Merged commit includes the following changes: (#7690)

275538818  by Sergio Guadarrama:

    Support grayscale input images in Slim model training

--
275355841  by Sergio Guadarrama:

    Fixed cases where tf.TensorShape was constructed with float dimensions

    This is a prerequisite for making TensorShape and Dimension more strict
    about the types of their arguments.

--
275131829  by Sergio Guadarrama:

    updates mobilenet/README.md to be github compatible adds V2+ reference to mobilenet_v1.md file and fixes invalid markdown

--

PiperOrigin-RevId: 275538818

research/slim/datasets/visualwakewords.py

@@ -39,7 +39,7 @@ _FILE_PATTERN = '%s.record-*'
 
 _SPLITS_TO_SIZES = {
     'train': 82783,
-    'validation': 40504,
+    'val': 40504,
 }
 
 

research/slim/eval_image_classifier.py

@@ -82,6 +82,9 @@ tf.app.flags.DEFINE_integer(
 tf.app.flags.DEFINE_bool(
     'quantize', False, 'whether to use quantized graph or not.')
 
+tf.app.flags.DEFINE_bool('use_grayscale', False,
+                         'Whether to convert input images to grayscale.')
+
 FLAGS = tf.app.flags.FLAGS
 
 
@@ -124,7 +127,8 @@ def main(_):
     preprocessing_name = FLAGS.preprocessing_name or FLAGS.model_name
     image_preprocessing_fn = preprocessing_factory.get_preprocessing(
         preprocessing_name,
-        is_training=False)
+        is_training=False,
+        use_grayscale=FLAGS.use_grayscale)
 
     eval_image_size = FLAGS.eval_image_size or network_fn.default_image_size
 

research/slim/export_inference_graph.py

@@ -110,6 +110,9 @@ tf.app.flags.DEFINE_integer(
 tf.app.flags.DEFINE_bool('write_text_graphdef', False,
                          'Whether to write a text version of graphdef.')
 
+tf.app.flags.DEFINE_bool('use_grayscale', False,
+                         'Whether to convert input images to grayscale.')
+
 FLAGS = tf.app.flags.FLAGS
 
 
@@ -128,11 +131,14 @@ def main(_):
         num_classes=(dataset.num_classes - FLAGS.labels_offset),
         is_training=FLAGS.is_training)
     image_size = FLAGS.image_size or network_fn.default_image_size
+    num_channels = 1 if FLAGS.use_grayscale else 3
     if FLAGS.is_video_model:
-      input_shape = [FLAGS.batch_size, FLAGS.num_frames,
-                     image_size, image_size, 3]
+      input_shape = [
+          FLAGS.batch_size, FLAGS.num_frames, image_size, image_size,
+          num_channels
+      ]
     else:
-      input_shape = [FLAGS.batch_size, image_size, image_size, 3]
+      input_shape = [FLAGS.batch_size, image_size, image_size, num_channels]
     placeholder = tf.placeholder(name='input', dtype=tf.float32,
                                  shape=input_shape)
     network_fn(placeholder)

research/slim/nets/mobilenet/mobilenet_v2_test.py

@@ -157,7 +157,8 @@ class MobilenetV2Test(tf.test.TestCase):
     new_def = copy.deepcopy(mobilenet_v2.V2_DEF)
 
     def inverse_multiplier(output_params, multiplier):
-      output_params['num_outputs'] /= multiplier
+      output_params['num_outputs'] = int(
+          output_params['num_outputs'] / multiplier)
 
     new_def['spec'][0] = op(
         slim.conv2d,

research/slim/nets/nets_factory.py

@@ -128,9 +128,9 @@ def get_network_fn(name, num_classes, weight_decay=0.0, is_training=False):
     network_fn: A function that applies the model to a batch of images. It has
       the following signature:
           net, end_points = network_fn(images)
-      The `images` input is a tensor of shape [batch_size, height, width, 3]
-      with height = width = network_fn.default_image_size. (The permissibility
-      and treatment of other sizes depends on the network_fn.)
+      The `images` input is a tensor of shape [batch_size, height, width, 3 or
+       1] with height = width = network_fn.default_image_size. (The
+      permissibility and treatment of other sizes depends on the network_fn.)
       The returned `end_points` are a dictionary of intermediate activations.
       The returned `net` is the topmost layer, depending on `num_classes`:
       If `num_classes` was a non-zero integer, `net` is a logits tensor

research/slim/preprocessing/cifarnet_preprocessing.py

@@ -31,7 +31,8 @@ def preprocess_for_train(image,
                          output_height,
                          output_width,
                          padding=_PADDING,
-                         add_image_summaries=True):
+                         add_image_summaries=True,
+                         use_grayscale=False):
   """Preprocesses the given image for training.
 
