Peng Wang comments

Results 8 comments of


                                            Peng Wang

API ownership

cc @rohan100jain

`tf.vectorized_map` fallback cause

We (the TF core team) don't monitor keras-team/keras-cv . The issue needs to be reposted to tensorflow/tensorflow to be under our radar.

`tf.vectorized_map` fallback cause

It can be just a title and a link to this one. TF has a system for alerting TF developers of tensorflow/tensorflow issues.

`tf.vectorized_map` fallback cause

CCing @ishark .

`tf.vectorized_map` fallback cause

> @wangpengmit Can I ask you why you have mentioned @ishark 1 month ago? Is there a new API ownership plan? Yeah, we still lack a primary owner of `tf.vectorized_map`,...

`tf.random.shuffle` is deterministic when XLA is used

For RNG in general, `tf.random.stateless_*` and `tf.random.Generator` are recommended over `tf.random.shuffle` and alike. Please see https://www.tensorflow.org/guide/random_numbers . We don't have `stateless_shuffle` or `Generator.shuffle` yet. We are working on adding the...

backing_device changes depending on the tensor's dtype

Hi @miguelusque, you've hit a long-running issue internally known as "the int32 problem". Basically TF carves out int32 as "the shape dtype", to workaround some design problems of the device-placement...

Internal seed of keras.layers.Dropout gets increased by +1 every time a tf.function gets created from the model

I don't have a strong opinion on this. Increasing a user-provided seed does surprise me a bit. OTOH the problem in https://github.com/keras-team/tf-keras/issues/60 does exist. Letting two legacy stateful RNG ops...