Shulou(Shulou.com)06/02 Report--

This article is about how to use PyTorch to implement WGAN in Python. The editor thinks it is very practical, so share it with you as a reference and follow the editor to have a look.

Brief introduction of 1.GAN

In GAN, there are two models, one is the generation model, which is used to generate samples, and the other is the discrimination model, which is used to judge whether the samples are true or false. However, because in GAN, the JS divergence is used to calculate the loss value, it is easy to lead to gradient dispersion, so it is impossible to update the parameters by gradient descent, so Wasserstein Distance is introduced into WGAN to make the training stable. In this paper, we take the data which obeys the Gaussian distribution as the sample.

two。 Generator module

Here from the two-dimensional data, the final generation of two-dimensional, the main purpose is to facilitate visualization. In other words, in the generation model, we enter disorganized two-dimensional data, and after training, we can generate a fake that mimics the Gaussian distribution.

3. Discriminator module

The discriminator also inputs two-dimensional data. For example, our generator above generates a two-dimensional fake, and after inputting it into the discriminator, it can finally output a result after sigmoid conversion, which is equivalent to a probability, so as to judge whether the fake can reach the degree of being fake or not.

4. Data generation module

Because we use the Gaussian model, we can generate the data we need directly. In this module, we generate 8 data that obey Gaussian distribution.

5. Discriminator training

Because when using JS divergence to calculate the loss, it is easy to occur that the gradient is very small and close to 0, which will make the gradient drop impossible, so when calculating the loss, Wasserstein Distance is used to measure the difference between the two distributions. So let's say we take the factor of gradient punishment.

The modules of gradient punishment are as follows:

6. Generator training

The training here is followed by discriminator training. In other words, in a cycle, train the discriminator first, and then train the generator.

7. Result visualization

The loss value is visualized by visdom and the prediction result of distribution is visualized by matplotlib.

Thank you for reading! This is the end of this article on "how to use PyTorch to achieve WGAN in Python". I hope the above content can be of some help to you, so that you can learn more knowledge. if you think the article is good, you can share it out for more people to see!

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.