   Note that the actual resizing scale is sampled from
@@ -43,6 +44,7 @@ def preprocess_for_train(image,
     output_width: The width of the image after preprocessing.
     padding: The amound of padding before and after each dimension of the image.
     add_image_summaries: Enable image summaries.
+    use_grayscale: Whether to convert the image from RGB to grayscale.
 
   Returns:
     A preprocessed image.
@@ -52,6 +54,8 @@ def preprocess_for_train(image,
 
   # Transform the image to floats.
   image = tf.to_float(image)
+  if use_grayscale:
+    image = tf.image.rgb_to_grayscale(image)
   if padding > 0:
     image = tf.pad(image, [[padding, padding], [padding, padding], [0, 0]])
   # Randomly crop a [height, width] section of the image.
@@ -74,8 +78,11 @@ def preprocess_for_train(image,
   return tf.image.per_image_standardization(distorted_image)
 
 
-def preprocess_for_eval(image, output_height, output_width,
-                        add_image_summaries=True):
+def preprocess_for_eval(image,
+                        output_height,
+                        output_width,
+                        add_image_summaries=True,
+                        use_grayscale=False):
   """Preprocesses the given image for evaluation.
 
   Args:
@@ -83,6 +90,7 @@ def preprocess_for_eval(image, output_height, output_width,
     output_height: The height of the image after preprocessing.
     output_width: The width of the image after preprocessing.
     add_image_summaries: Enable image summaries.
+    use_grayscale: Whether to convert the image from RGB to grayscale.
 
   Returns:
     A preprocessed image.
@@ -91,6 +99,8 @@ def preprocess_for_eval(image, output_height, output_width,
     tf.summary.image('image', tf.expand_dims(image, 0))
   # Transform the image to floats.
   image = tf.to_float(image)
+  if use_grayscale:
+    image = tf.image.rgb_to_grayscale(image)
 
   # Resize and crop if needed.
   resized_image = tf.image.resize_image_with_crop_or_pad(image,
@@ -103,8 +113,12 @@ def preprocess_for_eval(image, output_height, output_width,
   return tf.image.per_image_standardization(resized_image)
 
 
-def preprocess_image(image, output_height, output_width, is_training=False,
-                     add_image_summaries=True):
+def preprocess_image(image,
+                     output_height,
+                     output_width,
+                     is_training=False,
+                     add_image_summaries=True,
+                     use_grayscale=False):
   """Preprocesses the given image.
 
   Args:
@@ -114,15 +128,22 @@ def preprocess_image(image, output_height, output_width, is_training=False,
     is_training: `True` if we're preprocessing the image for training and
       `False` otherwise.
     add_image_summaries: Enable image summaries.
+    use_grayscale: Whether to convert the image from RGB to grayscale.
 
   Returns:
     A preprocessed image.
   """
   if is_training:
     return preprocess_for_train(
-        image, output_height, output_width,
-        add_image_summaries=add_image_summaries)
+        image,
+        output_height,
+        output_width,
+        add_image_summaries=add_image_summaries,
+        use_grayscale=use_grayscale)
   else:
     return preprocess_for_eval(
-        image, output_height, output_width,
-        add_image_summaries=add_image_summaries)
+        image,
+        output_height,
+        output_width,
+        add_image_summaries=add_image_summaries,
+        use_grayscale=use_grayscale)

research/slim/preprocessing/inception_preprocessing.py

@@ -160,7 +160,8 @@ def preprocess_for_train(image,
                          fast_mode=True,
                          scope=None,
                          add_image_summaries=True,
-                         random_crop=True):
+                         random_crop=True,
+                         use_grayscale=False):
   """Distort one image for training a network.
 
   Distorting images provides a useful technique for augmenting the data
@@ -186,6 +187,7 @@ def preprocess_for_train(image,
     add_image_summaries: Enable image summaries.
     random_crop: Enable random cropping of images during preprocessing for
       training.
+    use_grayscale: Whether to convert the image from RGB to grayscale.
   Returns:
     3-D float Tensor of distorted image used for training with range [-1, 1].
   """
@@ -242,6 +244,9 @@ def preprocess_for_train(image,
         lambda x, ordering: distort_color(x, ordering, fast_mode),
         num_cases=num_distort_cases)
 
+    if use_grayscale:
+      distorted_image = tf.image.rgb_to_grayscale(distorted_image)
+
     if add_image_summaries:
       tf.summary.image('final_distorted_image',
                        tf.expand_dims(distorted_image, 0))
@@ -255,7 +260,8 @@ def preprocess_for_eval(image,
                         width,
                         central_fraction=0.875,
                         scope=None,
-                        central_crop=True):
+                        central_crop=True,
+                        use_grayscale=False):
   """Prepare one image for evaluation.
 
