def testReuseArgScope(self):
func1_kwargs = {'a': 1, 'b': None, 'c': [1]}
key_op = _key_op(func1)
current_scope = {key_op: func1_kwargs.copy()}
with self.cached_session():
with arg_scope([func1], a=1, b=None, c=[1]) as scope1:
pass
with arg_scope(scope1) as scope:
self.assertDictEqual(scope, current_scope)
def overfeat_arg_scope(weight_decay=0.0005):
with arg_scope(
[layers.conv2d, layers_lib.fully_connected],
activation_fn=nn_ops.relu,
weights_regularizer=regularizers.l2_regularizer(weight_decay),
biases_initializer=init_ops.zeros_initializer()):
with arg_scope([layers.conv2d], padding='SAME'):
with arg_scope([layers_lib.max_pool2d], padding='VALID') as arg_sc:
return arg_sc
def testClearArgScope(self):
func1_kwargs = {'a': 1, 'b': None, 'c': [1]}
key_op = _key_op(func1)
func1_scope = {key_op: func1_kwargs.copy()}
with self.cached_session():
with arg_scope([func1], a=1, b=None, c=[1]) as sc1:
self.assertEqual(sc1, func1_scope)
with arg_scope({}) as sc2:
self.assertEqual(sc2, {})
with arg_scope([]) as current_arg_scope:
self.assertEqual(current_arg_scope, func1_scope)
def testNestedArgScopeObjectCreatedOutsideScopeOverridesArgScope(self):
def get_scope_object():
with arg_scope([func1], a=1, b=None, c=[1]) as sc:
return sc
scope_object = get_scope_object()
with arg_scope([func1], b=2, d=10):
with arg_scope(scope_object):
args, kwargs = func1(0)
self.assertTupleEqual(args, (0, ))
self.assertDictEqual(kwargs, {'a': 1, 'b': None, 'c': [1]})
def testCurrentArgScopeNested(self):
func1_kwargs = {'a': 1, 'b': None, 'c': [1]}
func2_kwargs = {'b': 2, 'd': [2]}
key = _key_op
current_scope = {
key(func1): func1_kwargs.copy(),
key(func2): func2_kwargs.copy()
}
with self.cached_session():
with arg_scope([func1], a=1, b=None, c=[1]):
with arg_scope([func2], b=2, d=[2]) as scope:
self.assertDictEqual(scope, current_scope)
def testNestedArgScope(self):
func1_args = (0, )
func1_kwargs = {'a': 1, 'b': None, 'c': [1]}
with arg_scope([func1], a=1, b=None, c=[1]):
args, kwargs = func1(0)
self.assertTupleEqual(args, func1_args)
self.assertDictEqual(kwargs, func1_kwargs)
func1_kwargs['b'] = 2
with arg_scope([func1], b=2):
args, kwargs = func1(0)
self.assertTupleEqual(args, func1_args)
self.assertDictEqual(kwargs, func1_kwargs)
def testPartiallySharedArgScope(self):
func1_args = (0, )
func1_kwargs = {'a': 1, 'b': None, 'c': [1]}
func2_args = (1, )
func2_kwargs = {'a': 1, 'b': None, 'd': [2]}
with arg_scope([func1, func2], a=1, b=None):
with arg_scope([func1], c=[1]):
with arg_scope([func2], d=[2]):
args, kwargs = func1(0)
self.assertTupleEqual(args, func1_args)
self.assertDictEqual(kwargs, func1_kwargs)
args, kwargs = func2(1)
self.assertTupleEqual(args, func2_args)
self.assertDictEqual(kwargs, func2_kwargs)
def inception_v3_arg_scope(weight_decay=0.00004,
batch_norm_var_collection='moving_vars',
batch_norm_decay=0.9997,
batch_norm_epsilon=0.001,
updates_collections=ops.GraphKeys.UPDATE_OPS,
use_fused_batchnorm=True):
"""Defines the default InceptionV3 arg scope.
