How to continue training without replacing the previous training?

[molo32]

How to continue training without replacing the previous training?
I have images of a face divided into part1 and part2

first I train part1 of images of face.

when I finish training and do a driver, you can see the face of the images in part1

Then I want to add the images from part 2.

then I load the pth and train from there with the images of part 2

The problem is that when I make a driver, only the images of part2 are chosen, it is as if part2 were to overwrite part 1.

What I expected was to have more variety of expressions, so expressions from part 1 and part 2 are chosen.

How to avoid overwriting images or expressions from a previous training session?

[shrubb]

Sorry, I didn't get it at all, what do you mean by "overwriting images"? And why do you want to train (or fine-tune?) two times. If you have two datasets, just fine-tune to them independently using a single meta-learned checkpoint

[molo32]

I want to train two sets of the same person separately because if I load the whole dataset,
cuda out of memory error.
so to avoid that error I split the dataset in A and B and first train a datasetA then datasetB.

dataset are images-cropped generated by preprocess data.py

images are expressions or the face of a person.

driver is to take a video driver, a chekpoint and produce a video with driver.py

When selecting images I mean which images from the data set are chosen to make the output video with driver.py

By overwriting images I mean overwriting expressions or faces.

If data set A has different illumination to data set B, if I refine a meta-learned model with data set A with python3 train.py, then I repeat but with data set B, when making a driver only images are seen with the illumination of B, then B overwritten A.

I don't want to train, I want to tune.
I want to fit two dataset A and B independently, I select the latent-pose-release.pth) to train with the first dataset A, in DATASET_ROOT = path set A, I run python3 train.py
At the end of training it gives me a checkpoint.pth, if I do a driver with that checkpoint the expressions of data set A are seen, then I load that chekpoint.pth to continue training from there to data set B, I finish training and
it gives me another chekpointB.pth, but when I do a driver with the chekpointB, only images are selected from the last dataset B and not from dataset A, that's what I mean by overwrite.

[shrubb]

because if I load the whole dataset, cuda out of memory error.

This means that you're doing something wrong: GPU memory doesn't depend on the dataset size. Just use smaller batches. For example, with a batch size of 1 you can fine-tune on as many images as you want.

As I understood, you're trying to fine-tune a meta-learned checkpoint to dataset A, then take that fine-tuned model and fine-tune it further to dataset B. Well, we never tried that. I don't know if that will even work -- that's a research question. You'll probably need to modify the code for that, and it's entirely at your risk, I'm afraid I can't help here.

[molo32]

ok I understand, another thing, can I make the checkpoint smaller?, output is always 1 gb size.

[shrubb]

Yes, it's not hard (just don't include discriminator, embedder, optimizer state etc. in the checkpoint) but for that you'll have to modify the code yourself.