Adding noise to a PIL or numpy image with numpy/pytorch

Question

I'm trying to reproduce this interesting paper: "Deep Image Prior". https://arxiv.org/abs/1711.10925

Now, I saw there is a colab version that I can run and it works pretty nice: DIP

The only think missing for me is this graph:

Now, I think I found the place in the code where this process is taking place: DIP restoration

Where is the noise addition?

Edit: The noise addition happens here:

Main loop

def closure(): global i, psrn_masked_last, last_net, net_input if reg_noise_std > 0: net_input = net_input_saved + (noise.normal_() * reg_noise_std)

Why, if I remark these lines I still see a noised image being plotted?

Thanks!

Answer 1

The noise addition happens here:

Main loop

def closure(): global i, psrn_masked_last, last_net, net_input if reg_noise_std > 0: net_input = net_input_saved + (noise.normal_() * reg_noise_std)