Merged commit includes the following changes: (#8077)

Internal cleanup (py2->py3) plus the following changes:

285513318  by Sergio Guadarrama:

    Adds a script for post-training quantization

284222305  by Sergio Guadarrama:

    Modified squeeze-excite operation to accommodate tensors of undefined (Nonetype) H/W.

282028343  by Sergio Guadarrama:

    Add MobilenetV3 and MobilenetEdgeTPU to the slim/nets_factory.

PiperOrigin-RevId: 289455329
Co-authored-by: Sergio Guadarrama <sguada@gmail.com>

research/slim/nets/nets_factory.py

@@ -32,13 +32,15 @@ from nets import resnet_v2
 from nets import s3dg
 from nets import vgg
 from nets.mobilenet import mobilenet_v2
+from nets.mobilenet import mobilenet_v3
 from nets.nasnet import nasnet
 from nets.nasnet import pnasnet
 
 
 slim = contrib_slim
 
-networks_map = {'alexnet_v2': alexnet.alexnet_v2,
+networks_map = {
+    'alexnet_v2': alexnet.alexnet_v2,
     'cifarnet': cifarnet.cifarnet,
     'overfeat': overfeat.overfeat,
     'vgg_a': vgg.vgg_a,
@@ -67,14 +69,21 @@ networks_map = {'alexnet_v2': alexnet.alexnet_v2,
     'mobilenet_v2': mobilenet_v2.mobilenet,
     'mobilenet_v2_140': mobilenet_v2.mobilenet_v2_140,
     'mobilenet_v2_035': mobilenet_v2.mobilenet_v2_035,
+    'mobilenet_v3_small': mobilenet_v3.small,
+    'mobilenet_v3_large': mobilenet_v3.large,
+    'mobilenet_v3_small_minimalistic': mobilenet_v3.small_minimalistic,
+    'mobilenet_v3_large_minimalistic': mobilenet_v3.large_minimalistic,
+    'mobilenet_edgetpu': mobilenet_v3.edge_tpu,
+    'mobilenet_edgetpu_075': mobilenet_v3.edge_tpu_075,
     'nasnet_cifar': nasnet.build_nasnet_cifar,
     'nasnet_mobile': nasnet.build_nasnet_mobile,
     'nasnet_large': nasnet.build_nasnet_large,
     'pnasnet_large': pnasnet.build_pnasnet_large,
     'pnasnet_mobile': pnasnet.build_pnasnet_mobile,
-               }
+}
 
-arg_scopes_map = {'alexnet_v2': alexnet.alexnet_v2_arg_scope,
+arg_scopes_map = {
+    'alexnet_v2': alexnet.alexnet_v2_arg_scope,
     'cifarnet': cifarnet.cifarnet_arg_scope,
     'overfeat': overfeat.overfeat_arg_scope,
     'vgg_a': vgg.vgg_arg_scope,
@@ -84,8 +93,7 @@ arg_scopes_map = {'alexnet_v2': alexnet.alexnet_v2_arg_scope,
     'inception_v2': inception.inception_v3_arg_scope,
     'inception_v3': inception.inception_v3_arg_scope,
     'inception_v4': inception.inception_v4_arg_scope,
-                  'inception_resnet_v2':
-                  inception.inception_resnet_v2_arg_scope,
+    'inception_resnet_v2': inception.inception_resnet_v2_arg_scope,
     'i3d': i3d.i3d_arg_scope,
     's3dg': s3dg.s3dg_arg_scope,
     'lenet': lenet.lenet_arg_scope,
@@ -104,12 +112,18 @@ arg_scopes_map = {'alexnet_v2': alexnet.alexnet_v2_arg_scope,
     'mobilenet_v2': mobilenet_v2.training_scope,
     'mobilenet_v2_035': mobilenet_v2.training_scope,
     'mobilenet_v2_140': mobilenet_v2.training_scope,
+    'mobilenet_v3_small': mobilenet_v3.training_scope,
+    'mobilenet_v3_large': mobilenet_v3.training_scope,
+    'mobilenet_v3_small_minimalistic': mobilenet_v3.training_scope,
+    'mobilenet_v3_large_minimalistic': mobilenet_v3.training_scope,
+    'mobilenet_edgetpu': mobilenet_v3.training_scope,
+    'mobilenet_edgetpu_075': mobilenet_v3.training_scope,
     'nasnet_cifar': nasnet.nasnet_cifar_arg_scope,
     'nasnet_mobile': nasnet.nasnet_mobile_arg_scope,
     'nasnet_large': nasnet.nasnet_large_arg_scope,
     'pnasnet_large': pnasnet.pnasnet_large_arg_scope,
     'pnasnet_mobile': pnasnet.pnasnet_mobile_arg_scope,
-                 }
+}
 
