PiperOrigin-RevId: 197225788

Author: aquariusjay

research/deeplab/README.md

@@ -104,14 +104,27 @@ Misc:
 
 To get help with issues you may encounter while using the DeepLab Tensorflow
 implementation, create a new question on
-[StackOverflow](https://stackoverflow.com/) with the tags "tensorflow" and
-"deeplab".
+[StackOverflow](https://stackoverflow.com/) with the tag "tensorflow".
 
 Please report bugs (i.e., broken code, not usage questions) to the
 tensorflow/models GitHub [issue
 tracker](https://github.com/tensorflow/models/issues), prefixing the issue name
 with "deeplab".
 
+## Change Logs
+
+### March 22, 2018
+
+Release checkpoints using MobileNet-V2 as network backbone and pretrained on
+PASCAL VOC 2012 and Cityscapes.
+
+
+### March 5, 2018
+
+First release of DeepLab in TensorFlow including deeper Xception network
+backbone. Include chekcpoints that have been pretrained on PASCAL VOC 2012
+and Cityscapes.
+
 ## References
 
 1.  **Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs**<br />

research/deeplab/common.py

@@ -95,6 +95,7 @@
 # Test set name.
 TEST_SET = 'test'
 
+
 class ModelOptions(
     collections.namedtuple('ModelOptions', [
         'outputs_to_num_classes',
@@ -109,7 +110,8 @@ class ModelOptions(
         'decoder_output_stride',
         'decoder_use_separable_conv',
         'logits_kernel_size',
-        'model_variant'
+        'model_variant',
+        'depth_multiplier',
     ])):
   """Immutable class to hold model options."""
 
@@ -139,4 +141,4 @@ def __new__(cls,
         FLAGS.aspp_with_batch_norm, FLAGS.aspp_with_separable_conv,
         FLAGS.multi_grid, FLAGS.decoder_output_stride,
         FLAGS.decoder_use_separable_conv, FLAGS.logits_kernel_size,
-        FLAGS.model_variant)
+        FLAGS.model_variant, FLAGS.depth_multiplier)

research/deeplab/core/feature_extractor.py

@@ -17,8 +17,9 @@
 import functools
 import tensorflow as tf
 
+from deeplab.core import resnet_v1_beta
 from deeplab.core import xception
-from nets.mobilenet import mobilenet as mobilenet_lib
+from tensorflow.contrib.slim.nets import resnet_utils
 from nets.mobilenet import mobilenet_v2
 
 
@@ -56,10 +57,12 @@ def _mobilenet_v2(net,
   """
   with tf.variable_scope(
       scope, 'MobilenetV2', [net], reuse=reuse) as scope:
-    return mobilenet_lib.mobilenet_base(
+    return mobilenet_v2.mobilenet_base(
         net,
         conv_defs=mobilenet_v2.V2_DEF,
-        multiplier=depth_multiplier,
+        depth_multiplier=depth_multiplier,
+        min_depth=8 if depth_multiplier == 1.0 else 1,
+        divisible_by=8 if depth_multiplier == 1.0 else 1,
         final_endpoint=final_endpoint or _MOBILENET_V2_FINAL_ENDPOINT,
         output_stride=output_stride,
         scope=scope)
@@ -68,13 +71,25 @@ def _mobilenet_v2(net,
 # A map from network name to network function.
 networks_map = {
     'mobilenet_v2': _mobilenet_v2,
+    'resnet_v1_50': resnet_v1_beta.resnet_v1_50,
+    'resnet_v1_50_beta': resnet_v1_beta.resnet_v1_50_beta,
+    'resnet_v1_101': resnet_v1_beta.resnet_v1_101,
+    'resnet_v1_101_beta': resnet_v1_beta.resnet_v1_101_beta,
+    'xception_41': xception.xception_41,
     'xception_65': xception.xception_65,
+    'xception_71': xception.xception_71,
 }
 
 # A map from network name to network arg scope.
 arg_scopes_map = {
     'mobilenet_v2': mobilenet_v2.training_scope,
+    'resnet_v1_50': resnet_utils.resnet_arg_scope,
+    'resnet_v1_50_beta': resnet_utils.resnet_arg_scope,
+    'resnet_v1_101': resnet_utils.resnet_arg_scope,
+    'resnet_v1_101_beta': resnet_utils.resnet_arg_scope,
+    'xception_41': xception.xception_arg_scope,
     'xception_65': xception.xception_arg_scope,
+    'xception_71': xception.xception_arg_scope,
 }
 
