Merged commit includes the following changes: (#7797)

279978375  by Sergio Guadarrama:

    Pass s=2 to the expanded_conv block so it can apply residual correctly in case of fused convolutions.  (Before it was relying on channel mismatch only)

--
279788358  by Sergio Guadarrama:

    Update README to add mobilenet-edgetpu details

--
279774392  by Sergio Guadarrama:

    Adds MobilenetV3-EdgeTpu definition.

--
278917344  by Sergio Guadarrama:

    Create visualwakewords dataset using slim scripts instead of custom scripts.

--
277940048  by Sergio Guadarrama:

    Internal changes to tf.contrib symbols

--

PiperOrigin-RevId: 279978375

research/slim/nets/mobilenet/mobilenet_v2.py

@@ -28,11 +28,13 @@ import copy
 import functools
 
 import tensorflow as tf
+from tensorflow.contrib import layers as contrib_layers
+from tensorflow.contrib import slim as contrib_slim
 
 from nets.mobilenet import conv_blocks as ops
 from nets.mobilenet import mobilenet as lib
 
-slim = tf.contrib.slim
+slim = contrib_slim
 op = lib.op
 
 expand_input = ops.expand_input_by_factor
@@ -84,18 +86,18 @@ V2_DEF = dict(
 # Mobilenet v2 Definition with group normalization.
 V2_DEF_GROUP_NORM = copy.deepcopy(V2_DEF)
 V2_DEF_GROUP_NORM['defaults'] = {
-    (tf.contrib.slim.conv2d, tf.contrib.slim.fully_connected,
-     tf.contrib.slim.separable_conv2d): {
-        'normalizer_fn': tf.contrib.layers.group_norm,  # pylint: disable=C0330
+    (contrib_slim.conv2d, contrib_slim.fully_connected,
+     contrib_slim.separable_conv2d): {
+        'normalizer_fn': contrib_layers.group_norm,  # pylint: disable=C0330
         'activation_fn': tf.nn.relu6,  # pylint: disable=C0330
     },  # pylint: disable=C0330
     (ops.expanded_conv,): {
         'expansion_size': ops.expand_input_by_factor(6),
         'split_expansion': 1,
-        'normalizer_fn': tf.contrib.layers.group_norm,
+        'normalizer_fn': contrib_layers.group_norm,
         'residual': True
     },
-    (tf.contrib.slim.conv2d, tf.contrib.slim.separable_conv2d): {
+    (contrib_slim.conv2d, contrib_slim.separable_conv2d): {
         'padding': 'SAME'
     }
 }
@@ -213,7 +215,7 @@ def mobilenet_base_group_norm(input_tensor, depth_multiplier=1.0, **kwargs):
   """Creates base of the mobilenet (no pooling and no logits) ."""
   kwargs['conv_defs'] = V2_DEF_GROUP_NORM
   kwargs['conv_defs']['defaults'].update({
-      (tf.contrib.layers.group_norm,): {
+      (contrib_layers.group_norm,): {
           'groups': kwargs.pop('groups', 8)
       }
   })

research/slim/nets/mobilenet/mobilenet_v2_test.py

@@ -19,12 +19,13 @@ from __future__ import division
 from __future__ import print_function
 import copy
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 from nets.mobilenet import conv_blocks as ops
 from nets.mobilenet import mobilenet
 from nets.mobilenet import mobilenet_v2
 
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 def find_ops(optype):

research/slim/nets/mobilenet/mobilenet_v3.py

@@ -23,11 +23,12 @@ import functools
 import numpy as np
 
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from nets.mobilenet import conv_blocks as ops
 from nets.mobilenet import mobilenet as lib
 
-slim = tf.contrib.slim
+slim = contrib_slim
 op = lib.op
 expand_input = ops.expand_input_by_factor
 
@@ -57,7 +58,7 @@ def reduce_to_1x1(input_tensor, default_size=7, **kwargs):
   return slim.avg_pool2d(input_tensor, kernel_size=k, **kwargs)
 
 
-def mbv3_op(ef, n, k, s=1, act=tf.nn.relu, se=None):
+def mbv3_op(ef, n, k, s=1, act=tf.nn.relu, se=None, **kwargs):
   """Defines a single Mobilenet V3 convolution block.
 
