Skip to content

fix: added Setup.sh to be MacOS Friendly - #10

Open
Ntropy86 wants to merge 1 commit into
Neuroprosthetics-Lab:mainfrom
Ntropy86:main
Open

fix: added Setup.sh to be MacOS Friendly#10
Ntropy86 wants to merge 1 commit into
Neuroprosthetics-Lab:mainfrom
Ntropy86:main

Conversation

@Ntropy86

@Ntropy86 Ntropy86 commented Oct 5, 2025

Copy link
Copy Markdown

For Macos users made the setup.sh file more friendly to use.
Issue Fixed: In Apple Silicon Macs, The CUDA version does not work, it requires MPS version.
Extra Fixes:

  • Added Auto detection OS and installs the correct files
  • Followed Literal Definitions standards
  • Added Comments

@nckcard

nckcard commented Oct 6, 2025

Copy link
Copy Markdown
Contributor

Hey, this looks great, thanks. One concern: Pytorch CTC loss is not supported on MPS (as far as I know), which would prevent someone using this repo from running the baseline model training on MPS. Have you tried doing so on MPS / is there anyway around this limitation other than using the much-slower CPU instead?

@Ntropy86

Ntropy86 commented Oct 8, 2025

Copy link
Copy Markdown
Author

@nckcard You are correct. It is not supported (I looked around and tried to see if there are any other solutions for CTC, did not find anything promising. So what I can do is add a WARN for macos users and make the run Hybrid approach using MPS normally but Running CTC on CPU.

Upon reading more, I believe we can use Attention Loss however that would be a massive architectural change and not a direct implementation of the paper.

@nckcard

nckcard commented Oct 9, 2025

Copy link
Copy Markdown
Contributor

If you can make it run hybrid, that seems fine to me. As is, I think the training script will crash if we try to train with CTC loss via MPS. Attention loss also seems very interesting, if you want to PR that separately I would be supportive! Thanks.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants