def testBuildPreLogitsLargeModel(self):
batch_size = 5
height, width = 331, 331
num_classes = None
inputs = tf.random_uniform((batch_size, height, width, 3))
tf.train.create_global_step()
with slim.arg_scope(nasnet.nasnet_large_arg_scope()):
net, end_points = nasnet.build_nasnet_large(inputs, num_classes)
self.assertFalse('AuxLogits' in end_points)
self.assertFalse('Predictions' in end_points)
self.assertTrue(net.op.name.startswith('final_layer/Mean'))
self.assertListEqual(net.get_shape().as_list(), [batch_size, 4032])
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 = nasnet.large_imagenet_config()
config.set_hparam('data_format', 'NCHW')
with slim.arg_scope(nasnet.nasnet_large_arg_scope()):
_, end_points = nasnet.build_nasnet_large(inputs,
num_classes,
config=config)
self.assertListEqual(end_points['Stem'].shape.as_list(),
[batch_size, 336, 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 = nasnet.large_imagenet_config()
config.set_hparam('use_aux_head', int(use_aux_head))
with slim.arg_scope(nasnet.nasnet_large_arg_scope()):
_, end_points = nasnet.build_nasnet_large(inputs,
num_classes,
config=config)
self.assertEqual('AuxLogits' in end_points, use_aux_head)
def testBuildLogitsLargeModel(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()
with slim.arg_scope(nasnet.nasnet_large_arg_scope()):
logits, end_points = nasnet.build_nasnet_large(inputs, num_classes)
auxlogits = end_points['AuxLogits']
predictions = end_points['Predictions']
self.assertListEqual(auxlogits.get_shape().as_list(),
[batch_size, num_classes])
self.assertListEqual(logits.get_shape().as_list(),
[batch_size, num_classes])
self.assertListEqual(predictions.get_shape().as_list(),
[batch_size, num_classes])
def nasnet_large_arg_scope_for_detection(is_batch_norm_training=False):
"""Defines the default arg scope for the NASNet-A Large for object detection.
This provides a small edit to switch batch norm training on and off.
Args:
is_batch_norm_training: Boolean indicating whether to train with batch norm.
Returns:
An `arg_scope` to use for the NASNet Large Model.
"""
imagenet_scope = nasnet.nasnet_large_arg_scope()
with arg_scope(imagenet_scope):
with arg_scope([slim.batch_norm],
is_training=is_batch_norm_training) as sc:
return sc
def testAllEndPointsShapesLargeModel(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()
with slim.arg_scope(nasnet.nasnet_large_arg_scope()):
_, end_points = nasnet.build_nasnet_large(inputs, num_classes)
endpoints_shapes = {
'Stem': [batch_size, 42, 42, 336],
'Cell_0': [batch_size, 42, 42, 1008],
'Cell_1': [batch_size, 42, 42, 1008],
'Cell_2': [batch_size, 42, 42, 1008],
'Cell_3': [batch_size, 42, 42, 1008],
'Cell_4': [batch_size, 42, 42, 1008],
'Cell_5': [batch_size, 42, 42, 1008],
'Cell_6': [batch_size, 21, 21, 2016],
'Cell_7': [batch_size, 21, 21, 2016],
'Cell_8': [batch_size, 21, 21, 2016],
'Cell_9': [batch_size, 21, 21, 2016],
'Cell_10': [batch_size, 21, 21, 2016],
'Cell_11': [batch_size, 21, 21, 2016],
'Cell_12': [batch_size, 11, 11, 4032],
'Cell_13': [batch_size, 11, 11, 4032],
'Cell_14': [batch_size, 11, 11, 4032],
'Cell_15': [batch_size, 11, 11, 4032],
'Cell_16': [batch_size, 11, 11, 4032],
'Cell_17': [batch_size, 11, 11, 4032],
'Reduction_Cell_0': [batch_size, 21, 21, 1344],
'Reduction_Cell_1': [batch_size, 11, 11, 2688],
'global_pool': [batch_size, 4032],
# Logits and predictions
'AuxLogits': [batch_size, num_classes],
'Logits': [batch_size, num_classes],
'Predictions': [batch_size, num_classes]
}
self.assertItemsEqual(endpoints_shapes.keys(), end_points.keys())
for endpoint_name in endpoints_shapes:
tf.logging.info('Endpoint name: {}'.format(endpoint_name))
expected_shape = endpoints_shapes[endpoint_name]
self.assertTrue(endpoint_name in end_points)
self.assertListEqual(
end_points[endpoint_name].get_shape().as_list(),
expected_shape)
def pnasnet_large_arg_scope(weight_decay=4e-5,
batch_norm_decay=0.9997,
batch_norm_epsilon=0.001):
"""Default arg scope for the PNASNet Large ImageNet model."""
return nasnet.nasnet_large_arg_scope(weight_decay, batch_norm_decay,
batch_norm_epsilon)
def _extract_box_classifier_features(self, proposal_feature_maps, scope):
"""Extracts second stage box classifier features.
This function reconstructs the "second half" of the NASNet-A
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
# 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_nasnet_large() within
# nasnet.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 = nasnet.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
# -- Add 2 for the reduction cells.
total_num_cells = hparams.num_cells + 2
# -- And add 2 for the stem cells for ImageNet training.
total_num_cells += 2
normal_cell = nasnet_utils.NasNetANormalCell(
hparams.num_conv_filters, hparams.drop_path_keep_prob,
total_num_cells, hparams.total_training_steps)
reduction_cell = nasnet_utils.NasNetAReductionCell(
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_11' used by
# by _extract_proposal_features().
start_cell_num = 12
# Note that this number equals:
# start_cell_num + 2 stem cells + 1 reduction cell
true_cell_num = 15
with slim.arg_scope(nasnet.nasnet_large_arg_scope()):
net = _build_nasnet_base(hidden_previous,
hidden,
normal_cell=normal_cell,
reduction_cell=reduction_cell,
hparams=hparams,
true_cell_num=true_cell_num,
start_cell_num=start_cell_num)
proposal_classifier_features = net
return proposal_classifier_features