   If height and width are specified it would output an image with that size by
@@ -275,12 +281,15 @@ def preprocess_for_eval(image,
     scope: Optional scope for name_scope.
     central_crop: Enable central cropping of images during preprocessing for
       evaluation.
+    use_grayscale: Whether to convert the image from RGB to grayscale.
   Returns:
     3-D float Tensor of prepared image.
   """
   with tf.name_scope(scope, 'eval_image', [image, height, width]):
     if image.dtype != tf.float32:
       image = tf.image.convert_image_dtype(image, dtype=tf.float32)
+    if use_grayscale:
+      image = tf.image.rgb_to_grayscale(image)
     # Crop the central region of the image with an area containing 87.5% of
     # the original image.
     if central_crop and central_fraction:
@@ -304,7 +313,8 @@ def preprocess_image(image,
                      bbox=None,
                      fast_mode=True,
                      add_image_summaries=True,
-                     crop_image=True):
+                     crop_image=True,
+                     use_grayscale=False):
   """Pre-process one image for training or evaluation.
 
   Args:
@@ -324,6 +334,7 @@ def preprocess_image(image,
     add_image_summaries: Enable image summaries.
     crop_image: Whether to enable cropping of images during preprocessing for
       both training and evaluation.
+    use_grayscale: Whether to convert the image from RGB to grayscale.
 
   Returns:
     3-D float Tensor containing an appropriately scaled image
@@ -339,6 +350,12 @@ def preprocess_image(image,
         bbox,
         fast_mode,
         add_image_summaries=add_image_summaries,
-        random_crop=crop_image)
+        random_crop=crop_image,
+        use_grayscale=use_grayscale)
   else:
-    return preprocess_for_eval(image, height, width, central_crop=crop_image)
+    return preprocess_for_eval(
+        image,
+        height,
+        width,
+        central_crop=crop_image,
+        use_grayscale=use_grayscale)

research/slim/preprocessing/lenet_preprocessing.py

@@ -23,7 +23,11 @@ import tensorflow as tf
 slim = tf.contrib.slim
 
 
-def preprocess_image(image, output_height, output_width, is_training):
+def preprocess_image(image,
+                     output_height,
+                     output_width,
+                     is_training,
+                     use_grayscale=False):
   """Preprocesses the given image.
 
   Args:
@@ -32,11 +36,15 @@ def preprocess_image(image, output_height, output_width, is_training):
     output_width: The width of the image after preprocessing.
     is_training: `True` if we're preprocessing the image for training and
       `False` otherwise.
+    use_grayscale: Whether to convert the image from RGB to grayscale.
 
   Returns:
     A preprocessed image.
   """
+  del is_training  # Unused argument
   image = tf.to_float(image)
+  if use_grayscale:
+    image = tf.image.rgb_to_grayscale(image)
   image = tf.image.resize_image_with_crop_or_pad(
       image, output_width, output_height)
   image = tf.subtract(image, 128.0)

research/slim/preprocessing/preprocessing_factory.py

@@ -28,13 +28,14 @@ from preprocessing import vgg_preprocessing
 slim = tf.contrib.slim
 
 
-def get_preprocessing(name, is_training=False):
+def get_preprocessing(name, is_training=False, use_grayscale=False):
   """Returns preprocessing_fn(image, height, width, **kwargs).
 
   Args:
     name: The name of the preprocessing function.
     is_training: `True` if the model is being used for training and `False`
       otherwise.
+    use_grayscale: Whether to convert the image from RGB to grayscale.
 
