Pix2Pix paper and get_gen_loss function from assingment C3W2B

Course Q&A Generative Adversarial Networks (GANS) Apply Generative Adversarial Networks
1

Hello,

I’m trying to understand the 4 equations described in the Pix2Pix paper, section 3.1 - Objective, and its implementation in the C3W2B assignment.

  • Equation 4 is the step 4) described in the assingment, inside the cell UNQ_C2 that implements the get_gen_loss function

  • I think equation 3 is the reconstruction loss, step 3) in the cell UNQ_C2, and in this case y in the equation refers to real in the get_gen_loss function parameters, with G(x,z) as the fake image generator, x refers to the condition parameter of the get_gen_loss function, and z is the noise

Let’s assume the interpretation above is correct.

Other than that, I’m having difficulty in understanding the equation 1) and how the terms, and values x,y,z are interpreted in the get_gen_loss function.

Many thanks in advance for any explanation or links to clarifications elsewhere.

2

I actually figured it out by re-reading the paper and re-watching the videos.

Actually, these equations refer better to the whole training procedure that we have in the Pix2Pix Training.

At the section 3.3 of the paper, the authors explicitly say to train the discriminator and the generator following the standard GAN’s train procedure. We then do the adaptations of passing the condition input as a parameter to the discriminator, and adding the L1 regularisation (multiplied by lambda) to the generator BCE loss.