pix2pix train model correction

In train.py of the Pix2Pix research paper implementation on line 34, the code should be opt_disc.zero_grad()

The current one also seems to work, but I have a feeling that this correction would work better. Please let me know if I'm mistaken.