Marc van Zee comments

Results 104 comments of


                                            Marc van Zee

[tfjs-converter] Unsupported Ops: InvertPermutation, TensorScatterUpdate

+1 for InvertPermutation.

Should Flax return FrozenDicts or regular dicts?

Thanks for the input @n2cholas! After chatting with @jheek offline, the consensus is that it is indeed useful to return regular dicts, but that we block implementing this on merging...

Should Flax return FrozenDicts or regular dicts?

Sorry for the delay -- I was on parental leave. @jheek could you tell us whether any progress has been made on merging the chex and flax dataclasses?

Return key from `traverse_util.Traversal.iterate()`

@Dsantra92 sure, feel free to give it a try. We would have to run it against all our internal tests as well, but we can do that once all our...

Showcase "transfer learning" with imagenet example

@cgarciae who is working on a transfer learning example

Improve lifted transforms docstrings

#1980 improves some docstrings (also adds a variation of the example above in the `nn.scan` docstring). I added placeholders with `#1977` in that PR where docstrings should further be improved.

Improve documentation of make_causal_mask

Thanks for the feedback @rongcuid, I have renamed this issue to clarify that we should improve the documentation of causal mask.

Improve RTD for initializers

Hi @billmark , thanks for your issue! Could you tell me in a bit more detail what you were missing from the description of `kernel_init` and `bias_init` arguments?

Improve RTD for initializers

Thanks for the feedback, I've updated the issue description. Indeed, we don't seem to have our initializers on RTD, while we do have the activation functions. We can simply add...

Flax avg_pool also includes padding tokens when computing the mean

Discussed this offline with @jheek and @cgarciae. We agreed that the current behavior is not desirable since we are assuming that padding tokens for avg_pool are 0's and we include...