search

libffcv / ffcv

FFCV: Fast Forward Computer Vision (and other ML workloads!)
https://ffcv.io
Apache License 2.0
2.81k stars 180 forks source link

fill parmeter in translate has no effect #175

Closed tfriedel closed 1 year ago

tfriedel commented 2 years ago

When using the translate operation, the empty part of the result image should be filled with the tuple provided with the "fill" argument. It's currently not being used and thus the background will always consist of (0,0,0).

GuillaumeLeclerc commented 2 years ago

Hello @tfriedel, Which transform are you referring to ? I don't think FFCV comes with a translate transform does it?

tfriedel commented 2 years ago

this one: https://github.com/libffcv/ffcv/blob/main/ffcv/transforms/translate.py

mujjingun commented 2 years ago

When using the translate operation, the empty part of the result image should be filled with the tuple provided with the "fill" argument. It's currently not being used and thus the background will always consist of (0,0,0).

It's actually worse than that, it's uninitialized and once in a while it contains a random value other than 0.0!

GuillaumeLeclerc commented 2 years ago

I didn't know we had that augmentation. Sorry about that, I'll try to fix that soon, but if someone wants to go ahead and make a PR it should be pretty straightforward

andrewilyas commented 2 years ago

This should be fixed in #184 !

andrewilyas commented 1 year ago

Now fixed on PyPI!