Args:
weight_decay: The weight decay to use for regularizing the model.
batch_norm_var_collection: The name of the collection for the batch norm
variables.
batch_norm_decay: Decay for batch norm moving average
batch_norm_epsilon: Small float added to variance to avoid division by zero
updates_collections: Collections for the update ops of the layer
use_fused_batchnorm: Enable fused batchnorm.
Returns:
An `arg_scope` to use for the inception v3 model.
"""
batch_norm_params = {
# Decay for the moving averages.
'decay': batch_norm_decay,
# epsilon to prevent 0s in variance.
'epsilon': batch_norm_epsilon,
# collection containing update_ops.
'updates_collections': updates_collections,
# Use fused batch norm if possible.
'fused': use_fused_batchnorm,
# collection containing the moving mean and moving variance.
'variables_collections': {
'beta': None,
'gamma': None,
'moving_mean': [batch_norm_var_collection],
'moving_variance': [batch_norm_var_collection],
}
}
# Set weight_decay for weights in Conv and FC layers.
with arg_scope(
[layers.conv2d, layers_lib.fully_connected],
weights_regularizer=regularizers.l2_regularizer(weight_decay)):
with arg_scope([layers.conv2d],
weights_initializer=initializers.
variance_scaling_initializer(),
activation_fn=nn_ops.relu,
normalizer_fn=layers_lib.batch_norm,
normalizer_params=batch_norm_params) as sc:
return sc
def testArgScopeObjectCreatedWithinScopeInheritsArgScope(self):
def get_scope_object():
with arg_scope([func1], a=1, b=None, c=[1]) as sc:
return sc
with arg_scope([func1], b=2, d=10):
with arg_scope(get_scope_object()):
args, kwargs = func1(0)
self.assertTupleEqual(args, (0, ))
self.assertDictEqual(kwargs, {
'a': 1,
'b': None,
'c': [1],
'd': 10
})
def inception_v1_arg_scope(weight_decay=0.00004,
use_batch_norm=True,
batch_norm_var_collection='moving_vars'):
"""Defines the default InceptionV1 arg scope.
Note: Althougth the original paper didn't use batch_norm we found it useful.
Args:
weight_decay: The weight decay to use for regularizing the model.
use_batch_norm: "If `True`, batch_norm is applied after each convolution.
batch_norm_var_collection: The name of the collection for the batch norm
variables.
Returns:
An `arg_scope` to use for the inception v3 model.
"""
batch_norm_params = {
# Decay for the moving averages.
'decay': 0.9997,
# epsilon to prevent 0s in variance.
'epsilon': 0.001,
# collection containing update_ops.
'updates_collections': ops.GraphKeys.UPDATE_OPS,
# collection containing the moving mean and moving variance.
'variables_collections': {
'beta': None,
'gamma': None,
'moving_mean': [batch_norm_var_collection],
'moving_variance': [batch_norm_var_collection],
}
}
if use_batch_norm:
normalizer_fn = layers_lib.batch_norm
normalizer_params = batch_norm_params
else:
normalizer_fn = None
normalizer_params = {}
# Set weight_decay for weights in Conv and FC layers.
with arg_scope(
[layers.conv2d, layers_lib.fully_connected],
weights_regularizer=regularizers.l2_regularizer(weight_decay)):
with arg_scope([layers.conv2d],
weights_initializer=initializers.
variance_scaling_initializer(),
activation_fn=nn_ops.relu,
normalizer_fn=normalizer_fn,
normalizer_params=normalizer_params) as sc:
return sc
def testNonDecorated(self):
def my_func(t, a=None):
return (t, a)
with self.assertRaises(ValueError):
with arg_scope([my_func], a=1):
pass
def testEndPointsV2(self):
"""Test the end points of a tiny v2 bottleneck network."""
