Brett Young
Brett Young
just wanted to follow up on this! Ive tried some fc layers and I notice complete underfitting of my data on cifar10. However a simple model with mainly conv layers...
i added it to my repo. check it out on my profile
just the final finetuning stage
Thanks!
I confirmed it does not work. This would a great addition.
Anyone able to find a solution?
You all able to solve this? - Brett
Well, I tried training with IJEPA on cifar10 and then using the pretrained model to fine tune on cifar10, using the labels during fine tuning (with just the target encoder)....
Around 10 or so, as after that train and validation loss began to rise.