In this repository, I have prepared examples of classical generative models - simple convolutional autoencoder (AE) and vanilla convolutional generative adversarial model (GAN). The objectives are two-fold:
- Use autoencoder to reconstruct input maps as close as possible.
- Use GAN to generate channelized examples given random input noise.
- In both cases, I used (unpacked) channelized (binary) maps.
- Furthermore, the models are too constructed using PyTorch Lighning which nicely and neatly wraps PyTorch.
- The AE model can be accessed here, which comes with saved weights that can be loaded using pre-trained weights.
- The GAN model can be accessed here, which comes with saved weights that can be loaded using pre-trained weights.
- Examples of reconstructed maps can be found below. First, I present test set that is followed by predicted maps using AE model. Then, I present a set of examples generated using GAN.
Test set - True Maps
Test set - Reconstructed Maps using AE
Generated Maps using GAN