   Args:
@@ -67,14 +68,44 @@ def mbv3_op(ef, n, k, s=1, act=tf.nn.relu, se=None):
     s: stride
     act: activation function in inner layers
     se: squeeze excite function.
+    **kwargs: passed to expanded_conv
 
   Returns:
     An object (lib._Op) for inserting in conv_def, representing this operation.
   """
-  return op(ops.expanded_conv, expansion_size=expand_input(ef),
-            kernel_size=(k, k), stride=s, num_outputs=n,
-            inner_activation_fn=act,
-            expansion_transform=se)
+  return op(
+      ops.expanded_conv,
+      expansion_size=expand_input(ef),
+      kernel_size=(k, k),
+      stride=s,
+      num_outputs=n,
+      inner_activation_fn=act,
+      expansion_transform=se,
+      **kwargs)
+
+
+def mbv3_fused(ef, n, k, s=1, **kwargs):
+  """Defines a single Mobilenet V3 convolution block.
+
+  Args:
+    ef: expansion factor
+    n: number of output channels
+    k: stride of depthwise
+    s: stride
+    **kwargs: will be passed to mbv3_op
+
+  Returns:
+    An object (lib._Op) for inserting in conv_def, representing this operation.
+  """
+  expansion_fn = functools.partial(slim.conv2d, kernel_size=k, stride=s)
+  return mbv3_op(
+      ef,
+      n,
+      k=1,
+      s=s,
+      depthwise_location=None,
+      expansion_fn=expansion_fn,
+      **kwargs)
 
 
 mbv3_op_se = functools.partial(mbv3_op, se=_se4)
@@ -206,6 +237,38 @@ V3_SMALL_MINIMALISTIC = dict(
     ]))
 
 
+# EdgeTPU friendly variant of MobilenetV3 that uses fused convolutions
+# instead of depthwise in the early layers.
+V3_EDGETPU = dict(
+    defaults=dict(DEFAULTS),
+    spec=[
+        op(slim.conv2d, stride=2, num_outputs=32, kernel_size=(3, 3)),
+        mbv3_fused(k=3, s=1, ef=1, n=16),
+        mbv3_fused(k=3, s=2, ef=8, n=32),
+        mbv3_fused(k=3, s=1, ef=4, n=32),
+        mbv3_fused(k=3, s=1, ef=4, n=32),
+        mbv3_fused(k=3, s=1, ef=4, n=32),
+        mbv3_fused(k=3, s=2, ef=8, n=48),
+        mbv3_fused(k=3, s=1, ef=4, n=48),
+        mbv3_fused(k=3, s=1, ef=4, n=48),
+        mbv3_fused(k=3, s=1, ef=4, n=48),
+        mbv3_op(k=3, s=2, ef=8, n=96),
+        mbv3_op(k=3, s=1, ef=4, n=96),
+        mbv3_op(k=3, s=1, ef=4, n=96),
+        mbv3_op(k=3, s=1, ef=4, n=96),
+        mbv3_op(k=3, s=1, ef=8, n=96, residual=False),
+        mbv3_op(k=3, s=1, ef=4, n=96),
+        mbv3_op(k=3, s=1, ef=4, n=96),
+        mbv3_op(k=3, s=1, ef=4, n=96),
+        mbv3_op(k=5, s=2, ef=8, n=160),
+        mbv3_op(k=5, s=1, ef=4, n=160),
+        mbv3_op(k=5, s=1, ef=4, n=160),
+        mbv3_op(k=5, s=1, ef=4, n=160),
+        mbv3_op(k=3, s=1, ef=8, n=192),
+        op(slim.conv2d, stride=1, num_outputs=1280, kernel_size=(1, 1)),
+    ])
+
+
 @slim.add_arg_scope
 def mobilenet(input_tensor,
               num_classes=1001,
@@ -275,15 +338,26 @@ def mobilenet_base(input_tensor, depth_multiplier=1.0, **kwargs):
       input_tensor, depth_multiplier=depth_multiplier, base_only=True, **kwargs)
 
 
-def wrapped_partial(func, *args, **kwargs):
-  partial_func = functools.partial(func, *args, **kwargs)
+def wrapped_partial(func, new_defaults=None,
+                    **kwargs):
+  """Partial function with new default parameters and updated docstring."""
+  if not new_defaults:
+    new_defaults = {}
+  def func_wrapper(*f_args, **f_kwargs):
+    new_kwargs = dict(new_defaults)
+    new_kwargs.update(f_kwargs)
+    return func(*f_args, **new_kwargs)
+  functools.update_wrapper(func_wrapper, func)
+  partial_func = functools.partial(func_wrapper, **kwargs)
   functools.update_wrapper(partial_func, func)
   return partial_func
 
 
 large = wrapped_partial(mobilenet, conv_defs=V3_LARGE)
 small = wrapped_partial(mobilenet, conv_defs=V3_SMALL)
-
+edge_tpu = wrapped_partial(mobilenet,
+                           new_defaults={'scope': 'MobilenetEdgeTPU'},
+                           conv_defs=V3_EDGETPU)
 
