Revert "Update usage of tf.contrib.data to tf.data.experimental" (#7654)

* Revert "Update tf.contrib.data to tf.data.experimental. (#7650)"

This reverts commit faf4bbb3.

* revert research

official/r1/resnet/resnet_run_loop.py

@@ -110,7 +110,7 @@ def process_record_dataset(dataset,
   # Operations between the final prefetch and the get_next call to the iterator
   # will happen synchronously during run time. We prefetch here again to
   # background all of the above processing work and keep it out of the
-  # critical training path. Setting buffer_size to tf.data.experimental.AUTOTUNE
+  # critical training path. Setting buffer_size to tf.contrib.data.AUTOTUNE
   # allows DistributionStrategies to adjust how many batches to fetch based
   # on how many devices are present.
   dataset = dataset.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)

official/transformer/utils/dataset.py

@@ -183,7 +183,7 @@ def _batch_examples(dataset, batch_size, max_length):
     # lengths as well. Resulting lengths of inputs and targets can differ.
     return grouped_dataset.padded_batch(bucket_batch_size, ([None], [None]))
 
-  return dataset.apply(tf.data.experimental.group_by_window(
+  return dataset.apply(tf.contrib.data.group_by_window(
       key_func=example_to_bucket_id,
       reduce_func=batching_fn,
       window_size=None,
@@ -223,7 +223,7 @@ def _read_and_batch_from_files(
   # Read files and interleave results. When training, the order of the examples
   # will be non-deterministic.
   dataset = dataset.apply(
-      tf.data.experimental.parallel_interleave(
+      tf.contrib.data.parallel_interleave(
           _load_records, sloppy=shuffle, cycle_length=num_parallel_calls))
 
   # Parse each tf.Example into a dictionary
@@ -235,9 +235,8 @@ def _read_and_batch_from_files(
   dataset = dataset.filter(lambda x, y: _filter_max_length((x, y), max_length))
 
   if static_batch:
-    dataset = dataset.padded_batch(
-        batch_size // max_length, ([max_length], [max_length]),
-        drop_remainder=True)
+    dataset = dataset.apply(tf.contrib.data.padded_batch_and_drop_remainder(
+        batch_size // max_length, ([max_length], [max_length])))
   else:
     # Group and batch such that each batch has examples of similar length.
     dataset = _batch_examples(dataset, batch_size, max_length)
@@ -245,7 +244,7 @@ def _read_and_batch_from_files(
   dataset = dataset.repeat(repeat)
 
   # Prefetch the next element to improve speed of input pipeline.
-  dataset = dataset.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
+  dataset = dataset.prefetch(buffer_size=tf.contrib.data.AUTOTUNE)
   return dataset
 
 

official/vision/detection/evaluation/coco_utils.py

@@ -318,7 +318,7 @@ class COCOGroundtruthGenerator(object):
             cycle_length=32,
             sloppy=False))
     dataset = dataset.map(self._parse_single_example, num_parallel_calls=64)
-    dataset = dataset.prefetch(tf.data.experimental.AUTOTUNE)
+    dataset = dataset.prefetch(tf.contrib.data.AUTOTUNE)
     dataset = dataset.batch(1, drop_remainder=False)
     return dataset
 

official/vision/image_classification/imagenet_preprocessing.py

@@ -128,7 +128,7 @@ def process_record_dataset(dataset,
   # Operations between the final prefetch and the get_next call to the iterator
   # will happen synchronously during run time. We prefetch here again to
   # background all of the above processing work and keep it out of the
-  # critical training path. Setting buffer_size to tf.data.experimental.AUTOTUNE
+  # critical training path. Setting buffer_size to tf.contrib.data.AUTOTUNE
   # allows DistributionStrategies to adjust how many batches to fetch based
   # on how many devices are present.
   dataset = dataset.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)