blocks = [
resnet_v2.resnet_v2_block(
'block1', base_depth=1, num_units=2, stride=2),
resnet_v2.resnet_v2_block(
'block2', base_depth=2, num_units=2, stride=1),
]
inputs = create_test_input(2, 32, 16, 3)
with arg_scope(resnet_utils.resnet_arg_scope()):
_, end_points = self._resnet_plain(inputs, blocks, scope='tiny')
expected = [
'tiny/block1/unit_1/bottleneck_v2/shortcut',
'tiny/block1/unit_1/bottleneck_v2/conv1',
'tiny/block1/unit_1/bottleneck_v2/conv2',
'tiny/block1/unit_1/bottleneck_v2/conv3',
'tiny/block1/unit_2/bottleneck_v2/conv1',
'tiny/block1/unit_2/bottleneck_v2/conv2',
'tiny/block1/unit_2/bottleneck_v2/conv3',
'tiny/block2/unit_1/bottleneck_v2/shortcut',
'tiny/block2/unit_1/bottleneck_v2/conv1',
'tiny/block2/unit_1/bottleneck_v2/conv2',
'tiny/block2/unit_1/bottleneck_v2/conv3',
'tiny/block2/unit_2/bottleneck_v2/conv1',
'tiny/block2/unit_2/bottleneck_v2/conv2',
'tiny/block2/unit_2/bottleneck_v2/conv3'
]
self.assertItemsEqual(expected, end_points)
def _resnet_plain(self, inputs, blocks, output_stride=None, scope=None):
"""A plain ResNet without extra layers before or after the ResNet blocks."""
with variable_scope.variable_scope(scope, values=[inputs]):
with arg_scope([layers.conv2d], outputs_collections='end_points'):
net = resnet_utils.stack_blocks_dense(inputs, blocks, output_stride)
end_points = utils.convert_collection_to_dict('end_points')
return net, end_points
def testAtrousFullyConvolutionalValues(self):
"""Verify dense feature extraction with atrous convolution."""
nominal_stride = 32
for output_stride in [4, 8, 16, 32, None]:
with arg_scope(resnet_utils.resnet_arg_scope()):
with ops.Graph().as_default():
with self.cached_session() as sess:
random_seed.set_random_seed(0)
inputs = create_test_input(2, 81, 81, 3)
# Dense feature extraction followed by subsampling.
output, _ = self._resnet_small(
inputs,
None,
is_training=False,
global_pool=False,
output_stride=output_stride)
if output_stride is None:
factor = 1
else:
factor = nominal_stride // output_stride
output = resnet_utils.subsample(output, factor)
# Make the two networks use the same weights.
variable_scope.get_variable_scope().reuse_variables()
# Feature extraction at the nominal network rate.
expected, _ = self._resnet_small(
inputs, None, is_training=False, global_pool=False)
sess.run(variables.global_variables_initializer())
self.assertAllClose(
output.eval(), expected.eval(), atol=1e-4, rtol=1e-4)
def vgg_arg_scope(weight_decay=0.0005):
"""Defines the VGG arg scope.
Args:
weight_decay: The l2 regularization coefficient.
Returns:
An arg_scope.
"""
with arg_scope(
[layers.conv2d, layers_lib.fully_connected],
activation_fn=nn_ops.relu,
weights_regularizer=regularizers.l2_regularizer(weight_decay),
biases_initializer=init_ops.zeros_initializer()):
with arg_scope([layers.conv2d], padding='SAME') as arg_sc:
return arg_sc
def testOverwriteArgScope(self):
func1_args = (0, )
func1_kwargs = {'a': 1, 'b': 2, 'c': [1]}
with arg_scope([func1], a=1, b=None, c=[1]):
args, kwargs = func1(0, b=2)
self.assertTupleEqual(args, func1_args)
self.assertDictEqual(kwargs, func1_kwargs)
def testAtrousValuesBottleneck(self):
"""Verify the values of dense feature extraction by atrous convolution.
Make sure that dense feature extraction by stack_blocks_dense() followed by
subsampling gives identical results to feature extraction at the nominal
network output stride using the simple self._stack_blocks_nondense() above.