 
 def get_network_fn(name, num_classes, weight_decay=0.0, is_training=False):

research/slim/nets/nets_factory_test.py

@@ -36,8 +36,7 @@ class NetworksTest(tf.test.TestCase):
         # Most networks use 224 as their default_image_size
         image_size = getattr(net_fn, 'default_image_size', 224)
         if net not in ['i3d', 's3dg']:
-          inputs = tf.random_uniform(
-              (batch_size, image_size, image_size, 3))
+          inputs = tf.random.uniform((batch_size, image_size, image_size, 3))
           logits, end_points = net_fn(inputs)
           self.assertTrue(isinstance(logits, tf.Tensor))
           self.assertTrue(isinstance(end_points, dict))
@@ -53,8 +52,7 @@ class NetworksTest(tf.test.TestCase):
         # Most networks use 224 as their default_image_size
         image_size = getattr(net_fn, 'default_image_size', 224)
         if net not in ['i3d', 's3dg']:
-          inputs = tf.random_uniform(
-              (batch_size, image_size, image_size, 3))
+          inputs = tf.random.uniform((batch_size, image_size, image_size, 3))
           logits, end_points = net_fn(inputs)
           self.assertTrue(isinstance(logits, tf.Tensor))
           self.assertTrue(isinstance(end_points, dict))
@@ -69,8 +67,7 @@ class NetworksTest(tf.test.TestCase):
         net_fn = nets_factory.get_network_fn(net, num_classes=num_classes)
         # Most networks use 224 as their default_image_size
         image_size = getattr(net_fn, 'default_image_size', 224) // 2
-        inputs = tf.random_uniform(
-            (batch_size, 10, image_size, image_size, 3))
+        inputs = tf.random.uniform((batch_size, 10, image_size, image_size, 3))
         logits, end_points = net_fn(inputs)
         self.assertTrue(isinstance(logits, tf.Tensor))
         self.assertTrue(isinstance(end_points, dict))

research/slim/nets/overfeat.py

@@ -35,14 +35,17 @@ import tensorflow as tf
 from tensorflow.contrib import slim as contrib_slim
 
 slim = contrib_slim
-trunc_normal = lambda stddev: tf.truncated_normal_initializer(0.0, stddev)
+
+# pylint: disable=g-long-lambda
+trunc_normal = lambda stddev: tf.compat.v1.truncated_normal_initializer(
+    0.0, stddev)
 
 
 def overfeat_arg_scope(weight_decay=0.0005):
   with slim.arg_scope([slim.conv2d, slim.fully_connected],
                       activation_fn=tf.nn.relu,
                       weights_regularizer=slim.l2_regularizer(weight_decay),
-                      biases_initializer=tf.zeros_initializer()):
+                      biases_initializer=tf.compat.v1.zeros_initializer()):
     with slim.arg_scope([slim.conv2d], padding='SAME'):
       with slim.arg_scope([slim.max_pool2d], padding='VALID') as arg_sc:
         return arg_sc
@@ -88,7 +91,7 @@ def overfeat(inputs,
       None).
     end_points: a dict of tensors with intermediate activations.
   """
-  with tf.variable_scope(scope, 'overfeat', [inputs]) as sc:
+  with tf.compat.v1.variable_scope(scope, 'overfeat', [inputs]) as sc:
     end_points_collection = sc.original_name_scope + '_end_points'
     # Collect outputs for conv2d, fully_connected and max_pool2d
     with slim.arg_scope([slim.conv2d, slim.fully_connected, slim.max_pool2d],
@@ -104,9 +107,10 @@ def overfeat(inputs,
       net = slim.max_pool2d(net, [2, 2], scope='pool5')
 
