 |
|---|
| Sample of images produced by the transposed convolutional GAN trained on the CIFAR-10 dataset |
In this repository you will find various materials associated with the semester long research project for CAP6610. In this project I wanted to examine if the conditional variations of the GAN and VAE were superior to their unconditional counterparts. I also wanted to see if the upscaling convolution was actually better than the transposed convolution.
To validate this I trained 8 models on the CIFAR-10 dataset: 4 Unconditioned GANs and VAEs with some consisting of uspcaling convolutions and others of transposed convolutions and 4 of the same GANs and VAEs but conditioned. To validate the results of this I utilized FID and IS.
It was found that the conditional variations of the VAE were significantly better than their non-conditioned counterparts but for GANs, it seemed conditioning made it worse. More research will need to be done to determine the cause of this but this offers a unique insight into the power of conditioning generative models.
This repository is divided into four major folders:
- Images
- models
- Notebooks
- Reports
You can find some sample output of each model in images, the trained weights of each model under models, and the notebooks used to train and evaluate the models under Notebooks. Reports as the name states contains all the reports I wrote and developed this semester for the project.
The following resources were used as a starting point and were heavily modified for my uses and experiments.
https://keras.io/examples/generative/conditional_gan/ https://keras.io/examples/generative/vae/ https://keras.io/examples/generative/dcgan_overriding_train_step/ https://www.tensorflow.org/tutorials/generative/cvae