[tf.data] Add option for setting intra-op parallelism on a private threadpool.
This changes the default behavior when using `PrivateThreadPool` with `override_threadpool()`. It now defaults to using a maximum intra-op parallelism of 1 (which tends to be the most effective setting for high-throughput pipelines that are otherwise parallelized in the `Dataset.map()` or `tf.contrib.data.map_and_batch()` transformations. PiperOrigin-RevId: 201561361
Loading
Please sign in to comment