Skip to content

Latest commit

 

History

History
18 lines (12 loc) · 1.21 KB

README.md

File metadata and controls

18 lines (12 loc) · 1.21 KB

Pytorch_GAN_MNIST

Pytorch implementation of Basic GAN in generating handwritten digits

Background

Generative Adversarial Net (GAN) is a deep learning architecture framework that consists of 2 main components - Generator and Discriminator. It is an extremely powerful architecture in many content-generation tasks (e.g. image generation, Low-light image enhancement).

GAN was introduced in a paper (https://arxiv.org/pdf/1406.2661.pdf) by Ian Goodfellow and other researchers including Yoshua Bengio in 2014. Facebook's AI research director Yann LeCun once said adversarial training being "the most interesting idea in the last 10 years in ML"

Training

In this project, I have carried out 300 epochs of training. In the process, I have encountered different problems (e.g. too high learning rate, overfitting of the Discriminator). At last, the training went well. The losses are shown below:

GAN_MNIST_losses

Results

Below is the evolution of the result in 300 epochs (sampled every 10 epochs)

GAN_MNIST_300_epochs