Lucas David

It failed again. I apologize for that. I'm doing my best, but its difficult to see what's wrong without being able to test this locally or see the logs. I'd...

@rchao I think I figured it out. There are two problems. Firstly, function `multi_worker_testing_utils.mnist_synthetic_dataset` had a glitch which ignored the parameter `batch_size` and set `batch=64` for all cases. (Other test...

I might have not expressed myself clearly: I figured out what the error was (I believe MultiWorkerMirroredStrategy#run is returning the expected loss value divided by num_workers/two), but I am unsure...

Is it possible that MWMS is doing reduction ReductionOp.MEAN? ([distribute/distribute_lib.py#L1449](https://github.com/tensorflow/tensorflow/blob/v2.10.0/tensorflow/python/distribute/distribute_lib.py#L1449)) That would explain why we are getting loss/2.

> The `loss` of the result of `Model.fit()` should not change with the strategy being used. Yes, that's the idea behind my proposed testcase: if all samples have associated loss...

@rchao I fixed it. The problem was that [experimental_local_results](https://www.tensorflow.org/api_docs/python/tf/distribute/experimental/MultiWorkerMirroredStrategy#experimental_local_results) was discarding the results from the second worker. More detailed: I mentioned above that results were already gathered when reduce_per_replica was...

@rchao done. While we are here, I would like to make two suggestions regarding the reduction property. Let me know if any of them interest you and I will refactor...

Sounds good. :-) Sorry I pushed the wrong button. I think you need to approve it again. :grimacing:

I believe `include ActiveModel::Serialization` is missing in the `AuthToken` model? Still, I managed to get it to work with this: ```rb class AuthTokenSerializer < ActiveModel::Serializer attributes :jwt, :payload def jwt...

`tf.math.divide_no_nan` would fix the first edge case. However, assuming a correct usage of the API, I'm not sure it would ever happen in the first place: it's impossible for the...