GAN implementations with PyTorch
The followings are a list of GAN implemented here.
- Vanilla GAN
- Vanilla GAN for MNIST
- Vanilla GAN for Fashion MNIST
- Conditional GAN for MNIST
- DCGAN for Fashion MNIST
- Improved GAN
- Feature Maching
It shows a little bit of mode collapse; a commonly encountered failure case for GANs where the generator produces samples with extremely low variety. In this case, the generator produces
1 with extremely high probability.
Conditional GAN reduces mode collapse issue by giving the model additional information.
DCGAN makes use of convolutions and transposed convolutions.
It uses feature mapping and LSTM.
Anomaly Detection with GAN
images generated by GAN Generator
The following images are generated by the GAN Generator. It is the same model as the DCGAN model.
Anomaly Detection for Normal Data
The following images are normal images.
It is good if red dots are less shown.