Doubt in gradient penalty image interpolation

Course Q&A Generative Adversarial Networks (GANS) Build Basic Generative Adversarial Networks
1

Sir I didn’t understood why we are using random interpolated image for calculating regularization term and how random interpolation is going to help in calculating gradient penalty can you please help me sir

Thanks in advance

2

Hey @enuguru_2002,
First of all, let me tell you that random interpolation doesn’t help in calculating gradient penalty. Instead, it helps to ensure that the gradient penalty term (which is already calculated) is encouraged to be 1L continuous. If you are confused with this much, then I urge you to rewatch the videos as this thing is clearly explained in the videos.

Now, it must be clear to you that the regularization term is added in order to encourage 1-Lipschitz continuity, i.e., the gradient of the derivatives at all times should be <= 1. Now, if you think about it, the critic is basically a model which sees both real and fake images. So, instead of calculating this regularization term for both real and fake images separately, it makes sense to use random interpolation of both, which will ensure that the regularization term represents both real and fake images simultaneously.

According to me, this is how random interpolation helps in calculating the regularization term. I hope this helps!

1 Like
3

Thanks a lot sir I understood now

4

Please don’t say, Sir, I am also a student just like you, and I am glad, I could help :blush:

5

I have some more doubts can I ask @Elemento

6

Sure @enuguru_2002, I would love to help you. But please note that I am also a student, so don’t take my answers as a reference. They are just a reflection of my understanding, which may be wrong at times. Nonetheless, I will try my best to answer all your questions, to the best of my knowledge.

7

Actually I am facing difficulty in understanding exact math behind gans I started reading paper “A MATHEMATICAL INTRODUCTION TO GENERATIVE
ADVERSARIAL NETS (GAN)” to understand the math but I am unable to understand the paper fully can you please help me from where should I need to start to understand the math behind gans much better please

8

@enuguru_2002 To be frank, I haven’t read the research paper of GANs myself. I found the lecture notes and notebooks to be quite sufficient. Though I will be wrong to admit that I understood the entire concept of GANs the first time, when I reviewed the lecture videos and notebooks, the second time, things started to fall in place. You may follow this approach if you would like to, otherwise, I am sure that you can find a great number of amazing tutorials on GANs on the web, if you are unable to comprehend the research paper, and not finding the provided resources sufficient!