Merge remote-tracking branch 'upstream/master' into vat

# Conflicts:
#	neural_structured_learning/research/gam/data/dataset.py
#	neural_structured_learning/research/gam/data/loaders.py
#	neural_structured_learning/research/gam/experiments/helper.py
#	neural_structured_learning/research/gam/experiments/run_train_gam.py
#	neural_structured_learning/research/gam/experiments/run_train_gam_graph.py
#	neural_structured_learning/research/gam/trainer/trainer_agreement.py
#	neural_structured_learning/research/gam/trainer/trainer_classification.py
#	neural_structured_learning/research/gam/trainer/trainer_cotrain.py

Author: otiliastr

neural_structured_learning/research/gam/data/preprocessing.py

@@ -28,30 +28,6 @@ def convert_image(image):
   image *= 2.
   return image
 
-def split_train_val(indices, ratio_val, rng, max_num_val=None):
-  """Split the train sample indices into train and validation.
-
-  Args:
-    indices: A numpy array containing the indices of the training samples.
-    ratio_val: A float number between (0, 1) representing the ratio of samples
-      to use for validation.
-    rng: A random number generator.
-    max_num_val: An integer representing the maximum number of samples to
-      include in the validation set.
-
-  Returns:
-    Two numpy arrays containing the subset of indices used for training, and
-    validation, respectively.
-  """
-  num_samples = indices.shape[0]
-  num_val = int(ratio_val * num_samples)
-  if max_num_val and num_val > max_num_val:
-    num_val = max_num_val
-  ind = np.arange(0, num_samples)
-  rng.shuffle(ind)
-  ind_val = ind[:num_val]
-  ind_train = ind[num_val:]
-  return ind_train, ind_val
 
 def split_train_val(indices, ratio_val, rng, max_num_val=None):
   """Split the train sample indices into train and validation.

neural_structured_learning/research/gam/trainer/trainer_classification.py

@@ -493,7 +493,6 @@ def _get_agreement_reg_loss(self, data, is_train, features_shape):
     # edges at the end of training, so the shapes don't match needs fixing.
     left = tf.concat((labels_ll_left, labels_lu_left, predictions_uu_left),
                      axis=0)
-    # left = tf.stop_gradient(left)
     right = tf.concat(
         (predictions_ll_right, predictions_lu_right, predictions_uu_right),
         axis=0)
@@ -517,7 +516,6 @@ def _get_agreement_reg_loss(self, data, is_train, features_shape):
         src_indices=indices_uu_left,
         tgt_indices=indices_uu_right)
     agreement = tf.concat((agreement_ll, agreement_lu, agreement_uu), axis=0)
-    # agreement = tf.stop_gradient(agreement)
     if self.penalize_neg_agr:
       # Since the agreement is predicting scores between [0, 1], anything
       # under 0.5 should represent disagreement. Therefore, we want to encourage
@@ -712,17 +710,17 @@ def edge_iterator(self, data, batch_size, labeling):
   def _evaluate(self, indices, split, session, summary_writer):
     """Evaluates the samples with the provided indices."""
     data_iterator_val = batch_iterator(
-      indices,
-      batch_size=self.batch_size,
-      shuffle=False,
-      allow_smaller_batch=True,
-      repeat=False)
+        indices,
+        batch_size=self.batch_size,
+        shuffle=False,
+        allow_smaller_batch=True,
+        repeat=False)
     feed_dict_val = self._construct_feed_dict(data_iterator_val, split)
     cummulative_acc = 0.0
     num_samples = 0
     while feed_dict_val is not None:
       val_acc, batch_size_actual = session.run(
-        (self.accuracy, self.batch_size_actual), feed_dict=feed_dict_val)
+          (self.accuracy, self.batch_size_actual), feed_dict=feed_dict_val)
       cummulative_acc += val_acc * batch_size_actual
       num_samples += batch_size_actual
       feed_dict_val = self._construct_feed_dict(data_iterator_val, split)
@@ -732,8 +730,8 @@ def _evaluate(self, indices, split, session, summary_writer):
     if self.enable_summaries:
       summary = tf.Summary()
       summary.value.add(
-        tag='ClassificationModel/' + split + '_acc',
-        simple_value=cummulative_acc)
+          tag='ClassificationModel/' + split + '_acc',
+          simple_value=cummulative_acc)
       iter_cls_total = session.run(self.iter_cls_total)
       summary_writer.add_summary(summary, iter_cls_total)
       summary_writer.flush()