 # Minimalistic model that does not have Squeeze Excite blocks,
 # Hardswish, or 5x5 depthwise convolution.

research/slim/nets/mobilenet/mobilenet_v3_test.py

@@ -42,6 +42,21 @@ class MobilenetV3Test(absltest.TestCase):
         conv_defs=mobilenet_v3.V3_SMALL)
     self.assertEqual(endpoints['layer_15'].shape, [1, 1, 1, 1024])
 
+  def testMobilenetEdgeTpu(self):
+    _, endpoints = mobilenet_v3.edge_tpu(
+        tf.placeholder(tf.float32, (1, 224, 224, 3)))
+    self.assertIn('Inference mode is created by default',
+                  mobilenet_v3.edge_tpu.__doc__)
+    self.assertEqual(endpoints['layer_24'].shape, [1, 7, 7, 1280])
+    self.assertStartsWith(
+        endpoints['layer_24'].name, 'MobilenetEdgeTPU')
+
+  def testMobilenetEdgeTpuChangeScope(self):
+    _, endpoints = mobilenet_v3.edge_tpu(
+        tf.placeholder(tf.float32, (1, 224, 224, 3)), scope='Scope')
+    self.assertStartsWith(
+        endpoints['layer_24'].name, 'Scope')
+
   def testMobilenetV3BaseOnly(self):
     result, endpoints = mobilenet_v3.mobilenet(
         tf.placeholder(tf.float32, (1, 224, 224, 3)),

research/slim/nets/mobilenet_v1.py

@@ -109,8 +109,10 @@ from collections import namedtuple
 import functools
 
 import tensorflow as tf
+from tensorflow.contrib import layers as contrib_layers
+from tensorflow.contrib import slim as contrib_slim
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 # Conv and DepthSepConv namedtuple define layers of the MobileNet architecture
 # Conv defines 3x3 convolution layers
@@ -307,7 +309,7 @@ def mobilenet_v1(inputs,
                  min_depth=8,
                  depth_multiplier=1.0,
                  conv_defs=None,
-                 prediction_fn=tf.contrib.layers.softmax,
+                 prediction_fn=contrib_layers.softmax,
                  spatial_squeeze=True,
                  reuse=None,
                  scope='MobilenetV1',
@@ -461,7 +463,7 @@ def mobilenet_v1_arg_scope(
 
   # Set weight_decay for weights in Conv and DepthSepConv layers.
   weights_init = tf.truncated_normal_initializer(stddev=stddev)
-  regularizer = tf.contrib.layers.l2_regularizer(weight_decay)
+  regularizer = contrib_layers.l2_regularizer(weight_decay)
   if regularize_depthwise:
     depthwise_regularizer = regularizer
   else:

research/slim/nets/mobilenet_v1_eval.py

@@ -20,12 +20,14 @@ from __future__ import print_function
 
 import math
 import tensorflow as tf
+from tensorflow.contrib import quantize as contrib_quantize
+from tensorflow.contrib import slim as contrib_slim
 
 from datasets import dataset_factory
 from nets import mobilenet_v1
 from preprocessing import preprocessing_factory
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 flags = tf.app.flags
 
@@ -124,7 +126,7 @@ def build_model():
           num_classes=FLAGS.num_classes)
 
     if FLAGS.quantize:
-      tf.contrib.quantize.create_eval_graph()
+      contrib_quantize.create_eval_graph()
 
     eval_ops = metrics(logits, labels)
 

research/slim/nets/mobilenet_v1_test.py

@@ -20,10 +20,11 @@ from __future__ import print_function
 
 import numpy as np
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from nets import mobilenet_v1
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class MobilenetV1Test(tf.test.TestCase):

research/slim/nets/mobilenet_v1_train.py

@@ -19,12 +19,14 @@ from __future__ import division
 from __future__ import print_function
 
 import tensorflow as tf
+from tensorflow.contrib import quantize as contrib_quantize
+from tensorflow.contrib import slim as contrib_slim
 
 from datasets import dataset_factory
 from nets import mobilenet_v1
 from preprocessing import preprocessing_factory
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 flags = tf.app.flags
 
