Add: Keras reproducibility guide in the call

[sayakpaul]

@fchollet I have some preliminary work done in this: https://bit.ly/34V5ZNz. I wanted to extend it and include a call for contribution for this particular guide.

I have some pointers in mind:

• Fixing seeds
• Ensuring development infrastructure unification
• tf-determinism for using deterministic CUDA kernels to trade away performance

With these, it's possible to attain at least some amount of reproducibility. But I am of course open to any suggestions you may have. I think it's time we had a guide on this topic.

[fchollet]

Great idea, this would be a good guide or perhaps a walkthrough for the keras-recipes section!

Related to this, we have been making changes and simplifications to determinism and have started adopting stateless RNGs everywhere. This means that all random layers/etc will own a seed variable and will only run ops that are fully-seeded (constant given the seed value). It makes randomness much easier to understand and control.

There are a couple new APIs we should bring up:

• A Keras API for seeding everything, keras.utils.set_random_seed(seed): a one-stop util for making everything seeded: numpy, tf, keras
• A TF API for enabling fully-deterministic computation at the cost of a performance degradation, tf.config.experimental.enable_op_determinism()

IMO this is something that we should publish around the time of the TF 2.8 release.

[sayakpaul]

Thank you @fchollet.

A TF API for enabling fully-deterministic computation at the cost of a performance degradation, tf.config.experimental.enable_op_determinism()

Sounds really good. We might want to even include a friendly warning message on the performance degradation part.