       # Use conv2d instead of fully_connected layers.
-      with slim.arg_scope([slim.conv2d],
+      with slim.arg_scope(
+          [slim.conv2d],
           weights_initializer=trunc_normal(0.005),
-                          biases_initializer=tf.constant_initializer(0.1)):
+          biases_initializer=tf.compat.v1.constant_initializer(0.1)):
         net = slim.conv2d(net, 3072, [6, 6], padding='VALID', scope='fc6')
         net = slim.dropout(net, dropout_keep_prob, is_training=is_training,
                            scope='dropout6')
@@ -115,15 +119,18 @@ def overfeat(inputs,
         end_points = slim.utils.convert_collection_to_dict(
             end_points_collection)
         if global_pool:
-          net = tf.reduce_mean(net, [1, 2], keep_dims=True, name='global_pool')
+          net = tf.reduce_mean(
+              input_tensor=net, axis=[1, 2], keepdims=True, name='global_pool')
           end_points['global_pool'] = net
         if num_classes:
           net = slim.dropout(net, dropout_keep_prob, is_training=is_training,
                              scope='dropout7')
-          net = slim.conv2d(net, num_classes, [1, 1],
+          net = slim.conv2d(
+              net,
+              num_classes, [1, 1],
               activation_fn=None,
               normalizer_fn=None,
-                            biases_initializer=tf.zeros_initializer(),
+              biases_initializer=tf.compat.v1.zeros_initializer(),
               scope='fc8')
           if spatial_squeeze:
             net = tf.squeeze(net, [1, 2], name='fc8/squeezed')

research/slim/nets/overfeat_test.py

@@ -32,7 +32,7 @@ class OverFeatTest(tf.test.TestCase):
     height, width = 231, 231
     num_classes = 1000
     with self.test_session():
-      inputs = tf.random_uniform((batch_size, height, width, 3))
+      inputs = tf.random.uniform((batch_size, height, width, 3))
       logits, _ = overfeat.overfeat(inputs, num_classes)
       self.assertEquals(logits.op.name, 'overfeat/fc8/squeezed')
       self.assertListEqual(logits.get_shape().as_list(),
@@ -43,7 +43,7 @@ class OverFeatTest(tf.test.TestCase):
     height, width = 281, 281
     num_classes = 1000
     with self.test_session():
-      inputs = tf.random_uniform((batch_size, height, width, 3))
+      inputs = tf.random.uniform((batch_size, height, width, 3))
       logits, _ = overfeat.overfeat(inputs, num_classes, spatial_squeeze=False)
       self.assertEquals(logits.op.name, 'overfeat/fc8/BiasAdd')
       self.assertListEqual(logits.get_shape().as_list(),
@@ -54,7 +54,7 @@ class OverFeatTest(tf.test.TestCase):
     height, width = 281, 281
     num_classes = 1000
     with self.test_session():
-      inputs = tf.random_uniform((batch_size, height, width, 3))
+      inputs = tf.random.uniform((batch_size, height, width, 3))
       logits, _ = overfeat.overfeat(inputs, num_classes, spatial_squeeze=False,
                                     global_pool=True)
       self.assertEquals(logits.op.name, 'overfeat/fc8/BiasAdd')
@@ -66,7 +66,7 @@ class OverFeatTest(tf.test.TestCase):
     height, width = 231, 231
     num_classes = 1000
     with self.test_session():
-      inputs = tf.random_uniform((batch_size, height, width, 3))
+      inputs = tf.random.uniform((batch_size, height, width, 3))
       _, end_points = overfeat.overfeat(inputs, num_classes)
       expected_names = ['overfeat/conv1',
                         'overfeat/pool1',
@@ -87,7 +87,7 @@ class OverFeatTest(tf.test.TestCase):
     height, width = 231, 231
     num_classes = None
     with self.test_session():
-      inputs = tf.random_uniform((batch_size, height, width, 3))
+      inputs = tf.random.uniform((batch_size, height, width, 3))
       net, end_points = overfeat.overfeat(inputs, num_classes)
       expected_names = ['overfeat/conv1',
                         'overfeat/pool1',
@@ -108,7 +108,7 @@ class OverFeatTest(tf.test.TestCase):
     height, width = 231, 231
     num_classes = 1000
     with self.test_session():
-      inputs = tf.random_uniform((batch_size, height, width, 3))
+      inputs = tf.random.uniform((batch_size, height, width, 3))
       overfeat.overfeat(inputs, num_classes)
       expected_names = ['overfeat/conv1/weights',
                         'overfeat/conv1/biases',
@@ -135,11 +135,11 @@ class OverFeatTest(tf.test.TestCase):
     height, width = 231, 231
     num_classes = 1000
     with self.test_session():
-      eval_inputs = tf.random_uniform((batch_size, height, width, 3))
+      eval_inputs = tf.random.uniform((batch_size, height, width, 3))
       logits, _ = overfeat.overfeat(eval_inputs, is_training=False)
       self.assertListEqual(logits.get_shape().as_list(),
                            [batch_size, num_classes])
-      predictions = tf.argmax(logits, 1)
+      predictions = tf.argmax(input=logits, axis=1)
       self.assertListEqual(predictions.get_shape().as_list(), [batch_size])
 