@@ -136,7 +138,7 @@ def build_model():
     # quant_delay delays start of quantization till quant_delay steps, allowing
     # for better model accuracy.
     if FLAGS.quantize:
-      tf.contrib.quantize.create_training_graph(quant_delay=get_quant_delay())
+      contrib_quantize.create_training_graph(quant_delay=get_quant_delay())
 
     total_loss = tf.losses.get_total_loss(name='total_loss')
     # Configure the learning rate using an exponential decay.

research/slim/nets/nasnet/nasnet.py

@@ -22,11 +22,15 @@ from __future__ import print_function
 
 import copy
 import tensorflow as tf
+from tensorflow.contrib import framework as contrib_framework
+from tensorflow.contrib import layers as contrib_layers
+from tensorflow.contrib import slim as contrib_slim
+from tensorflow.contrib import training as contrib_training
 
 from nets.nasnet import nasnet_utils
 
-arg_scope = tf.contrib.framework.arg_scope
-slim = tf.contrib.slim
+arg_scope = contrib_framework.arg_scope
+slim = contrib_slim
 
 
 # Notes for training NASNet Cifar Model
@@ -37,7 +41,7 @@ slim = tf.contrib.slim
 # auxiliary head loss weighting: 0.4
 # clip global norm of all gradients by 5
 def cifar_config():
-  return tf.contrib.training.HParams(
+  return contrib_training.HParams(
       stem_multiplier=3.0,
       drop_path_keep_prob=0.6,
       num_cells=18,
@@ -67,7 +71,7 @@ def cifar_config():
 # label smoothing: 0.1
 # clip global norm of all gradients by 10
 def large_imagenet_config():
-  return tf.contrib.training.HParams(
+  return contrib_training.HParams(
       stem_multiplier=3.0,
       dense_dropout_keep_prob=0.5,
       num_cells=18,
@@ -94,7 +98,7 @@ def large_imagenet_config():
 # label smoothing: 0.1
 # clip global norm of all gradients by 10
 def mobile_imagenet_config():
-  return tf.contrib.training.HParams(
+  return contrib_training.HParams(
       stem_multiplier=1.0,
       dense_dropout_keep_prob=0.5,
       num_cells=12,
@@ -138,8 +142,8 @@ def nasnet_cifar_arg_scope(weight_decay=5e-4,
       'scale': True,
       'fused': True,
   }
-  weights_regularizer = tf.contrib.layers.l2_regularizer(weight_decay)
-  weights_initializer = tf.contrib.layers.variance_scaling_initializer(
+  weights_regularizer = contrib_layers.l2_regularizer(weight_decay)
+  weights_initializer = contrib_layers.variance_scaling_initializer(
       mode='FAN_OUT')
   with arg_scope([slim.fully_connected, slim.conv2d, slim.separable_conv2d],
                  weights_regularizer=weights_regularizer,
@@ -174,8 +178,8 @@ def nasnet_mobile_arg_scope(weight_decay=4e-5,
       'scale': True,
       'fused': True,
   }
-  weights_regularizer = tf.contrib.layers.l2_regularizer(weight_decay)
-  weights_initializer = tf.contrib.layers.variance_scaling_initializer(
+  weights_regularizer = contrib_layers.l2_regularizer(weight_decay)
+  weights_initializer = contrib_layers.variance_scaling_initializer(
       mode='FAN_OUT')
   with arg_scope([slim.fully_connected, slim.conv2d, slim.separable_conv2d],
                  weights_regularizer=weights_regularizer,
@@ -210,8 +214,8 @@ def nasnet_large_arg_scope(weight_decay=5e-5,
       'scale': True,
       'fused': True,
   }
-  weights_regularizer = tf.contrib.layers.l2_regularizer(weight_decay)
-  weights_initializer = tf.contrib.layers.variance_scaling_initializer(
+  weights_regularizer = contrib_layers.l2_regularizer(weight_decay)
+  weights_initializer = contrib_layers.variance_scaling_initializer(
       mode='FAN_OUT')
   with arg_scope([slim.fully_connected, slim.conv2d, slim.separable_conv2d],
                  weights_regularizer=weights_regularizer,
@@ -244,7 +248,7 @@ def _build_aux_head(net, end_points, num_classes, hparams, scope):
       aux_logits = slim.conv2d(aux_logits, 768, shape, padding='VALID')
       aux_logits = slim.batch_norm(aux_logits, scope='aux_bn1')
       aux_logits = activation_fn(aux_logits)
-      aux_logits = tf.contrib.layers.flatten(aux_logits)
+      aux_logits = contrib_layers.flatten(aux_logits)
       aux_logits = slim.fully_connected(aux_logits, num_classes)
       end_points['AuxLogits'] = aux_logits
 