   Returns:
     preprocessing_fn: A function that preprocessing a single image (pre-batch).
@@ -80,6 +81,11 @@ def get_preprocessing(name, is_training=False):
 
   def preprocessing_fn(image, output_height, output_width, **kwargs):
     return preprocessing_fn_map[name].preprocess_image(
-        image, output_height, output_width, is_training=is_training, **kwargs)
+        image,
+        output_height,
+        output_width,
+        is_training=is_training,
+        use_grayscale=use_grayscale,
+        **kwargs)
 
   return preprocessing_fn

research/slim/preprocessing/vgg_preprocessing.py

@@ -287,7 +287,8 @@ def preprocess_for_train(image,
                          output_height,
                          output_width,
                          resize_side_min=_RESIZE_SIDE_MIN,
-                         resize_side_max=_RESIZE_SIDE_MAX):
+                         resize_side_max=_RESIZE_SIDE_MAX,
+                         use_grayscale=False):
   """Preprocesses the given image for training.
 
   Note that the actual resizing scale is sampled from
@@ -301,6 +302,7 @@ def preprocess_for_train(image,
       aspect-preserving resizing.
     resize_side_max: The upper bound for the smallest side of the image for
       aspect-preserving resizing.
+    use_grayscale: Whether to convert the image from RGB to grayscale.
 
   Returns:
     A preprocessed image.
@@ -312,11 +314,17 @@ def preprocess_for_train(image,
   image = _random_crop([image], output_height, output_width)[0]
   image.set_shape([output_height, output_width, 3])
   image = tf.to_float(image)
+  if use_grayscale:
+    image = tf.image.rgb_to_grayscale(image)
   image = tf.image.random_flip_left_right(image)
   return _mean_image_subtraction(image, [_R_MEAN, _G_MEAN, _B_MEAN])
 
 
-def preprocess_for_eval(image, output_height, output_width, resize_side):
+def preprocess_for_eval(image,
+                        output_height,
+                        output_width,
+                        resize_side,
+                        use_grayscale=False):
   """Preprocesses the given image for evaluation.
 
   Args:
@@ -324,6 +332,7 @@ def preprocess_for_eval(image, output_height, output_width, resize_side):
     output_height: The height of the image after preprocessing.
     output_width: The width of the image after preprocessing.
     resize_side: The smallest side of the image for aspect-preserving resizing.
+    use_grayscale: Whether to convert the image from RGB to grayscale.
 
   Returns:
     A preprocessed image.
@@ -332,12 +341,18 @@ def preprocess_for_eval(image, output_height, output_width, resize_side):
   image = _central_crop([image], output_height, output_width)[0]
   image.set_shape([output_height, output_width, 3])
   image = tf.to_float(image)
+  if use_grayscale:
+    image = tf.image.rgb_to_grayscale(image)
   return _mean_image_subtraction(image, [_R_MEAN, _G_MEAN, _B_MEAN])
 
 
-def preprocess_image(image, output_height, output_width, is_training=False,
+def preprocess_image(image,
+                     output_height,
+                     output_width,
+                     is_training=False,
                      resize_side_min=_RESIZE_SIDE_MIN,
-                     resize_side_max=_RESIZE_SIDE_MAX):
+                     resize_side_max=_RESIZE_SIDE_MAX,
+                     use_grayscale=False):
   """Preprocesses the given image.
 
   Args:
@@ -353,13 +368,15 @@ def preprocess_image(image, output_height, output_width, is_training=False,
       aspect-preserving resizing. If `is_training` is `False`, this value is
       ignored. Otherwise, the resize side is sampled from
         [resize_size_min, resize_size_max].
+    use_grayscale: Whether to convert the image from RGB to grayscale.
 
   Returns:
     A preprocessed image.
   """
   if is_training:
     return preprocess_for_train(image, output_height, output_width,
-                                resize_side_min, resize_side_max)
+                                resize_side_min, resize_side_max,
+                                use_grayscale)
   else:
     return preprocess_for_eval(image, output_height, output_width,
-                               resize_side_min)
+                               resize_side_min, use_grayscale)

research/slim/train_image_classifier.py

@@ -206,6 +206,9 @@ tf.app.flags.DEFINE_integer(
 tf.app.flags.DEFINE_integer('max_number_of_steps', None,
                             'The maximum number of training steps.')
 
+tf.app.flags.DEFINE_bool('use_grayscale', False,
+                         'Whether to convert input images to grayscale.')
+
 #####################
 # Fine-Tuning Flags #
 #####################
@@ -433,7 +436,8 @@ def main(_):
     preprocessing_name = FLAGS.preprocessing_name or FLAGS.model_name
     image_preprocessing_fn = preprocessing_factory.get_preprocessing(
         preprocessing_name,
-        is_training=True)
+        is_training=True,
+        use_grayscale=FLAGS.use_grayscale)
 
     ##############################################################
     # Create a dataset provider that loads data from the dataset #