Reproduction
I am currently trying to reproduce the results using your supplied hyperparameters. While it works great on imagenette or imagewoof, I can really suspect values on in1k. I suspect, that I am using the wrong version of in1k:
I downloaded the dataset directly from kaggle as suggested by the official website, I noticed that compared to imagenette/imagewoof the directory layout is different. For the validation set, all images are in one directory and the labels are in a different subfolder saved in a csv. Do I have to use a different in1k version or how can I manually adapt the dataset?
I found a script to help me organize my validation directory:
https://github.com/jiweibo/ImageNet/blob/master/valprep.sh
I do wonder though. Why was there no error / warning. Did all images just get assigned to a "default" class? Might be a good addition to throw a error or warning when images have no label.
@tony0278611 I'm not sure, valprep should work if you have the original imagenet validation tar file, I vaguely recall the one on kaggle being a bit different than the original. Each validation sample needs to be put in a folder with the synset name that mirrors the train folder structure.
It did indeed work, Ill close the discussion. For future, it could be a nice addition to add a warning when abnormal datasets are detected (e.g. every sample of the same class)