MNIST_GAN

Reproduct of GAN for MNIST. Just for learning the basic knowledges about Generative Adversarial Networks (GAN).

Results

Epoch 2500 17900 30200 40000

Basic Thoughts

GAN has a generator generates fake instances from noises, and a discriminator distinguishes the fake instances and real instances.

Training both two networks reaches a Nash Equilibrium the GAN model is converged.

Structure

The generator is a U-Net-like backbone with two downsampling stages and two upsampling stages. The discriminator is a VGG-like backbone with an average pooling after two convolutions and a global average pooling before fully connected.

Usage

Simple:

python mnist_gan.py

Costume:

python mnist_gan.py \
--learning_rate 0.01 \
--batch_szie 256 \
--epochs 40000 \
--smooth 0.9 \
--dataset_path /path/to/your/MNIST/dir/mnist.npz \
--model_dir /path/to/your/saved/model/dir \
--log_dir /path/to/your/saved/log/dir

Tips

This GAN project is build up following these advices.