Passing a filepath for models

[CopaceticMeatbag]

I'm trying to use the facebook m2m 1.2B model, but python can't download it in time before it times out and crashes the download. Is it possible to pass a filepath to a model I've already downloaded? It looks like it's possible in dl-translate, but I'm not sure how I'd access that model_or_path feature via the subsai options.

[abdeladim-s]

Hi @CopaceticMeatbag,

I think you can just pass the model path instead of the model name in the translate method.

Have you tried this and didn't work ?

[CopaceticMeatbag]

Thanks @abdeladim-s - I get an error: Unable to infer the model_family from "C:/Users/Administrator/m2m100_1.2B/". Try explicitly setting the value of model_family to "mbart50" or "m2m100"

The code I'm using is as below - I tried several variations on the path string with back/forward slash or raw string, no end slash, etc:

import os

from pathlib import Path

import pysubs2

from subsai import Tools

subtitles_file = 'V:\TestAutoSubs.srt'

subs = pysubs2.load(subtitles_file)

translation_model = 'C:/Users/Administrator/m2m100_1.2B/'

source_language = 'English'

target_language = 'Spanish'

format = 'srt'

translated_file = f"{subtitles_file}-{source_language}-{target_language}.{format}"

translated_subs = Tools.translate(subs, source_language=source_language, target_language=target_language, model=translation_model)

translated_subs.save(translated_file)

print(f"translated file saved to {translated_file}")

[abdeladim-s]

I think you will need to pass a Pytorch model : i.e something that ends with pth ! Where did you download the model ? from the code above, the translation_model var seems to be a folder!

[CopaceticMeatbag]

Ah, I thought we had to point to the whole directory! I downloaded from here: https://huggingface.co/facebook/m2m100_1.2B/tree/main

The script tries to download the pytorch_model.bin file (which is where it crashes), I also tried making the translation_model point to C:/Users/Administrator/m2m100_1.2B/pytorch_model.bin with the same results.

From what I can tell looking at dl-translate, it accepts a path to a directory containing the model files, but I think the integration into subsai bypasses this and uses the translate function directly. I'll have a bit more of a crack at figuring this out today!

edit: I'm looking at this snippet from dl-translate readme:

By default, dlt.TranslationModel will download the model from the huggingface repo for mbart50 or m2m100 and cache it. It's possible to load the model from a path or a model with a similar format, but you will need to specify the model_family:

mt = dlt.TranslationModel("/path/to/model/directory/", model_family="mbart50")
mt = dlt.TranslationModel("facebook/m2m100_1.2B", model_family="m2m100")

[abdeladim-s]

yeah basically it should work with the pytorch_model.bin file I have checked the dl-translate source code as well, and from what I can see you just need the model name or the path to a pretrained model, In subsai I just take the model var and pass it directly to that TranslationModel class, so it should be the same! There is also a create_translation_model which returns an instant of that class, you can use it in the translate method as well.

[abdeladim-s]

oh, so the model_family is not inferred automatically ? Please try the dl-translate snippet and let me know if it works ?

[abdeladim-s]

@CopaceticMeatbag I have added the model_family to the translate model, try to reinstall the latest commit and give it a try. Hope it will work now!

[CopaceticMeatbag]

You've only gone and bloody done it!! Nice work, all works perfectly :) thank you very much!