A utility function to help create a `tf.distribute.DistributedDataset`. Args: dataset_or_fn: A instance of `tf.data.Dataset`, or a "dataset function" returning a `tf.data.Dataset`. If it is a function, it may optionally have an argument named `input_context` which will be
(self, dataset_or_fn, *args, **kwargs)
| 96 | return eval_loop_fn |
| 97 | |
| 98 | def distribute_dataset(self, dataset_or_fn, *args, **kwargs): |
| 99 | """A utility function to help create a `tf.distribute.DistributedDataset`. |
| 100 | |
| 101 | Args: |
| 102 | dataset_or_fn: A instance of `tf.data.Dataset`, or a "dataset function" |
| 103 | returning a `tf.data.Dataset`. If it is a function, it may optionally |
| 104 | have an argument named `input_context` which will be passed a |
| 105 | `tf.distribute.InputContext` instance. |
| 106 | *args: Any positional arguments to pass through to `dataset_or_fn`. |
| 107 | **kwargs: Any keyword arguments to pass through to `dataset_or_fn`. |
| 108 | |
| 109 | Returns: |
| 110 | A distributed Dataset. |
| 111 | """ |
| 112 | if getattr(self, "_is_async", False): |
| 113 | per_worker_dataset_fn = functools.partial( |
| 114 | orbit.utils.make_distributed_dataset, self._strategy, dataset_or_fn, |
| 115 | *args, **kwargs) |
| 116 | per_worker_dataset_fn = tf.function(per_worker_dataset_fn) |
| 117 | |
| 118 | return self.coordinator_for_async().create_per_worker_dataset( |
| 119 | per_worker_dataset_fn |
| 120 | ) |
| 121 | else: |
| 122 | return orbit.utils.make_distributed_dataset(self._strategy, dataset_or_fn, |
| 123 | *args, **kwargs) |
| 124 | |
| 125 | |
| 126 | def get_runtime_options(config: ExperimentConfig): |