research/slim/nets/nasnet/nasnet_test.py

@@ -18,10 +18,11 @@ from __future__ import division
 from __future__ import print_function
 
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from nets.nasnet import nasnet
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class NASNetTest(tf.test.TestCase):

research/slim/nets/nasnet/nasnet_utils.py

@@ -32,10 +32,11 @@ from __future__ import division
 from __future__ import print_function
 
 import tensorflow as tf
+from tensorflow.contrib import framework as contrib_framework
+from tensorflow.contrib import slim as contrib_slim
 
-
-arg_scope = tf.contrib.framework.arg_scope
-slim = tf.contrib.slim
+arg_scope = contrib_framework.arg_scope
+slim = contrib_slim
 
 DATA_FORMAT_NCHW = 'NCHW'
 DATA_FORMAT_NHWC = 'NHWC'
@@ -55,14 +56,14 @@ def calc_reduction_layers(num_cells, num_reduction_layers):
   return reduction_layers
 
 
-@tf.contrib.framework.add_arg_scope
+@contrib_framework.add_arg_scope
 def get_channel_index(data_format=INVALID):
   assert data_format != INVALID
   axis = 3 if data_format == 'NHWC' else 1
   return axis
 
 
-@tf.contrib.framework.add_arg_scope
+@contrib_framework.add_arg_scope
 def get_channel_dim(shape, data_format=INVALID):
   assert data_format != INVALID
   assert len(shape) == 4
@@ -74,7 +75,7 @@ def get_channel_dim(shape, data_format=INVALID):
     raise ValueError('Not a valid data_format', data_format)
 
 
-@tf.contrib.framework.add_arg_scope
+@contrib_framework.add_arg_scope
 def global_avg_pool(x, data_format=INVALID):
   """Average pool away the height and width spatial dimensions of x."""
   assert data_format != INVALID
@@ -86,7 +87,7 @@ def global_avg_pool(x, data_format=INVALID):
     return tf.reduce_mean(x, [2, 3])
 
 
-@tf.contrib.framework.add_arg_scope
+@contrib_framework.add_arg_scope
 def factorized_reduction(net, output_filters, stride, data_format=INVALID):
   """Reduces the shape of net without information loss due to striding."""
   assert data_format != INVALID
@@ -129,7 +130,7 @@ def factorized_reduction(net, output_filters, stride, data_format=INVALID):
   return final_path
 
 
-@tf.contrib.framework.add_arg_scope
+@contrib_framework.add_arg_scope
 def drop_path(net, keep_prob, is_training=True):
   """Drops out a whole example hiddenstate with the specified probability."""
   if is_training:
@@ -422,7 +423,7 @@ class NasNetABaseCell(object):
     net = tf.concat(values=states_to_combine, axis=concat_axis)
     return net
 
-  @tf.contrib.framework.add_arg_scope  # No public API. For internal use only.
+  @contrib_framework.add_arg_scope  # No public API. For internal use only.
   def _apply_drop_path(self, net, current_step=None,
                        use_summaries=False, drop_connect_version='v3'):
     """Apply drop_path regularization.

research/slim/nets/nasnet/pnasnet.py

@@ -23,17 +23,20 @@ from __future__ import print_function
 
 import copy
 import tensorflow as tf
+from tensorflow.contrib import framework as contrib_framework
+from tensorflow.contrib import slim as contrib_slim
+from tensorflow.contrib import training as contrib_training
 
 from nets.nasnet import nasnet
 from nets.nasnet import nasnet_utils
 
-arg_scope = tf.contrib.framework.arg_scope
-slim = tf.contrib.slim
+arg_scope = contrib_framework.arg_scope
+slim = contrib_slim
 
 
 def large_imagenet_config():
   """Large ImageNet configuration based on PNASNet-5."""
-  return tf.contrib.training.HParams(
+  return contrib_training.HParams(
       stem_multiplier=3.0,
       dense_dropout_keep_prob=0.5,
       num_cells=12,
@@ -51,7 +54,7 @@ def large_imagenet_config():
 
 def mobile_imagenet_config():
   """Mobile ImageNet configuration based on PNASNet-5."""
-  return tf.contrib.training.HParams(
+  return contrib_training.HParams(
       stem_multiplier=1.0,
       dense_dropout_keep_prob=0.5,
       num_cells=9,

research/slim/nets/nasnet/pnasnet_test.py

@@ -18,10 +18,11 @@ from __future__ import division
 from __future__ import print_function
 
