Deep convolutional Generative Adversarial Network

Neural network generating hand written numbers

Custom model trained with custom data loader

Model

• Fully convolutional, no Linear (Fully contected) layers
• Generator with 6x TransposeConv, batchnorm, ReLu activation
• Discriminator with 6x Conv2d, batchnorm, Leaky ReLu activation with 0.2 negative_slope
• Orthogonal weight init with 0.1 gain
• Zero bias init
• Trained on NVIDIA RTX 3060

Loss function

Instead of the usual BCE function, I used BCEWithLogitsLoss for better numerical stability.
Loss formula -

The course of the loss function

From this figure we can see that 50 epoch is enough for these models. Generator can not get any better.
For better results you should try wasserstein loss for example. But i think for this aplication this is enough

Result

Fully generated hand digits from Generator neural network

How to use

Download dataset here - Dataset
In run_training.py change path to dataset.
Run run_training.py.