 # Names for end point features.
@@ -86,19 +101,49 @@ def _mobilenet_v2(net,
         # The provided checkpoint does not include decoder module.
         DECODER_END_POINTS: None,
     },
+    'resnet_v1_50': {
+        DECODER_END_POINTS: ['block1/unit_2/bottleneck_v1/conv3'],
+    },
+    'resnet_v1_50_beta': {
+        DECODER_END_POINTS: ['block1/unit_2/bottleneck_v1/conv3'],
+    },
+    'resnet_v1_101': {
+        DECODER_END_POINTS: ['block1/unit_2/bottleneck_v1/conv3'],
+    },
+    'resnet_v1_101_beta': {
+        DECODER_END_POINTS: ['block1/unit_2/bottleneck_v1/conv3'],
+    },
+    'xception_41': {
+        DECODER_END_POINTS: [
+            'entry_flow/block2/unit_1/xception_module/'
+            'separable_conv2_pointwise',
+        ],
+    },
     'xception_65': {
         DECODER_END_POINTS: [
             'entry_flow/block2/unit_1/xception_module/'
             'separable_conv2_pointwise',
         ],
-    }
+    },
+    'xception_71': {
+        DECODER_END_POINTS: [
+            'entry_flow/block2/unit_1/xception_module/'
+            'separable_conv2_pointwise',
+        ],
+    },
 }
 
 # A map from feature extractor name to the network name scope used in the
 # ImageNet pretrained versions of these models.
 name_scope = {
     'mobilenet_v2': 'MobilenetV2',
+    'resnet_v1_50': 'resnet_v1_50',
+    'resnet_v1_50_beta': 'resnet_v1_50',
+    'resnet_v1_101': 'resnet_v1_101',
+    'resnet_v1_101_beta': 'resnet_v1_101',
+    'xception_41': 'xception_41',
     'xception_65': 'xception_65',
+    'xception_71': 'xception_71',
 }
 
 # Mean pixel value.
@@ -118,7 +163,13 @@ def _preprocess_zero_mean_unit_range(inputs):
 
 _PREPROCESS_FN = {
     'mobilenet_v2': _preprocess_zero_mean_unit_range,
+    'resnet_v1_50': _preprocess_subtract_imagenet_mean,
+    'resnet_v1_50_beta': _preprocess_zero_mean_unit_range,
+    'resnet_v1_101': _preprocess_subtract_imagenet_mean,
+    'resnet_v1_101_beta': _preprocess_zero_mean_unit_range,
+    'xception_41': _preprocess_zero_mean_unit_range,
     'xception_65': _preprocess_zero_mean_unit_range,
+    'xception_71': _preprocess_zero_mean_unit_range,
 }
 
 
@@ -140,7 +191,8 @@ def mean_pixel(model_variant=None):
   Returns:
     Mean pixel value.
   """
-  if model_variant is None:
+  if model_variant in ['resnet_v1_50',
+                       'resnet_v1_101'] or model_variant is None:
     return _MEAN_RGB
   else:
     return [127.5, 127.5, 127.5]
@@ -159,7 +211,8 @@ def extract_features(images,
                      regularize_depthwise=False,
                      preprocess_images=True,
                      num_classes=None,
-                     global_pool=False):
+                     global_pool=False,
+                     use_bounded_activations=False):
   """Extracts features by the particular model_variant.
 
   Args:
@@ -184,6 +237,8 @@ def extract_features(images,
       to None for dense prediction tasks.
     global_pool: Global pooling for image classification task. Defaults to
       False, since dense prediction tasks do not use this.
+    use_bounded_activations: Whether or not to use bounded activations. Bounded
+      activations better lend themselves to quantized inference.
 
   Returns:
     features: A tensor of size [batch, feature_height, feature_width,
@@ -195,7 +250,25 @@ def extract_features(images,
   Raises:
     ValueError: Unrecognized model variant.
   """
-  if 'xception' in model_variant:
+  if 'resnet' in model_variant:
+    arg_scope = arg_scopes_map[model_variant](
+        weight_decay=weight_decay,
+        batch_norm_decay=0.95,
+        batch_norm_epsilon=1e-5,
+        batch_norm_scale=True,
+        activation_fn=tf.nn.relu6 if use_bounded_activations else tf.nn.relu)
+    features, end_points = get_network(
+        model_variant, preprocess_images, arg_scope)(
+            inputs=images,
+            num_classes=num_classes,
+            is_training=(is_training and fine_tune_batch_norm),
+            global_pool=global_pool,
+            output_stride=output_stride,
+            multi_grid=multi_grid,
+            reuse=reuse,
+            scope=name_scope[model_variant],
+            use_bounded_activations=use_bounded_activations)
+  elif 'xception' in model_variant:
     arg_scope = arg_scopes_map[model_variant](
         weight_decay=weight_decay,
         batch_norm_decay=0.9997,