   def testTrainEvalWithReuse(self):
@@ -149,29 +149,29 @@ class OverFeatTest(tf.test.TestCase):
     eval_height, eval_width = 281, 281
     num_classes = 1000
     with self.test_session():
-      train_inputs = tf.random_uniform(
+      train_inputs = tf.random.uniform(
           (train_batch_size, train_height, train_width, 3))
       logits, _ = overfeat.overfeat(train_inputs)
       self.assertListEqual(logits.get_shape().as_list(),
                            [train_batch_size, num_classes])
-      tf.get_variable_scope().reuse_variables()
-      eval_inputs = tf.random_uniform(
+      tf.compat.v1.get_variable_scope().reuse_variables()
+      eval_inputs = tf.random.uniform(
           (eval_batch_size, eval_height, eval_width, 3))
       logits, _ = overfeat.overfeat(eval_inputs, is_training=False,
                                     spatial_squeeze=False)
       self.assertListEqual(logits.get_shape().as_list(),
                            [eval_batch_size, 2, 2, num_classes])
-      logits = tf.reduce_mean(logits, [1, 2])
-      predictions = tf.argmax(logits, 1)
+      logits = tf.reduce_mean(input_tensor=logits, axis=[1, 2])
+      predictions = tf.argmax(input=logits, axis=1)
       self.assertEquals(predictions.get_shape().as_list(), [eval_batch_size])
 
   def testForward(self):
     batch_size = 1
     height, width = 231, 231
     with self.test_session() as sess:
-      inputs = tf.random_uniform((batch_size, height, width, 3))
+      inputs = tf.random.uniform((batch_size, height, width, 3))
       logits, _ = overfeat.overfeat(inputs)
-      sess.run(tf.global_variables_initializer())
+      sess.run(tf.compat.v1.global_variables_initializer())
       output = sess.run(logits)
       self.assertTrue(output.any())
 

research/slim/nets/pix2pix.py

@@ -58,7 +58,8 @@ def pix2pix_arg_scope():
       [layers.conv2d, layers.conv2d_transpose],
       normalizer_fn=layers.instance_norm,
       normalizer_params=instance_norm_params,
-      weights_initializer=tf.random_normal_initializer(0, 0.02)) as sc:
+      weights_initializer=tf.compat.v1.random_normal_initializer(0,
+                                                                 0.02)) as sc:
     return sc
 
 
@@ -80,13 +81,14 @@ def upsample(net, num_outputs, kernel_size, method='nn_upsample_conv'):
   Raises:
     ValueError: if `method` is not recognized.
   """
-  net_shape = tf.shape(net)
+  net_shape = tf.shape(input=net)
   height = net_shape[1]
   width = net_shape[2]
 
   if method == 'nn_upsample_conv':
-    net = tf.image.resize_nearest_neighbor(
-        net, [kernel_size[0] * height, kernel_size[1] * width])
+    net = tf.image.resize(
+        net, [kernel_size[0] * height, kernel_size[1] * width],
+        method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
     net = layers.conv2d(net, num_outputs, [4, 4], activation_fn=None)
   elif method == 'conv2d_transpose':
     net = layers.conv2d_transpose(
@@ -166,7 +168,7 @@ def pix2pix_generator(net,
   ###########
   # Encoder #
   ###########
-  with tf.variable_scope('encoder'):
+  with tf.compat.v1.variable_scope('encoder'):
     with contrib_framework.arg_scope([layers.conv2d],
                                      kernel_size=[4, 4],
                                      stride=2,
@@ -194,7 +196,7 @@ def pix2pix_generator(net,
   reversed_blocks = list(blocks)
   reversed_blocks.reverse()
 
