Experiments done with autoencoders for my Master's Degree dissertation. The autoencoders are based on the Pix2Pix and CycleGAN generators.
The objective of the experiment is to create autoencoders capable of generating a latent vector that can be used for image manipulation.
The complete experiment can be accessed on (PT-BR only):
https://www.notion.so/vinitrevisan/Supervisionada-d26a162bcfa547adb17bf73716ae97c9
main.py
File with the experiment core. Training, test and validation of the networks. Parameters of the experiment.
networks.py
File with the classes and functions to create the generators and discriminators used on the experiments.
losses.py
File with the functions used to evaluate the losses to train the GANs
metrics.py
File with the functions used to evaluate the quality metrics (FID, IS, L1, Accuracy).
utils.py
File with all utilities functions, such as plot control, image processing, and exception handling.
transferlearning.py
File with the functions used to train networks using the Transfer Learning approach.
validate.py
Tests vector interpolation to see how the reconstruction of interpolated images is working for a given generator.
Autoencoders (github)
Pix2Pix-CycleGAN (github)
Unsupervised-GANs (github)
Main experiment page (Notion [PT-BR])