This project is inspired by the Huggingface. While working on projects unrelated to NLP, I really miss great documentation and well thought API that Hugginface contirbuted to the NLP echosystem. I therefore decided to start the project focused on image processing. The overall idea is to collect (if exist) or to create a a bunch of image models with easy to use APIs and decent documentation as installable Python packages. All code is in PyTorch.

The first model available is a very basic GAN based on fully connected network architecture. This is not something anyone would ever use in production yet it is useful if for building an overall understanding of generative adversarial network. The model demonstrates decent performance on MNIST dataset (28x28 pixels images). Using GPU is highly recommended, minimum 100 iterations are suggested

Installation from Github(currently from develop branch) -
pip install git+https://github.com/pastrop/GANS.git@main#egg=GANS
Model Package - GANS
Model Class - GAN_MLP(z_dim = 64, lr = 0.00001,device = 'cpu')
z_dim - noize vector dimensionality
lr - learning rate
device - 'gpu' or 'tpu' is recommended
Example: Initialzing

from GANS import model
example_model = model.GAN_MLP(device = 'cuda')

Methods & Variables
example_model.gen_opt - holds the generator gradient (Adam is used)
example_model.desc_opt - holds the descriminator gradient (Adam is used)
example_model.get_disc_loss(batch_size, batch_of_real_images) - calculates descriminator loss example_model.get_gen_loss(batch_size) - calculates generator loss

Example: Calling Methods & Using Class Variables

example_model.gen_opt
example_model.desc_opt
example_model.get_disc_loss(batch_size, batch_of_real_images) 
example_model.get_gen_loss(batch_size)

Dependencies
Torch
Torchvision & Pyplot - if you want to have vizualization