-
https://git.ecker.tech/ aims to provide a place to share my efforts while maintaining true ownership of my code, as I do not trust GitHub.
XMR: 4B9TQdkAkBFYrbj5ztvTx89e5LpucPeTSPzemCihdDi9EBnx7btn8RDNZTBz2zihWsjMnDkzn5As1LU6gLv3KQy8BLsZ8SG
- Joined on
2022-10-10
7911b780e4
consistent-ish VRAM usage, finally, maybe, hopefully, by explicitly deleting non-arg tensors when explicitly used
0d5cc441e4
reverted prom length strategy back to duration +- offset, and added GC every global step (every X grad_accum size) because I really, really need to get my batch size to 12.
I meant to send this earlier, but I kept getting sidetracked. Oops.
If you still need more data, I'd recommend checking out the VoxCeleb dataset.
Mmm, my only qualm with that is:
…
149aaca554
fixed the whisperx has no attribute named load_model whatever because I guess whisperx has as stable of an API as I do
AttributeError: module 'whisperx' has no attribute 'load_model'
Oh, it seems that also affects the master branch and not just the v3 one. I'll try and push a fix that doesn't outright break…
572f964ec6
modified algo to adjust prom length to fit prom_length + resp_length = max(prompt_duration, min_prompt_duration)*75), because I cannot find a good and easy way to have a little dataset for large batch size, and a large dataset for small batch size
Forgot to mention, should be fixed in commit e416b0fe6f186f8d6e69c79664d996361b5fdfb8.
Apologies, I'm not sure why/how it worked in my install.
5003bc89d3
cleaned up brain worms with wrapping around gradio progress by instead just using tqdm directly (slight regressions with some messages not getting pushed)
c90ee7c529
removed kludgy wrappers for passing progress when I was a pythonlet and didn't know gradio can hook into tqdm outputs anyways
c90ee7c529
removed kludgy wrappers for passing progress when I was a pythonlet and didn't know gradio can hook into tqdm outputs anyways
Did a better bandaid in commit 09d849a78f3df96e7d1e522b779bd7e71d16f3af (be sure to also update submodules), but I'll still clean up the kludge in my TorToiSe fork itself.