Skip to content

CNN, AutoEncoder, Generative Models, Score Matching, StyleGan, Self Supervision.

Notifications You must be signed in to change notification settings

itamar-sh/Neural-Network-for-Images

Repository files navigation

Neural-Network-for-Images

CNN, AutoEncoder, Generative Models, Score Matching, StyleGan, Self Supervision.
For all those projects I used pytorch, tensorflow and matplotlib.

CNNs

Classification using convolutional neural networks.
Working on CIFAR10 dataset, track results using wandb.
Exploring different architectures, understaing importance of non-linearity and cascaded receptive field.

Example of variance-bias tradeoff:

AutoEncoder with Conv layers

Auto-Encoding (AE) and transfer-learning with the AE over the MNIST digits dataset.
We tested aspects of AE such as dimensionality reduction, Interpolation, Decorrelation.
I used BCELoss as criterion and Adam for the optimizer.
The encoder is Conv layers with sigmoid and FC layer in the end, The decoder has one FC and Deconv layers.

Demonstration of reconstruction success in relation to latent space size:

Generative Adversarial Networks (GANs)

In this part I trained a generative adversarial model to produce images of a particular distribution.
generator network can follow the DCGAN architecture. Generator: FC with convs. Discrimator: classification network which decide: real or fake image..

Main points: Loss saturation, Model Inversion and Image restoration including denoise and inpainting.

Example of denoising using Gan and latent optimization:

About

CNN, AutoEncoder, Generative Models, Score Matching, StyleGan, Self Supervision.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published