Dongwan Kim

I've updated the code. ```python from copy import deepcopy import numpy as np import torch from ffcv import Loader from ffcv.fields.basics import IntDecoder from ffcv.fields.rgb_image import SimpleRGBImageDecoder from ffcv.loader import...

@andrewilyas I am not aware of the CuDNN simultaneous batch norm bug - could you link me to a reference? If it is a BN related bug, what I don't...

@GuillaumeLeclerc, thanks for looking into this. It does seem to be a very weird bug. Please let me know if there are any ways I can help

@ByungKwanLee the tests above were conducted on single-gpu, so I don't think SyncBN applies in this case.

@andrewilyas @GuillaumeLeclerc Just tested the code - this seems to be fixed in the `fix-198` branch. Doesn't seem to be merged in the `v1.0.0` branch yet. Thanks for the fix!...

@vfdev-5 Thanks for the reply! As you mentioned, upsampling from 64 --> 224 is not a usual case for training, but it was part of the baseline method (not sure...

@vfdev-5 thanks for the explanation and the updates. The updated version seems to be a great performance boost (for upsampling)! Also, did not know `torchvision.io.read_image == cv2.imread != PIL read`....

I see - I'll have to keep an eye out for updates in the future 😄

Issue seems to be persistent. You could implement your own distributed average meter or use something like [TorchMetrics](https://torchmetrics.readthedocs.io/en/stable/pages/lightning.html) which supports metrics across GPUs - I haven't used this myself but...

Willing to contribute (examples and perhaps installation?), but I have a lot of things going on rn so will need a few weeks to start working