Neil Stoker
AI & ML interests
Recent Activity
Organizations
This is a great start @benjamin-paine . However won't it trip up on relatively common cases such as:
- Hello Dr. Jones.
- The answer is 10.45s.
- Well... that needs extra attention.
- They wanted pets, small animals etc. to be brought in for checks.
- The filename is Report2023.xls.
- You should look at www.example.com.
I'm AFK right now so can't verify for sure but given this seems to use a basic sequence of regex's it seems likely it won't cater for these sorts of things (unless there's some subtle aspect I'm overlooking)
Something like SaT would be too heavy weight. Maybe a compromise closer to PySBD but in JS would work well here?
Generate 10 seconds of speech in ~1 second for $0.
What will you build? ๐ฅ
webml-community/kokoro-webgpu
The most difficult part was getting the model running in the first place, but the next steps are simple:
โ๏ธ Implement sentence splitting, allowing for streamed responses
๐ Multilingual support (only phonemization left)
Who wants to help?
I built a web app to interactively explore the self-attention maps produced by ViTs. This explains what the model is focusing on when making predictions, and provides insights into its inner workings! ๐คฏ
Try it out yourself! ๐
webml-community/attention-visualization
Source code: https://github.com/huggingface/transformers.js-examples/tree/main/attention-visualization
Quick update: I was able to confirm that I can get some other WebGPU demos working in Chrome on this particular laptop.
Eg managed to get this: https://toji.github.io/webgpu-clustered-shading/ running.
I made sure Vulkan was installed and gives sensible output in the terminal from: sudo vulkaninfo and then I made sure it was enabled in Chrome (the flag options for Vulkan won't appear in Chrome flags without also setting "Temporarily unexpire M130 flags") along with the flag for enable-unsafe-webgpu.
Unfortunately no change in the error messages from the Moonshine demo link.
@Xenova - Happy New Year! ๐
With the demo link above I'm seeing issues when running on Linux with Chrome on a laptop which make me wonder if the WASM fallback is working in my case.
I get this message:
An error occurred
Requested device not found
The message about the requested device is quickly replaced with:
Uncaught Error: no available backend found. ERR: [webgpu] Error: Failed to get GPU adapter. You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.
I've tried setting the Chrome flags, but that doesn't seem to make any difference.
Is there anything you'd recommend I check to try to diagnose the issue? Or a way to force the WASM fallback?
I'd previously had great results on my main PC (but this has a 3090 GPU) and on mobile (Android Pixel 8 Pro which also has GPU support) and it was only when I went to test on some slightly older devices I came upon this issue.
Problem laptop has Chrome Version 131.0.6778.204 (Official Build) (64-bit), it's running Arch Linux. CPU is an i7-7500U from 2017 which has Intelยฎ HD Graphics 620.
No examples in readme
Wonderful news!
Any indication on how easily v2 code etc will work under v3 and what the main changes are? (am guessing there'll be a blog post in due course ๐)
Install it from NPM with:
๐๐๐ ๐ @๐๐๐๐๐๐๐๐๐๐๐/๐๐๐๐๐๐๐๐๐๐๐๐
or via CDN, for example: https://v2.scrimba.com/s0lmm0qh1q
Segment Anything demo: webml-community/segment-anything-webgpu
Any chance of some pointers for how to use the model in plain JavaScript? I had a look at the react/ts and wasn't sure where to begin (I'm not the strongest js coder so I may be missing something obvious!)
Amazing progress, thank you!
For anyone trying this on Chrome on Linux, you may need to set some flags: a comment appears in the Console when the model loading gets stuck suggesting a startup flag, but you can just switch those with chrome://flags directly and relaunch - for me, I needed to enable both Vulkan and Unsafe WebGPU, and then it works (seriously fast I should note!!)