"""
block = resnet_v1.resnet_v1_block
blocks = [
block('block1', base_depth=1, num_units=2, stride=2),
block('block2', base_depth=2, num_units=2, stride=2),
block('block3', base_depth=4, num_units=2, stride=2),
block('block4', base_depth=8, num_units=2, stride=1),
]
nominal_stride = 8
# Test both odd and even input dimensions.
height = 30
width = 31
with arg_scope(resnet_utils.resnet_arg_scope()):
with arg_scope([layers.batch_norm], is_training=False):
for output_stride in [1, 2, 4, 8, None]:
with ops.Graph().as_default():
with self.cached_session() as sess:
random_seed.set_random_seed(0)
inputs = create_test_input(1, height, width, 3)
# Dense feature extraction followed by subsampling.
output = resnet_utils.stack_blocks_dense(
inputs, blocks, output_stride)
if output_stride is None:
factor = 1
else:
factor = nominal_stride // output_stride
output = resnet_utils.subsample(output, factor)
# Make the two networks use the same weights.
variable_scope.get_variable_scope(
).reuse_variables()
# Feature extraction at the nominal network rate.
expected = self._stack_blocks_nondense(
inputs, blocks)
sess.run(variables.global_variables_initializer())
output, expected = sess.run([output, expected])
self.assertAllClose(output,
expected,
atol=1e-4,
rtol=1e-4)
def resnet_arg_scope(weight_decay=0.0001,
batch_norm_decay=0.997,
batch_norm_epsilon=1e-5,
batch_norm_scale=True):
"""Defines the default ResNet arg scope.
TODO(gpapan): The batch-normalization related default values above are
appropriate for use in conjunction with the reference ResNet models
released at https://github.com/KaimingHe/deep-residual-networks. When
training ResNets from scratch, they might need to be tuned.
Args:
weight_decay: The weight decay to use for regularizing the model.
batch_norm_decay: The moving average decay when estimating layer activation
statistics in batch normalization.
batch_norm_epsilon: Small constant to prevent division by zero when
normalizing activations by their variance in batch normalization.
batch_norm_scale: If True, uses an explicit `gamma` multiplier to scale the
activations in the batch normalization layer.
Returns:
An `arg_scope` to use for the resnet models.
"""
batch_norm_params = {
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
'scale': batch_norm_scale,
'updates_collections': ops.GraphKeys.UPDATE_OPS,
}
with arg_scope(
[layers_lib.conv2d],
weights_regularizer=regularizers.l2_regularizer(weight_decay),
weights_initializer=initializers.variance_scaling_initializer(),
activation_fn=nn_ops.relu,
normalizer_fn=layers.batch_norm,
normalizer_params=batch_norm_params):
with arg_scope([layers.batch_norm], **batch_norm_params):
# The following implies padding='SAME' for pool1, which makes feature
# alignment easier for dense prediction tasks. This is also used in
# https://github.com/facebook/fb.resnet.torch. However the accompanying
# code of 'Deep Residual Learning for Image Recognition' uses
# padding='VALID' for pool1. You can switch to that choice by setting
# tf.contrib.framework.arg_scope([tf.contrib.layers.max_pool2d], padding='VALID').
with arg_scope([layers.max_pool2d], padding='SAME') as arg_sc:
return arg_sc
def testSimpleArgScopeWithTuple(self):
func1_args = (0, )
func1_kwargs = {'a': 1, 'b': None, 'c': [1]}
with self.cached_session():
with arg_scope((func1, ), a=1, b=None, c=[1]):
args, kwargs = func1(0)
self.assertTupleEqual(args, func1_args)
self.assertDictEqual(kwargs, func1_kwargs)
def testModelHasExpectedNumberOfParameters(self):
batch_size = 5
height, width = 224, 224
inputs = random_ops.random_uniform((batch_size, height, width, 3))
with arg_scope(inception_v2.inception_v2_arg_scope()):
inception_v2.inception_v2_base(inputs)
total_params, _ = model_analyzer.analyze_vars(
variables_lib.get_model_variables())
self.assertAlmostEqual(10173112, total_params)
def testSharedArgScopeTuple(self):
func1_args = (0, )
func1_kwargs = {'a': 1, 'b': None, 'c': [1]}
with arg_scope((func1, func2), a=1, b=None, c=[1]):
args, kwargs = func1(0)
self.assertTupleEqual(args, func1_args)
self.assertDictEqual(kwargs, func1_kwargs)
args, kwargs = func2(0)
self.assertTupleEqual(args, func1_args)
self.assertDictEqual(kwargs, func1_kwargs)
def testClassificationEndPoints(self):
global_pool = True
num_classes = 10
inputs = create_test_input(2, 224, 224, 3)
with arg_scope(resnet_utils.resnet_arg_scope()):
logits, end_points = self._resnet_small(
inputs, num_classes, global_pool=global_pool, scope='resnet')
self.assertTrue(logits.op.name.startswith('resnet/logits'))
self.assertListEqual(logits.get_shape().as_list(), [2, 1, 1, num_classes])
self.assertTrue('predictions' in end_points)
self.assertListEqual(end_points['predictions'].get_shape().as_list(),
[2, 1, 1, num_classes])
def inception_v2_arg_scope(weight_decay=0.00004,
batch_norm_var_collection='moving_vars'):
"""Defines the default InceptionV2 arg scope.
Args:
weight_decay: The weight decay to use for regularizing the model.
batch_norm_var_collection: The name of the collection for the batch norm
variables.
Returns:
An `arg_scope` to use for the inception v3 model.
"""
batch_norm_params = {
# Decay for the moving averages.
'decay': 0.9997,
# epsilon to prevent 0s in variance.
'epsilon': 0.001,
# collection containing update_ops.
'updates_collections': ops.GraphKeys.UPDATE_OPS,
# collection containing the moving mean and moving variance.
'variables_collections': {
'beta': None,
'gamma': None,
'moving_mean': [batch_norm_var_collection],
'moving_variance': [batch_norm_var_collection],
}
}
# Set weight_decay for weights in Conv and FC layers.
with arg_scope(
[layers.conv2d, layers_lib.fully_connected],
weights_regularizer=regularizers.l2_regularizer(weight_decay)):
with arg_scope([layers.conv2d],
weights_initializer=initializers.
variance_scaling_initializer(),
activation_fn=nn_ops.relu,
normalizer_fn=layers_lib.batch_norm,
normalizer_params=batch_norm_params) as sc:
return sc
def testFullyConvolutionalUnknownHeightWidth(self):
batch = 2
height, width = 65, 65
global_pool = False
inputs = create_test_input(batch, None, None, 3)
with arg_scope(resnet_utils.resnet_arg_scope()):
output, _ = self._resnet_small(inputs, None, global_pool=global_pool)
self.assertListEqual(output.get_shape().as_list(), [batch, None, None, 32])
images = create_test_input(batch, height, width, 3)
with self.cached_session() as sess:
sess.run(variables.global_variables_initializer())
output = sess.run(output, {inputs: images.eval()})
self.assertEqual(output.shape, (batch, 3, 3, 32))
def testClassificationShapes(self):
global_pool = True
num_classes = 10
inputs = create_test_input(2, 224, 224, 3)
with arg_scope(resnet_utils.resnet_arg_scope()):
_, end_points = self._resnet_small(
inputs, num_classes, global_pool=global_pool, scope='resnet')
endpoint_to_shape = {
'resnet/block1': [2, 28, 28, 4],
'resnet/block2': [2, 14, 14, 8],
'resnet/block3': [2, 7, 7, 16],
'resnet/block4': [2, 7, 7, 32]
}
for endpoint in endpoint_to_shape:
shape = endpoint_to_shape[endpoint]
self.assertListEqual(end_points[endpoint].get_shape().as_list(), shape)
def testFullyConvolutionalEndpointShapes(self):
global_pool = False
num_classes = 10
inputs = create_test_input(2, 321, 321, 3)
with arg_scope(resnet_utils.resnet_arg_scope()):
_, end_points = self._resnet_small(
inputs, num_classes, global_pool=global_pool, scope='resnet')
endpoint_to_shape = {
'resnet/block1': [2, 41, 41, 4],
'resnet/block2': [2, 21, 21, 8],
'resnet/block3': [2, 11, 11, 16],
'resnet/block4': [2, 11, 11, 32]
}
for endpoint in endpoint_to_shape:
shape = endpoint_to_shape[endpoint]
self.assertListEqual(end_points[endpoint].get_shape().as_list(), shape)
def _recomputing_grad_fn(compute_fn, original_args, original_vars,
output_grads, grad_fn_variables, use_data_dep,
tupleize_grads, arg_scope, var_scope,
has_is_recompute_kwarg):
"""Grad fn for recompute_grad."""
