search

caoscott / SReC

PyTorch Implementation of "Lossless Image Compression through Super-Resolution"
MIT License
1.05k stars 94 forks source link

overfit models yeild domain specific image compressors? #11

Closed lidless-vision closed 3 years ago

lidless-vision commented 3 years ago

excellent work! this is awesome.

Ive been looking at this project and reading the paper and i think i have an experiment id like to run. being a lazy scientist, i'd like your opinion of the idea. my idea was also partially inspired by the paper: COIN: COmpression with Implicit Neural representations (https://arxiv.org/abs/2103.03123)

i'm sorta a newb here so... sorry if this is dumb, but heres the gist of it:

in data science we are typically building models that generalize to unseen data, which you've done very well. but in the use case of storing large amounts of domain specific images (like a database of cat pictures). it would seem that a little overfitting would benefit the performance of the model. also, the model could be continuously trained as new images are added to the database?

does anyone else think this might work?

lidless-vision commented 3 years ago

ok wow, i guess they did mention dataset compression in the paper, im running it right now.