Clay Mullis
Clay Mullis
Another thing to consider - the loss curve. The lowest amount of loss possible on your dataset will never change. Your dataset, however, can. So will loss curves be different...
@lucidrains How easy/possible would it be to use a custom mask "structure"? perhaps the target format could be whatever the typical coco-style segmentation data looks like; and then maybe you...
Once we get that working - this would be a great PR to include in the larger DALLE-pytorch codebase. It is part of the original paper - after all.
I have _one last_ idea, but I'm now thinking it may not work actually because it works with git clone, but not necessarily pip ```sh cd clip_rerank; git status #...
The authors of this repo should _really_ fix this stuff. It was _deeply_ frustrating trying to set up this project. _Deeply_. I've never seen such a terrible codebase in ~7...
Probably still a good idea to update the README.md accordingly, no? I would do it myself but I suspect it won't be merged.