variables = grad_fn_variables or []
# Identity ops around the inputs ensures correct gradient graph-walking.
inputs = [array_ops.identity(x) for x in list(original_args)]
# Recompute outputs
# Use a control dependency to ensure that the recompute is not eliminated by
# CSE and that it happens on the backwards pass.
ctrl_dep_grads = [g for g in output_grads if g is not None]
with framework_ops.control_dependencies(ctrl_dep_grads):
if use_data_dep:
inputs = _force_data_dependency(output_grads, inputs)
# Re-enter scopes
with contrib_framework_ops.arg_scope(arg_scope):
with variable_scope.variable_scope(var_scope, reuse=True):
# Re-call the function and ensure that the touched variables are the
# same as in the first call.
with backprop.GradientTape() as tape:
fn_kwargs = {}
if has_is_recompute_kwarg:
fn_kwargs["is_recomputing"] = True
outputs = compute_fn(*inputs, **fn_kwargs)
recompute_vars = set(tape.watched_variables())
if original_vars != recompute_vars:
raise ValueError(_WRONG_VARS_ERR)
if not isinstance(outputs, (list, tuple)):
outputs = [outputs]
outputs = list(outputs)
# Compute gradients
grads = gradients_impl.gradients(outputs, inputs + variables, output_grads)
if tupleize_grads:
if use_data_dep:
grads = _tuple_with_data_dep(grads)
else:
grads = control_flow_ops.tuple(grads)
grad_inputs = grads[:len(inputs)]
grad_vars = grads[len(inputs):]
return grad_inputs, grad_vars
def testUnknownBatchSize(self):
batch = 2
height, width = 65, 65
global_pool = True
num_classes = 10
inputs = create_test_input(None, height, width, 3)
with arg_scope(resnet_utils.resnet_arg_scope()):
logits, _ = self._resnet_small(
inputs, num_classes, global_pool=global_pool, scope='resnet')
self.assertTrue(logits.op.name.startswith('resnet/logits'))
self.assertListEqual(logits.get_shape().as_list(),
[None, 1, 1, num_classes])
images = create_test_input(batch, height, width, 3)
with self.cached_session() as sess:
sess.run(variables.global_variables_initializer())
output = sess.run(logits, {inputs: images.eval()})
self.assertEqual(output.shape, (batch, 1, 1, num_classes))
def test_conv_layer(self):
with compat.forward_compatibility_horizon(2019, 6, 7):
g = ops.Graph()
with g.as_default():
inputs = array_ops.placeholder(dtypes.float32, shape=[8, 5, 5, 3])
with contrib_ops.arg_scope([layers.batch_norm],
fused=True,
is_training=True,
trainable=True):
return layers.convolution(
inputs,
num_outputs=16,
kernel_size=3,
stride=1,
padding='VALID',
activation_fn=nn_ops.relu,
normalizer_fn=layers.batch_norm,
normalizer_params={},
weights_initializer=initializers.xavier_initializer(),
weights_regularizer=None,
biases_initializer=init_ops.zeros_initializer(),
biases_regularizer=None,
reuse=None,
trainable=True,
scope=None)
inputs_pattern = graph_matcher.OpTypePattern('*', name='inputs')
relu_pattern = graph_matcher.OpTypePattern(
'Relu',
name='relu',
inputs=[
graph_matcher.OpTypePattern(
'FusedBatchNormV3',
inputs=[
graph_matcher.OpTypePattern(
'Conv2D', inputs=[inputs_pattern, '*']), '*', '*',
'*', '*'
])
])
matcher = graph_matcher.GraphMatcher(relu_pattern)
match_results = list(matcher.match_graph(g))
self.assertEqual(1, len(match_results))
match_result = match_results[0]
self.assertEqual(match_result.get_tensor(inputs_pattern), inputs)
self.assertEqual(match_result.get_tensor('inputs'), inputs)
def overfeat(inputs,
num_classes=1000,
is_training=True,
dropout_keep_prob=0.5,
spatial_squeeze=True,
scope='overfeat'):
"""Contains the model definition for the OverFeat network.
