Merge pull request tensorflow#3608 from sachinprasadhs:patch-2

PiperOrigin-RevId: 417833128
sputney13 · Dec 22, 2021 · ef6cd07 · ef6cd07
2 parents ac907cc + 92ac5f3
commit ef6cd07
Show file tree

Hide file tree

Showing 4 changed files with 7 additions and 9 deletions.
diff --git a/docs/determinism.ipynb b/docs/determinism.ipynb
@@ -319,7 +319,7 @@
         "id": "gAJTLLsuFeuP"
       },
       "source": [
-        "Note: Setting `shuffle_files=True` also [disable](https://github.com/tensorflow/datasets/tree/master/tensorflow_datasets/core/dataset_builder.py?l=676\u0026rcl=354322021)  `experimental_deterministic` in [`tf.data.Options`](https://www.tensorflow.org/api_docs/python/tf/data/Options) to give some performance boost. So even small datasets which only have a single shard (like mnist), become non-deterministic.\n",
+        "Note: Setting `shuffle_files=True` also [disable](https://github.com/tensorflow/datasets/tree/master/tensorflow_datasets/core/dataset_builder.py?l=676\u0026rcl=354322021)  `deterministic` in [`tf.data.Options`](https://www.tensorflow.org/api_docs/python/tf/data/Options) to give some performance boost. So even small datasets which only have a single shard (like mnist), become non-deterministic.\n",
         "\n",
         "See recipe below to get deterministic file shuffling."
       ]

diff --git a/docs/performances.md b/docs/performances.md
@@ -116,11 +116,10 @@ ds = tfds.load('imagenet2012', split='train', shuffle_files=True)
 ```
 
 Additionally, when `shuffle_files=True`, TFDS disables
-[`options.experimental_deterministic`](https://www.tensorflow.org/api_docs/python/tf/data/Options#experimental_deterministic),
+[`options.deterministic`](https://www.tensorflow.org/api_docs/python/tf/data/Options#deterministic),
 which may give a slight performance boost. To get deterministic shuffling, it is
 possible to opt-out of this feature with `tfds.ReadConfig`: either by setting
-`read_config.shuffle_seed` or overwriting
-`read_config.options.experimental_deterministic`.
+`read_config.shuffle_seed` or overwriting `read_config.options.deterministic`.
 
 ### Auto-shard your data across workers (TF)
 

diff --git a/tensorflow_datasets/core/dataset_builder.py b/tensorflow_datasets/core/dataset_builder.py
@@ -646,11 +646,10 @@ def lookup_nest(features: Dict[str, Any]) -> Tuple[Any, ...]:
     # non-deterministic
     # This code should probably be moved inside tfreader, such as
     # all the tf.data.Options are centralized in a single place.
-    if (shuffle_files and
-        read_config.options.experimental_deterministic is None and
+    if (shuffle_files and read_config.options.deterministic is None and
         read_config.shuffle_seed is None):
       options = tf.data.Options()
-      options.experimental_deterministic = False
+      options.deterministic = False
       ds = ds.with_options(options)
     # If shuffle is False, keep the default value (deterministic), which
     # allow the user to overwritte it.

diff --git a/tensorflow_datasets/core/utils/read_config.py b/tensorflow_datasets/core/utils/read_config.py
@@ -38,8 +38,8 @@ class ReadConfig:
 
   Attributes:
     options: `tf.data.Options()`, dataset options to use. Note that when
-      `shuffle_files` is True and no seed is defined, experimental_deterministic
-      will be set to False internally, unless it is defined here.
+      `shuffle_files` is True and no seed is defined, deterministic will be set
+      to False internally, unless it is defined here.
     try_autocache: If True (default) and the dataset satisfy the right
       conditions (dataset small enough, files not shuffled,...) the dataset will
       be cached during the first iteration (through `ds = ds.cache()`).