-  with tf.variable_scope('decoder'):
+  with tf.compat.v1.variable_scope('decoder'):
     # Dropout is used at both train and test time as per 'Image-to-Image',
     # Section 2.1 (last paragraph).
     with contrib_framework.arg_scope([layers.dropout], is_training=True):
@@ -210,7 +212,7 @@ def pix2pix_generator(net,
           net = layers.dropout(net, keep_prob=block.decoder_keep_prob)
         end_points['decoder%d' % block_id] = net
 
-  with tf.variable_scope('output'):
+  with tf.compat.v1.variable_scope('output'):
     # Explicitly set the normalizer_fn to None to override any default value
     # that may come from an arg_scope, such as pix2pix_arg_scope.
     logits = layers.conv2d(
@@ -249,11 +251,11 @@ def pix2pix_discriminator(net, num_filters, padding=2, pad_mode='REFLECT',
 
   def padded(net, scope):
     if padding:
-      with tf.variable_scope(scope):
+      with tf.compat.v1.variable_scope(scope):
         spatial_pad = tf.constant(
             [[0, 0], [padding, padding], [padding, padding], [0, 0]],
             dtype=tf.int32)
-        return tf.pad(net, spatial_pad, pad_mode)
+        return tf.pad(tensor=net, paddings=spatial_pad, mode=pad_mode)
     else:
       return net
 

research/slim/nets/pix2pix_test.py

@@ -42,7 +42,7 @@ class GeneratorTest(tf.test.TestCase):
           upsample_method='nn_upsample_conv')
 
     with self.test_session() as session:
-      session.run(tf.global_variables_initializer())
+      session.run(tf.compat.v1.global_variables_initializer())
       np_outputs = session.run(logits)
       self.assertListEqual([batch_size, height, width, num_outputs],
                            list(np_outputs.shape))
@@ -59,7 +59,7 @@ class GeneratorTest(tf.test.TestCase):
           upsample_method='conv2d_transpose')
 
     with self.test_session() as session:
-      session.run(tf.global_variables_initializer())
+      session.run(tf.compat.v1.global_variables_initializer())
       np_outputs = session.run(logits)
       self.assertListEqual([batch_size, height, width, num_outputs],
                            list(np_outputs.shape))

research/slim/nets/resnet_utils.py

@@ -117,8 +117,9 @@ def conv2d_same(inputs, num_outputs, kernel_size, stride, rate=1, scope=None):
     pad_total = kernel_size_effective - 1
     pad_beg = pad_total // 2
     pad_end = pad_total - pad_beg
-    inputs = tf.pad(inputs,
-                    [[0, 0], [pad_beg, pad_end], [pad_beg, pad_end], [0, 0]])
+    inputs = tf.pad(
+        tensor=inputs,
+        paddings=[[0, 0], [pad_beg, pad_end], [pad_beg, pad_end], [0, 0]])
     return slim.conv2d(inputs, num_outputs, kernel_size, stride=stride,
                        rate=rate, padding='VALID', scope=scope)
 
@@ -180,7 +181,7 @@ def stack_blocks_dense(net, blocks, output_stride=None,
   rate = 1
 
   for block in blocks:
-    with tf.variable_scope(block.scope, 'block', [net]) as sc:
+    with tf.compat.v1.variable_scope(block.scope, 'block', [net]) as sc:
       block_stride = 1
       for i, unit in enumerate(block.args):
         if store_non_strided_activations and i == len(block.args) - 1:
@@ -188,7 +189,7 @@ def stack_blocks_dense(net, blocks, output_stride=None,
           block_stride = unit.get('stride', 1)
           unit = dict(unit, stride=1)
 
-        with tf.variable_scope('unit_%d' % (i + 1), values=[net]):
+        with tf.compat.v1.variable_scope('unit_%d' % (i + 1), values=[net]):
           # If we have reached the target output_stride, then we need to employ
           # atrous convolution with stride=1 and multiply the atrous rate by the
           # current unit's stride for use in subsequent layers.
@@ -220,13 +221,14 @@ def stack_blocks_dense(net, blocks, output_stride=None,
   return net
 
 
-def resnet_arg_scope(weight_decay=0.0001,
+def resnet_arg_scope(
+    weight_decay=0.0001,
     batch_norm_decay=0.997,
     batch_norm_epsilon=1e-5,
     batch_norm_scale=True,
     activation_fn=tf.nn.relu,
     use_batch_norm=True,
-                     batch_norm_updates_collections=tf.GraphKeys.UPDATE_OPS):
+    batch_norm_updates_collections=tf.compat.v1.GraphKeys.UPDATE_OPS):
   """Defines the default ResNet arg scope.
 
