Why discriminator loss need to be separately backward in pix2pix train.py

[jehovahxu]

I tried backward errD((errD_fake + errD_real)*0.5) and original code only backward total generator loss, but in this code, the result worse than separate backward errD_fake and errD_real.