How gradient is back propagated in WGAN-GP if D set to false?

falmasri · February 15, 2018, 12:09pm

How is the gradient is backpropogated from the critic to the generator if we set

for p in netD.parameters():
        p.requires_grad = False  # to avoid computation

in the generator training?
in this example

github.com

caogang/wgan-gp/blob/master/gan_cifar10.py

import os, sys
sys.path.append(os.getcwd())

import time
import tflib as lib
import tflib.save_images
import tflib.mnist
import tflib.cifar10
import tflib.plot
import tflib.inception_score

import numpy as np


import torch
import torchvision
from torch import nn
from torch import autograd
from torch import optim

This file has been truncated. show original

jpeg729 · February 15, 2018, 12:15pm

That just tells pytorch not to calculate the gradients with respect to the parameters of D. The gradient still flows back via the inputs of D to the parameters of G.

falmasri · February 15, 2018, 12:22pm

isn’t the gradient of G dependent on the gradient flow from the D ?

jpeg729 · February 15, 2018, 12:37pm

Yes, but it only depends on the gradient with respect to the input of D, it does not depend on the gradient with respect to the parameters of D.