acsweet
acsweet
It's in the PR, but for reference: - `psnr`, `ctc_loss`, and `ctc_decode` implemented - `ctc_decode` with `beam_search` is not the most efficient without a `unique` function from mlx - `norm`...
Once these latest two PRs are merged, I'd like to try merging the master branch into mlx (if that's okay)
Merged the Keras `master` branch into `mlx` and patched a few files for pytest to work Going to add new functions and check tests starting with `nn.py` Will start adjusting...
I'm currently working through getting the layer tests to pass (`keras/src/layers`) including updates to `ops` functions as needed, and skipping unsupported tests. Updates to `ops` are mostly in `math`, `nn`,...
We're continuing to work through the remaining functions in `linalg` and `numpy`, and adjusting tests where appropriate. I'm marking quantize tests as unsupported by mlx for now. The quantized matmul...
Sorry Awni, I think I spoke too soon! I'm still wrapping my head around mlx's quantization and Keras' quantization related methods (and quantization in general). It looked like to call...
I think I'll see if I can work on a fix for this.
If I have an idea should I submit a PR? And we can discuss what you think?
Thank you! Let me see what I can do along those lines. The tests definitely make sense, I wasn't sure what performance benchmark made sense along these lines. If the...
I'm still working on some good simple tests, I ran in to a few more errors with the prior proposed changes. But I wanted to ask what you thought of...