metavoice-src
metavoice-src copied to clipboard
metavoiceio
is there a way to easily load the model ?
I'm working on automating the huggingface repo to automatically call the model and I would like ask for some help with creating a custom HF friendly architecture for this. All i need is a simplified version of creating the models in python. This is nice blogpost about creating custom architectures that I wrote and could be used as a reference https://huggingface.co/blog/not-lain/custom-architectures-with-huggingface
So sorry @not-lain for some reason this issue fell through, sorry about the delay! I'm not fully sure what's needed here, i'll check out the blogpost!
in the meantime, you can check the fast_model.py and fast_inference.py which contains slightly cleaner code, in case that helps?
it's ok, don't worry about it, also i have been swamped with some work lately, also thanks for the info, i only needed some pointers to help me speed up the process of creating the architecture, i'll try to contribute when i get some time off. also since this is related to huggingface and not github i'll let you decide if you want to close this one
Hey @not-lain, this issue has been open for a while, sorry for not revisiting earlier! Did you end up having a look?
@lucapericlp aplogies, I have been swamped a little lately and I think I found a really interesting solution. huggingface updated their PyTorchModelHubMixin class and now you can add your model more easily with this class, I'll contact them later to add a load button for your library (this repo) as soon as the PR is merged. the final lookout would be something like this :
- model https://huggingface.co/pyp1/VoiceCraft_830M_TTSEnhanced
- repo https://github.com/jasonppy/VoiceCraft
will create a PR using the same concept ASAP
The reason why I recommend using PyTorchModelHubMixin is because this repo has multiple dependencies that are distributed across multiple files and will require lots of work to adapt it to a single file, plus you will have your own library integrated with huggingface so it will help you stand out even more.
Also for future finetuned models, you can keep track of all the community models since you will have they ability to filter all the models using the metavoice-src library which can be used as a tag in huggingface in the following list.