François Chollet
François Chollet
Thanks for the detailed analysis. The lack of the issue with other eager backends, and the disappearance of the issue when using a tf.function, strongly indicate that the leak may...
We could consider casting the values to `floatx()` in `update_state()` -- would you like to open a PR @eli-osherovich ?
Thanks for the PR! @hertschuh, @jeffcarp, do you remember why this line was inserted? I believe it was intended to fix a bug, however removing it does not seem to...
Thank you, @james77777778 and @hertschuh ! @TheMGGdev are you able to add a basic, minimal unit test for your use case, so we avoid breaking it in the future? We...
Are you able to provide a unit test?
The code snippet you posted (2nd one) can be turned into an appropriate unit test I think.
Right -- random ops in Keras Core are basically always intended to be called with a `SeedGenerator` instance as the seed argument, since `seed=None` defaults to an integer seed, which...
The difficulty is that there will be some reference to a RNG seed variable that you'll have to take into account, something like -- ``` def fn(variables): trainable_variables = ......
I looked at this more closely. What I can propose is this: 1. Unseeded random ops use a global `SeedGenerator` 2. The state of the global seed generator (`KerasVariable` of...
That's roughly how it works, except it's actually much simpler and more intuitive. - You don't need to work about new types like `PRNGKey`. RNG state is a regular variable...