The definition for the network was obtained from:
OverFeat: Integrated Recognition, Localization and Detection using
Convolutional Networks
Pierre Sermanet, David Eigen, Xiang Zhang, Michael Mathieu, Rob Fergus and
Yann LeCun, 2014
http://arxiv.org/abs/1312.6229
Note: All the fully_connected layers have been transformed to conv2d layers.
To use in classification mode, resize input to 231x231. To use in fully
convolutional mode, set spatial_squeeze to false.
Args:
inputs: a tensor of size [batch_size, height, width, channels].
num_classes: number of predicted classes.
is_training: whether or not the model is being trained.
dropout_keep_prob: the probability that activations are kept in the dropout
layers during training.
spatial_squeeze: whether or not should squeeze the spatial dimensions of the
outputs. Useful to remove unnecessary dimensions for classification.
scope: Optional scope for the variables.
Returns:
the last op containing the log predictions and end_points dict.
"""
with variable_scope.variable_scope(scope, 'overfeat', [inputs]) as sc:
end_points_collection = sc.name + '_end_points'
# Collect outputs for conv2d, fully_connected and max_pool2d
with arg_scope(
[layers.conv2d, layers_lib.fully_connected, layers_lib.max_pool2d],
outputs_collections=end_points_collection):
net = layers.conv2d(
inputs, 64, [11, 11], 4, padding='VALID', scope='conv1')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool1')
net = layers.conv2d(net, 256, [5, 5], padding='VALID', scope='conv2')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool2')
net = layers.conv2d(net, 512, [3, 3], scope='conv3')
net = layers.conv2d(net, 1024, [3, 3], scope='conv4')
net = layers.conv2d(net, 1024, [3, 3], scope='conv5')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool5')
with arg_scope(
[layers.conv2d],
weights_initializer=trunc_normal(0.005),
biases_initializer=init_ops.constant_initializer(0.1)):
# Use conv2d instead of fully_connected layers.
net = layers.conv2d(net, 3072, [6, 6], padding='VALID', scope='fc6')
net = layers_lib.dropout(
net, dropout_keep_prob, is_training=is_training, scope='dropout6')
net = layers.conv2d(net, 4096, [1, 1], scope='fc7')
net = layers_lib.dropout(
net, dropout_keep_prob, is_training=is_training, scope='dropout7')
net = layers.conv2d(
net,
num_classes, [1, 1],
activation_fn=None,
normalizer_fn=None,
biases_initializer=init_ops.zeros_initializer(),
scope='fc8')
# Convert end_points_collection into a end_point dict.
end_points = utils.convert_collection_to_dict(end_points_collection)
if spatial_squeeze:
net = array_ops.squeeze(net, [1, 2], name='fc8/squeezed')
end_points[sc.name + '/fc8'] = net
return net, end_points
