mnist-gan

Naive Generative Adversarial Network (GAN) implementation able to generate images of digits from the MNIST digit dataset using fully connected layers. Implemented using the Rust dfdx crate.

Training Data

Input data is expected in the following file structure:

├──mnist-gan
│  ├──data
│  │  ├──training
│  │  │  ├──0
│  │  │  │  ├──digit_0_0.jpg
│  │  │  │  ├──digit_0_1.jpg
│  │  │  ├──1
│  │  │  │  ├──digit_1_0.jpg
│  │  │  │  ├──digit_1_1.jpg
│  │  │  ├──2
│  │  │  │  ├──digit_2_0.jpg
│  │  │  │  ├──digit_2_1.jpg
...

File names don't matter. All files within the folders will be considered for training. This source offers the data in such structure.

Results

After 500,000 iterations the network is able to produce results like this: