Fix a bug in seek_next due to which seeks were performed 2^(num_epochs - 1)

times instead of num_epochs times.

expanded_list and string_list were copies by reference which mean the list was
being doubled with each call to extend().

Also don't add the variables created to the TRAINABLE variables collection.

PiperOrigin-RevId: 164020330

2 files changed, 11 insertions, 8 deletions

diff --git a/tensorflow/contrib/input_pipeline/python/ops/input_pipeline_ops.py b/tensorflow/contrib/input_pipeline/python/ops/input_pipeline_ops.py
index be22b5c579..0092ae74b2 100644
--- a/tensorflow/contrib/input_pipeline/python/ops/input_pipeline_ops.py
+++ b/tensorflow/contrib/input_pipeline/python/ops/input_pipeline_ops.py
@@ -55,7 +55,7 @@ def _maybe_randomize_list(string_list, shuffle):
 def _create_list(string_list, shuffle, seed, num_epochs):
   if shuffle and seed:
     random.seed(seed)
-  expanded_list = _maybe_randomize_list(string_list, shuffle)
+  expanded_list = _maybe_randomize_list(string_list, shuffle)[:]
   if num_epochs:
     for _ in range(num_epochs - 1):
       expanded_list.extend(_maybe_randomize_list(string_list, shuffle))
@@ -89,18 +89,21 @@ def seek_next(string_list, shuffle=False, seed=None, num_epochs=None):
         name="obtain_next_counter",
         initializer=constant_op.constant(
             -1, dtype=dtypes.int64),
-        dtype=dtypes.int64)
+        dtype=dtypes.int64,
+        trainable=False)
     with ops.colocate_with(counter):
       string_tensor = variable_scope.get_variable(
           name="obtain_next_expanded_list",
           initializer=constant_op.constant(expanded_list),
-          dtype=dtypes.string)
+          dtype=dtypes.string,
+          trainable=False)
     if num_epochs:
       filename_counter = variable_scope.get_variable(
           name="obtain_next_filename_counter",
           initializer=constant_op.constant(
               0, dtype=dtypes.int64),
-          dtype=dtypes.int64)
+          dtype=dtypes.int64,
+          trainable=False)
       c = filename_counter.count_up_to(len(expanded_list))
       with ops.control_dependencies([c]):
         return obtain_next(string_tensor, counter)
diff --git a/tensorflow/contrib/input_pipeline/python/ops/input_pipeline_ops_test.py b/tensorflow/contrib/input_pipeline/python/ops/input_pipeline_ops_test.py
index d6c0bd62de..9ed017592a 100644
--- a/tensorflow/contrib/input_pipeline/python/ops/input_pipeline_ops_test.py
+++ b/tensorflow/contrib/input_pipeline/python/ops/input_pipeline_ops_test.py
@@ -73,16 +73,16 @@ class InputPipelineOpsTest(test.TestCase):
       ])
       self._assert_output([b"a", b"b", b"c"], session, elem)
 
-  def testSeekNextLimitEpochsTwo(self):
+  def testSeekNextLimitEpochsThree(self):
     string_list = ["a", "b", "c"]
     with self.test_session() as session:
-      elem = input_pipeline_ops.seek_next(string_list, num_epochs=2)
+      elem = input_pipeline_ops.seek_next(string_list, num_epochs=3)
       session.run([
           variables.local_variables_initializer(),
           variables.global_variables_initializer()
       ])
-      # Expect to see [a, b, c] two times.
-      self._assert_output([b"a", b"b", b"c"] * 2, session, elem)
+      # Expect to see [a, b, c] three times.
+      self._assert_output([b"a", b"b", b"c"] * 3, session, elem)
 
 
 if __name__ == "__main__":