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from nets.nasnet import pnasnet
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class PNASNetTest(tf.test.TestCase):

research/slim/nets/nets_factory.py

@@ -18,8 +18,7 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 import functools
-
-import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from nets import alexnet
 from nets import cifarnet
@@ -37,7 +36,7 @@ from nets.nasnet import nasnet
 from nets.nasnet import pnasnet
 
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 networks_map = {'alexnet_v2': alexnet.alexnet_v2,
                 'cifarnet': cifarnet.cifarnet,

research/slim/nets/overfeat.py

@@ -32,8 +32,9 @@ from __future__ import division
 from __future__ import print_function
 
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
-slim = tf.contrib.slim
+slim = contrib_slim
 trunc_normal = lambda stddev: tf.truncated_normal_initializer(0.0, stddev)
 
 

research/slim/nets/overfeat_test.py

@@ -18,10 +18,11 @@ from __future__ import division
 from __future__ import print_function
 
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from nets import overfeat
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class OverFeatTest(tf.test.TestCase):

research/slim/nets/pix2pix.py

@@ -33,8 +33,10 @@ import collections
 import functools
 
 import tensorflow as tf
+from tensorflow.contrib import framework as contrib_framework
+from tensorflow.contrib import layers as contrib_layers
 
-layers = tf.contrib.layers
+layers = contrib_layers
 
 
 def pix2pix_arg_scope():
@@ -52,7 +54,7 @@ def pix2pix_arg_scope():
       'epsilon': 0.00001,
   }
 
-  with tf.contrib.framework.arg_scope(
+  with contrib_framework.arg_scope(
       [layers.conv2d, layers.conv2d_transpose],
       normalizer_fn=layers.instance_norm,
       normalizer_params=instance_norm_params,
@@ -165,11 +167,10 @@ def pix2pix_generator(net,
   # Encoder #
   ###########
   with tf.variable_scope('encoder'):
-    with tf.contrib.framework.arg_scope(
-        [layers.conv2d],
-        kernel_size=[4, 4],
-        stride=2,
-        activation_fn=tf.nn.leaky_relu):
+    with contrib_framework.arg_scope([layers.conv2d],
+                                     kernel_size=[4, 4],
+                                     stride=2,
+                                     activation_fn=tf.nn.leaky_relu):
 
       for block_id, block in enumerate(blocks):
         # No normalizer for the first encoder layers as per 'Image-to-Image',
@@ -196,7 +197,7 @@ def pix2pix_generator(net,
   with tf.variable_scope('decoder'):
     # Dropout is used at both train and test time as per 'Image-to-Image',
     # Section 2.1 (last paragraph).
-    with tf.contrib.framework.arg_scope([layers.dropout], is_training=True):
+    with contrib_framework.arg_scope([layers.dropout], is_training=True):
 
       for block_id, block in enumerate(reversed_blocks):
         if block_id > 0:
@@ -256,12 +257,11 @@ def pix2pix_discriminator(net, num_filters, padding=2, pad_mode='REFLECT',
     else:
       return net
 
-  with tf.contrib.framework.arg_scope(
-      [layers.conv2d],
-      kernel_size=[4, 4],
-      stride=2,
-      padding='valid',
-      activation_fn=activation_fn):
+  with contrib_framework.arg_scope([layers.conv2d],
+                                   kernel_size=[4, 4],
+                                   stride=2,
+                                   padding='valid',
+                                   activation_fn=activation_fn):
 
     # No normalization on the input layer.
     net = layers.conv2d(

research/slim/nets/pix2pix_test.py

@@ -19,6 +19,7 @@ from __future__ import division
 from __future__ import print_function
 
 import tensorflow as tf
+from tensorflow.contrib import framework as contrib_framework
 from nets import pix2pix
 
