Sam Anklesaria

Results 25 comments of Sam Anklesaria

Is there still interest in adding this feature? I might be interested in taking it on.

The reshape/broadcasting implementation doesn't use any of the nice structure from the module. What about the following: ```julia vec(A) = reshape(A, :) Zygote.@adjoint kronecker(A, B) = kronecker(A,B), dy-> ( dy.A...

Most of the errors are of the form 'RuntimeError: Error saving audio file: failed to open file /tmp/...'. These don't occur when I run the tests locally. I assume this...

It seems like the CI hasn't been working for a while: the last release didn't seem to work either (https://github.com/pytorch/audio/actions/runs/14930335102).

It's odd that `torchaudio_unittest/io/stream_writer_test.py` is still running: I thought pytest's `-k 'not /io/'` should exclude that. It seems to on my machine.

The `padding` argument to `nnx.ConvTranspose` is serving a different role than the `padding` argument of `torch.ConvTranspose2d`. According to the pytorch docs: > The padding argument effectively adds dilation * (kernel_size...

After some discussion, it seems like this unintuitive behavior of `lax.conv_transpose` (in which string paddings are interpreted as the paddings for the ordinary convolution by integer paddings are interpreted as...

Hi @nilq - we'd like to finally land this PR on top of all the recent changes to nnx! We just merged support for SpectralNorm (#4623), and your WeightNorm code...

The PR #4948 dealt with calculating flops for each method in a `tabulate` call. To get the number of flops, we need to compile the method first. But certain methods...