crowsonkb / vgg_loss Goto Github PK

Hi, thanks for your great project. I have a question about the initialization in vgg_loss_demo.py

where the torch.rand_like would generate a tensor with uniform distribution over [0, 1] and the result is further divided by 255, leading to small values. Does this kind of initialization help the optimization since my output is all black with other initialization or learning rate which I am not sure about.