Uniform regularization; Visualization

bgshih · Dec 21, 2017 · adc6b65 · adc6b65
1 parent 62af1ff
commit adc6b65
Show file tree

Hide file tree

Showing 5 changed files with 104 additions and 8 deletions.
diff --git a/core/feature_extractor.py b/core/feature_extractor.py
@@ -7,6 +7,7 @@
 
 from rare.core import feature_extractor_pb2
 from rare.core import hyperparams
+from rare.utils import visualization_utils
 
 
 class FeatureExtractor(object):
@@ -22,6 +23,8 @@ def __init__(self,
   def preprocess(self, resized_inputs, scope=None):
     with tf.variable_scope(scope, 'FeatureExtractorPreprocess', [resized_inputs]):
       preprocessed_inputs = (2.0 / 255.0) * resized_inputs - 1.0
+    if self._summarize_inputs:
+      tf.summary.image('preprocessed_inputs', preprocessed_inputs, max_outputs=1)
     return preprocessed_inputs
 
   @abstractmethod
@@ -168,16 +171,14 @@ class BaselineFeatureExtractor(FeatureExtractor):
 
   def __init__(self,
                conv_hyperparams=None,
-               summarize_inputs=False):
-    super(BaselineFeatureExtractor, self).__init__()
+               summarize_inputs=False,
+               is_training=None):
+    super(BaselineFeatureExtractor, self).__init__(
+      summarize_inputs=summarize_inputs,
+      is_training=is_training)
     self._conv_hyperparams = conv_hyperparams # FIXME: add it back
     self._summarize_inputs = summarize_inputs
 
-  def preprocess(self, resized_inputs, scope=None):
-    with tf.variable_scope(scope, 'ModelPreprocess', [resized_inputs]):
-      preprocessed_inputs = (2.0 / 255.0) * resized_inputs - 1.0
-    return preprocessed_inputs
-
   def extract_features(self, preprocessed_inputs, scope=None):
     """Extract features
     Args:
@@ -215,7 +216,12 @@ def extract_features(self, preprocessed_inputs, scope=None):
         if self._summarize_inputs:
           for layer in [conv1, pool1, conv2, pool2, conv3,
                         conv4, pool4, conv5, conv6, pool6, conv7]:
-            tf.summary.histogram(layer.name, layer)
+            tf.summary.histogram(layer.op.name, layer)
+            # tf.summary.image(
+            #   layer.name,
+            #   visualization_utils.tile_activation_maps_max_dimensions(layer, 512, 512),
+            #   max_outputs=1
+            # )
     return [conv7]
 
 

diff --git a/core/hyperparams.proto b/core/hyperparams.proto
@@ -66,6 +66,7 @@ message Initializer {
     TruncatedNormalInitializer truncated_normal_initializer = 1;
     VarianceScalingInitializer variance_scaling_initializer = 2;
     OrthogonalInitializer orthogonal_initializer = 3;
+    UniformInitializer uniform_initializer = 4;
   }
 }
 
@@ -95,6 +96,11 @@ message OrthogonalInitializer {
   optional int32 seed = 2;
 }
 
+message UniformInitializer {
+  optional float minval = 1 [default = -0.1];
+  optional float maxval = 2 [default = 0.1];
+}
+
 // Configuration proto for batch norm to apply after convolution op. See
 // https://www.tensorflow.org/api_docs/python/tf/contrib/layers/batch_norm
 message BatchNorm {

diff --git a/core/hyperparams.py b/core/hyperparams.py
@@ -150,6 +150,10 @@ def _build_initializer(initializer):
       gain=initializer.orthogonal_initializer.gain,
       seed=initializer.orthogonal_initializer.seed
     )
+  if initializer_oneof == 'uniform_initializer':
+    return tf.random_uniform_initializer(
+      minval=initializer.uniform_initializer.minval,
+      maxval=initializer.uniform_initializer.maxval)
   raise ValueError('Unknown initializer function: {}'.format(
       initializer_oneof))
 

diff --git a/utils/visualization_utils.py b/utils/visualization_utils.py
@@ -0,0 +1,46 @@
+import tensorflow as tf
+
+from rare.utils import shape_utils
+
+
+def tile_activation_maps_max_dimensions(maps, max_height, max_width):
+  batch_size, map_height, map_width, map_depth = \
+    shape_utils.combined_static_and_dynamic_shape(maps)
+  num_rows = max_height // map_height
+  num_cols = max_width // map_width
+  return tile_activation_maps_rows_cols(maps, num_rows, num_cols)
+
+
+def tile_activation_maps_rows_cols(maps, num_rows, num_cols):
+  """
+  Args:
+    maps: [batch_size, map_height, map_width, map_depth]
+  Return:
+    tiled_map: [batch_size, tiled_height, tiled_width]
+  """
+  batch_size, map_height, map_width, map_depth = \
+    shape_utils.combined_static_and_dynamic_shape(maps)
+
+  # padding
+  num_maps = num_rows * num_cols
+  padded_map = tf.cond(
+    tf.greater(num_maps, map_depth),
+    true_fn=lambda: tf.pad(maps, [[0, 0], [0, 0], [0, 0], [0, tf.maximum(num_maps - map_depth, 0)]]),
+    false_fn=lambda: maps[:,:,:,:num_maps]
+  )
+
+  # reshape to [batch_size, map_height, map_width, num_rows, num_cols]
+  reshaped_map = tf.reshape(padded_map, [batch_size, map_height, map_width, num_rows, num_cols])
+
+  # unstack and concat along widths
+  width_concated_maps = tf.concat(
+    tf.unstack(reshaped_map, axis=4), # => list of [batch_size, map_height, map_width, num_rows]
+    axis=2) # => [batch_size, map_height, map_width * num_cols, num_rows]
+
+  tiled_map = tf.concat(
+    tf.unstack(width_concated_maps, axis=3), # => list of [batch_size, map_height, map_width * num_cols]
+    axis=1) # => [batch_size, map_height * num_rows, map_width * num_cols]
+
+  tiled_map = tf.expand_dims(tiled_map, axis=3)
+
+  return tiled_map
diff --git a/utils/visualization_utils_test.py b/utils/visualization_utils_test.py
@@ -0,0 +1,34 @@
+import tensorflow as tf
+
+from rare.utils import visualization_utils
+
+class VisualizationUtilsTest(tf.test.TestCase):
+
+  def test_tile_activation_maps_with_padding(self):
+    test_maps = tf.random_uniform([64, 32, 100, 16])
+    tiled_map = visualization_utils.tile_activation_maps_rows_cols(test_maps, 5, 5)
+
+    with self.test_session() as sess:
+      tiled_map_output = tiled_map.eval()
+      self.assertAllEqual(tiled_map_output.shape, [64, 32 * 5, 100 * 5, 1])
+
+  def test_tile_activation_maps_with_slicing(self):
+    test_maps = tf.random_uniform([64, 32, 100, 16])
+    tiled_map = visualization_utils.tile_activation_maps_rows_cols(test_maps, 5, 1)
+
+    with self.test_session() as sess:
+      tiled_map_output = tiled_map.eval()
+      self.assertAllEqual(tiled_map_output.shape, [64, 32 * 5, 100 * 1, 1])
+
+  def test_tile_activation_maps_max_sizes(self):
+    test_maps = tf.random_uniform([64, 32, 100, 16])
+    tiled_map = visualization_utils.tile_activation_maps_max_dimensions(
+      test_maps, 512, 512)
+
+    with self.test_session() as sess:
+      tiled_map_output = tiled_map.eval()
+      self.assertAllEqual(tiled_map_output.shape, [64, 512, 500, 1])
+
+
+if __name__ == '__main__':
+  tf.test.main()