 
@@ -35,7 +36,7 @@ class GeneratorTest(tf.test.TestCase):
     num_outputs = 4
 
     images = tf.ones((batch_size, height, width, 3))
-    with tf.contrib.framework.arg_scope(pix2pix.pix2pix_arg_scope()):
+    with contrib_framework.arg_scope(pix2pix.pix2pix_arg_scope()):
       logits, _ = pix2pix.pix2pix_generator(
           images, num_outputs, blocks=self._reduced_default_blocks(),
           upsample_method='nn_upsample_conv')
@@ -52,7 +53,7 @@ class GeneratorTest(tf.test.TestCase):
     num_outputs = 4
 
     images = tf.ones((batch_size, height, width, 3))
-    with tf.contrib.framework.arg_scope(pix2pix.pix2pix_arg_scope()):
+    with contrib_framework.arg_scope(pix2pix.pix2pix_arg_scope()):
       logits, _ = pix2pix.pix2pix_generator(
           images, num_outputs, blocks=self._reduced_default_blocks(),
           upsample_method='conv2d_transpose')
@@ -73,7 +74,7 @@ class GeneratorTest(tf.test.TestCase):
         pix2pix.Block(64, 0.5),
         pix2pix.Block(128, 0),
     ]
-    with tf.contrib.framework.arg_scope(pix2pix.pix2pix_arg_scope()):
+    with contrib_framework.arg_scope(pix2pix.pix2pix_arg_scope()):
       _, end_points = pix2pix.pix2pix_generator(
           images, num_outputs, blocks)
 
@@ -105,7 +106,7 @@ class DiscriminatorTest(tf.test.TestCase):
     output_size = self._layer_output_size(output_size, stride=1)
 
     images = tf.ones((batch_size, input_size, input_size, 3))
-    with tf.contrib.framework.arg_scope(pix2pix.pix2pix_arg_scope()):
+    with contrib_framework.arg_scope(pix2pix.pix2pix_arg_scope()):
       logits, end_points = pix2pix.pix2pix_discriminator(
           images, num_filters=[64, 128, 256, 512])
     self.assertListEqual([batch_size, output_size, output_size, 1],
@@ -124,7 +125,7 @@ class DiscriminatorTest(tf.test.TestCase):
     output_size = self._layer_output_size(output_size, stride=1, pad=0)
 
     images = tf.ones((batch_size, input_size, input_size, 3))
-    with tf.contrib.framework.arg_scope(pix2pix.pix2pix_arg_scope()):
+    with contrib_framework.arg_scope(pix2pix.pix2pix_arg_scope()):
       logits, end_points = pix2pix.pix2pix_discriminator(
           images, num_filters=[64, 128, 256, 512], padding=0)
     self.assertListEqual([batch_size, output_size, output_size, 1],
@@ -137,7 +138,7 @@ class DiscriminatorTest(tf.test.TestCase):
     input_size = 256
 
     images = tf.ones((batch_size, input_size, input_size, 3))
-    with tf.contrib.framework.arg_scope(pix2pix.pix2pix_arg_scope()):
+    with contrib_framework.arg_scope(pix2pix.pix2pix_arg_scope()):
       with self.assertRaises(TypeError):
         pix2pix.pix2pix_discriminator(
             images, num_filters=[64, 128, 256, 512], padding=1.5)
@@ -147,7 +148,7 @@ class DiscriminatorTest(tf.test.TestCase):
     input_size = 256
 
     images = tf.ones((batch_size, input_size, input_size, 3))
-    with tf.contrib.framework.arg_scope(pix2pix.pix2pix_arg_scope()):
+    with contrib_framework.arg_scope(pix2pix.pix2pix_arg_scope()):
       with self.assertRaises(ValueError):
         pix2pix.pix2pix_discriminator(
             images, num_filters=[64, 128, 256, 512], padding=-1)

research/slim/nets/resnet_utils.py

@@ -39,8 +39,9 @@ from __future__ import print_function
 
 import collections
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class Block(collections.namedtuple('Block', ['scope', 'unit_fn', 'args'])):

research/slim/nets/resnet_v1.py

@@ -57,12 +57,13 @@ from __future__ import division
 from __future__ import print_function
 
 import tensorflow as tf
+from tensorflow.contrib import slim as contrib_slim
 
 from nets import resnet_utils
 
 
 resnet_arg_scope = resnet_utils.resnet_arg_scope
-slim = tf.contrib.slim
+slim = contrib_slim
 
 
 class NoOpScope(object):