Curious why this creates 2GB .ckpt files

[EricBintner]

I'm curious what would cause this to create a 2GB .ckpt file when the HuggingFace v1-4 file is 4GB? Is this expected?

[TheLastBen]

The 2GB is in half precision which is what is widely used, the quality is exactly the same and a faster loading

[hashnag]

For the record for anyone that's curious: it's not just this implementation, the JoePenna Dreambooth also produces 2GB checkpoints (and probably most others).

[SAC020]

Trained a model on standard google colab GPU, it maxed out GPU memory but run without crashing.

However, the resulting model is just 2GB, and the quality of generated images is poor.

When I train on "premium" GPU, I get 4GB files and the quality is better.

Is there a correlation?

Thx

[TheLastBen]

In the cell 'start dreambooth', click twice and change fp16 = True to fp16 = False and see if there is any difference, if so, I'll add back the checkbox

[SAC020]

I have obtained much better results, but I've also increased the number of steps, so maybe not a like-for-like experiment. Regardless, I think it's worthwhile adding back the checkbox

[giteeeeee]

increased the number of steps, so maybe not a like-for-like experiment. Regardless, I think it's worthwhile adding

maybe try using the exact same number of steps just to compare in parallel? I'm really curious