search

rohitgandikota / erasing

Erasing Concepts from Diffusion Models
https://erasing.baulab.info
MIT License
559 stars 34 forks source link

Some questions and potential improvement on the code #7

Closed xjtupanda closed 1 year ago

xjtupanda commented 1 year ago

Good code and is easy to understand!

  1. These two actually are the same, I'm not sure why it's called twice. I think copy operation might work as well and save some computation. https://github.com/rohitgandikota/erasing/blob/724514a05b55e047dc783d1e42d444989230620d/train-scripts/train-esd.py#L221-L222

  2. These two lines are unnecessary. I've checked it and their requires_grad property is already set to False because of the previous with torch.no_grad() context. https://github.com/rohitgandikota/erasing/blob/724514a05b55e047dc783d1e42d444989230620d/train-scripts/train-esd.py#L244-L245

rohitgandikota commented 1 year ago

Yes you are right! Although getting token embeddings is usually not expensive. But yes, it will improve the efficiency, given they are the same. Thanks for catching these! we will take a note of this for our next update.