Can you add to an already trained model? #193

This isn't really an issue as much as a discussion topic...Once a model is trained, can you go back later with new datasets and add to it? Or do you have to recreate a new dataset with old and new, and then retrain the whole thing?

I'm trying to build some character voiced, english-speaking with accents, and trial and error takes a long time.

This isn't really an issue as much as a discussion topic...Once a model is trained, can you go back later with new datasets and add to it? Or do you have to recreate a new dataset with old and new, and then retrain the whole thing? I'm trying to build some character voiced, english-speaking with accents, and trial and error takes a long time.

You can go back and "add" to finetune by reusing the finetunes existing weights, yeah. I haven't done it specifically for TorToiSe, but I don't see it wouldn't work outside of "quality" concerns.

Your best bet is to use the previous finetuned model as your Source Model, and it will start training from iteration 0. I would heavily advise to move the model out from the ./training/{voice}/finetunes/ folder and into ./models/finetunes/, as the folder would get renamed / "backed up" when you start training and breaks things.

You can go back and "add" to finetune by reusing the finetunes existing weights, yeah. I haven't done it specifically for TorToiSe, but I don't see it wouldn't work outside of "quality" concerns. Your best bet is to use the previous finetuned model as your `Source Model`, and it will start training from iteration 0. I would heavily advise to move the model out from the `./training/{voice}/finetunes/` folder and into `./models/finetunes/`, as the folder would get renamed / "backed up" when you start training and breaks things.

Thanks for that. I'll give it a shot.
When you say "Source Model" is that the Autoregressive model under settings?

Thanks for that. I'll give it a shot. When you say "Source Model" is that the Autoregressive model under settings?

Crap, nevermind. I see it in the Config settings.

I did an additional 500 iterations with a new dataset of the same voice over the original model I trained as the Source Model... No change at all. Getting an accent to shine through is difficult. Not sure if the dataset is too small (300 wav files)... or if I'm not training long enough. I set it for 500 and it stopped at 493 as if it was just done.

Occasionally I get a British accent when I don't want one (and there's nothing in the dataset to cause one)...and then today I trained a voice with a British accent and the voice had no accent. Wish this was more of an exact science.

I did an additional 500 iterations with a new dataset of the same voice over the original model I trained as the Source Model... No change at all. Getting an accent to shine through is difficult. Not sure if the dataset is too small (300 wav files)... or if I'm not training long enough. I set it for 500 and it stopped at 493 as if it was just done. Occasionally I get a British accent when I don't want one (and there's nothing in the dataset to cause one)...and then today I trained a voice with a British accent and the voice had no accent. Wish this was more of an exact science.

Labels Milestones

Can you add to an already trained model? #193