   TODO(gpapan): The batch-normalization related default values above are

research/slim/nets/resnet_v1.py

@@ -109,7 +109,7 @@ def bottleneck(inputs,
   Returns:
     The ResNet unit's output.
   """
-  with tf.variable_scope(scope, 'bottleneck_v1', [inputs]) as sc:
+  with tf.compat.v1.variable_scope(scope, 'bottleneck_v1', [inputs]) as sc:
     depth_in = slim.utils.last_dimension(inputs.get_shape(), min_rank=4)
     if depth == depth_in:
       shortcut = resnet_utils.subsample(inputs, stride, 'shortcut')
@@ -219,7 +219,8 @@ def resnet_v1(inputs,
   Raises:
     ValueError: If the target output_stride is not valid.
   """
-  with tf.variable_scope(scope, 'resnet_v1', [inputs], reuse=reuse) as sc:
+  with tf.compat.v1.variable_scope(
+      scope, 'resnet_v1', [inputs], reuse=reuse) as sc:
     end_points_collection = sc.original_name_scope + '_end_points'
     with slim.arg_scope([slim.conv2d, bottleneck,
                          resnet_utils.stack_blocks_dense],
@@ -242,7 +243,8 @@ def resnet_v1(inputs,
 
         if global_pool:
           # Global average pooling.
-          net = tf.reduce_mean(net, [1, 2], name='pool5', keep_dims=True)
+          net = tf.reduce_mean(
+              input_tensor=net, axis=[1, 2], name='pool5', keepdims=True)
           end_points['global_pool'] = net
         if num_classes:
           net = slim.conv2d(net, num_classes, [1, 1], activation_fn=None,

research/slim/nets/resnet_v2.py

@@ -84,7 +84,7 @@ def bottleneck(inputs, depth, depth_bottleneck, stride, rate=1,
   Returns:
     The ResNet unit's output.
   """
-  with tf.variable_scope(scope, 'bottleneck_v2', [inputs]) as sc:
+  with tf.compat.v1.variable_scope(scope, 'bottleneck_v2', [inputs]) as sc:
     depth_in = slim.utils.last_dimension(inputs.get_shape(), min_rank=4)
     preact = slim.batch_norm(inputs, activation_fn=tf.nn.relu, scope='preact')
     if depth == depth_in:
@@ -181,7 +181,8 @@ def resnet_v2(inputs,
   Raises:
     ValueError: If the target output_stride is not valid.
   """
-  with tf.variable_scope(scope, 'resnet_v2', [inputs], reuse=reuse) as sc:
+  with tf.compat.v1.variable_scope(
+      scope, 'resnet_v2', [inputs], reuse=reuse) as sc:
     end_points_collection = sc.original_name_scope + '_end_points'
     with slim.arg_scope([slim.conv2d, bottleneck,
                          resnet_utils.stack_blocks_dense],
@@ -211,7 +212,8 @@ def resnet_v2(inputs,
 
         if global_pool:
           # Global average pooling.
-          net = tf.reduce_mean(net, [1, 2], name='pool5', keep_dims=True)
+          net = tf.reduce_mean(
+              input_tensor=net, axis=[1, 2], name='pool5', keepdims=True)
           end_points['global_pool'] = net
         if num_classes:
           net = slim.conv2d(net, num_classes, [1, 1], activation_fn=None,

research/slim/preprocessing/preprocessing_factory.py

@@ -56,6 +56,12 @@ def get_preprocessing(name, is_training=False, use_grayscale=False):
       'mobilenet_v1': inception_preprocessing,
       'mobilenet_v2': inception_preprocessing,
       'mobilenet_v2_035': inception_preprocessing,
+      'mobilenet_v3_small': inception_preprocessing,
+      'mobilenet_v3_large': inception_preprocessing,
+      'mobilenet_v3_small_minimalistic': inception_preprocessing,
+      'mobilenet_v3_large_minimalistic': inception_preprocessing,
+      'mobilenet_edgetpu': inception_preprocessing,
+      'mobilenet_edgetpu_075': inception_preprocessing,
       'mobilenet_v2_140': inception_preprocessing,
       'nasnet_mobile': inception_preprocessing,
       'nasnet_large': inception_preprocessing,