[Feature request] Using a single Torch binary for all the model related content

🚀 Feature Description

We consider using a single torch binary file as the model release format as opposed to the current zip files compressing all the other files used by a model at inference.

It’d help to keep things more compact and easy to manage by the ModelManager such as avoiding setting the relative paths of the dependencies (normalization stats, speaker IDs, etc.) in the config.json.

Solution

We’ll use a single model.pth that includes all the data dependencies and the model itself. Then you can access the content as python attributes like

import torch
chkp = torch.load("model.pth.tar")
speaker_ids = chkp['speaker_ids']
speaker_encoder = chkp['speaker_encoder']
...
model = chkp['model]'

So let us know what you think about this proposed change under this issue.

[erogol]

How would fine-tuning work? Say I want to FT an existing model with another dataset. I need to load the 1gb checkpoint in python do chkp['config']['datasets'][0] = BaseDatasetConfig(name="vctk", meta_file_train="", path="../VCTK/") then save a 1Gb copy of it if I don't want to overwrite the previous one, and then continue with that checkpoint? Instead of editing the config.json

@WeberJulian How about if we keep config.json separate?

How about having this compact format for inference only.

Nothing stays inference only 😄. Also, it creates a decoupling that we need to manage, maintain and book. IMO a single format that works everywhere (train, inference, fine-tune) is a more convenient and reliable approach.

[erogol]

Yep only a .pth file (same file but we don’t use .tar any more)

Yes we 'll update all the models the same way