After downloading the seamless-m4t-large model file, How to load and inference them locally?

[jiafuz]

After downloading all the files listed in the Files and versions? How to load them locally? When passing the folder name, I got the error:

ValueError: name must be a valid filename, but is './pretrained_models/seamless-m4t-large' instead.

The code I used are:

model_path = './pretrained_models/seamless-m4t-large' vocoder_path = './pretrained_models/vocoder' translator = Translator(model_name_or_card=model_path, vocoder_name_or_card=vocoder_path, device=torch.device('cuda:0'), dtype=None)

I also tried the following code, but no luck: model_path = './pretrained_models/seamless-m4t-large/multitask_unity_large.pt' vocoder_path = './pretrained_models/vocoder/vocoder_36langs.pt' translator = Translator(model_name_or_card=model_path, vocoder_name_or_card=vocoder_path, device=torch.device('cuda:0'), dtype=None)

[Ori-Replication]

Having the same problem. I can not load local models with translater function either. It always throw a ValueError like name must be a valid filename when I include slash in the string.

[huangwei2913]

Having the same problem, the model pt file is over 10GB, and there is no demo about how to launch with local pt file

[zachNA2]

114

Check my comment

[dsp6414]

How to use local checkpoint for inference? #99

see the topic closed #99

[mohsenoon]

@jiafuz

After downloading all the files, you should load model and vocoder retrieve_card and save them in the desired path:

import fairseq2, joblib

model_retrieve_card = fairseq2.assets.asset_store.retrieve_card('seamlessM4T_large') vocoder_retrieve_card = fairseq2.assets.asset_store.retrieve_card('vocoder_36langs') joblib.dump(model_retrieve_card, 'seamlessM4T_large_retrieve_card.pkl') joblib.dump(vocoder_retrieve_card, 'vocoder_36langs_retrieve_card.pkl')

Then do this to load the model and vocoder locally:

from seamless_communication.inference import Translator

model_retrieve_card = joblib.load('./seamlessM4T_large_retrieve_card.pkl') model_retrieve_card.metadata['checkpoint'] = './pretrained_models/seamless-m4t-large/multitask_unity_large.pt'

vocoder_retrieve_card = joblib.load('./vocoder_36langs_retrieve_card.pkl') vocoder_retrieve_card.metadata['checkpoint'] = './pretrained_models/vocoder/vocoder_36langs.pt'

translator = Translator(model_name_or_card=model_retrieve_card, vocoder_name_or_card=vocoder_retrieve_card, device=torch.device('cuda:0'), dtype=None)

[dsp6414]

已收到！