell_num = 8
true_cell_num = start_cell_num + num_stem_cells
with slim.arg_scope(pnasnet.pnasnet_large_arg_scope()):
net = _build_pnasnet_base(
hidden_previous,
hidden,
normal_cell=normal_cell,
hparams=hparams,
true_cell_num=true_cell_num,
start_cell_num=start_cell_num)
proposal_classifier_features = net
return proposal_classifier_features
