Training and Evaluation of pix2pix

The training of pix2pix is very similar to conditional GANs. Here, we will discuss the training and evaluation process of pix2pix model.

Training of pix2pix model

When training the discriminator network, a pair of real data and a label should be mapped to 1, whereas a pair of generated data and a label (that fake data is generated from) is mapped to 0. When training the generator network, the gradients are passed through both the discriminator and generator networks when the parameters in the generator network are updated. This generated data and the label should be mapped to 1 by the discriminator network. The major difference is that the labels are image-wise in CGAN and pixel-wise in pix2pix. This process is described in the following diagram: