ymalitsky / adaptive_gd Goto Github PK

Hello, excuse me for asking out of the blue.

I am very interested in your paper and am now thinking about its development.

I have a question about the code in this repository, which I think is an error.

In adaptive_GD/pytorch/optimizers.py,

Line 39, "grad_dif_norm += (d_p - prev_d_p).norm().item() ** 2" is not be, "grad_dif_norm += (d_p - prev_d_p).pow(2).sum().item()"?

and also Line 40, "grad_dif_norm += (p.data - prev_p.data).norm().item() ** 2" is not be, "param_dif_norm += (p.data - prev_p.data).pow(2).sum().item()"?

This is because in torch.Tensor.norm(), all the elements of the tensor are squared, added together, and the square root is taken.

In the original code, the Euclidean norm of the difference of each layer parameter (or gradient) is found, the result is squared, the squared Euclidean norm of all layers is added, and the square root of the result is taken.

This is not the same as the formula in the paper, is it?

Lastly, I am sorry if my English is not good enough to convey my intentions well.

I hope you understand that I am not malicious in my questions and that I am asking them out of simple doubt and interest in your research.