gradfeat20

gradfeat20 copied to clipboard

XiaTianqi

Thanks for this great work! Why jvp has to be calculated via NTK parameterized linear layers? What if just using vanilla Conv2d for the jvp calculation? https://github.com/fmu2/gradfeat20/blob/master/src/model.py#L106

chenwydj

----- Training phase ----- Traceback (most recent call last): File "./src/benchmark.py", line 228, in main() File "./src/benchmark.py", line 219, in main args.niter, args.stepsize, losses, it=it) TypeError: train() missing 1 required...

toosi

## Cannot load .pt model files.

1

I tried to run the script, but i cannot use torch.load('name.pt') to load the pre-trained model.

YujunLiao