CNN hessian eigenvalues are too large?

dgcnz / dl2

Code for "Effect of equivariance on training dynamics"

1 stars 0 forks source link

Open dgcnz opened 2 months ago

dgcnz commented 2 months ago

Even after 24 epochs, we have many times larger magnitudes for CNNs than [Park 2022] (warmup=5 epochs)