non-determinism on tf.keras.layers.UpSampling2D(..., interpolation='nearest')

[emiliocoutinho]

I believe there is a source of non-determinism on tf.keras.layers.UpSampling2D with interpolation='nearest', as you suggest on Other Possible GPU-Specific Sources of Non-Determinism.

In my case, I was building an auto-encoder, and on the decoder part, there were some procedures that I was not satisfied with and I was just waiting for the right time to change them. Turns out that this part was responsible for the non-deterministic behavior and as in #12, I replaced:

tf.keras.layers.UpSampling2D(..., interpolation='nearest')
ReflectionPadding2D()  # that is nothing more than a tf.pad(..., mode='REFLECT')
tf.keras.layers.Conv2D(..., padding='valid')

by:

tf.keras.layers.Conv2DTranspose(...,padding='same')

and now I have reproducible results on GPU and on CPU.

Thanks for the effort in building this package, it is very useful.

Best Regards.

[duncanriach]

Awesome work. Thanks Emilio. I'm going to move this source up into the confirmed list and review its priority relative to other ops, given that it's probably going to be a similar solution to that for bilinear filtering (referencing TF PR 39243). Thanks also for reporting the work-around.

[duncanriach]

With this commit, I have updated README.md to reflect all of this. I have also added fixing nearest neighbor resizing to my to-do list. This feature addition is in the pipeline to get added to TensorFlow. I'm going to close this issue now.

[duncanriach]

Closing.