Understanding CNNs and Generative Adversarial Networks
Implemented GAN and understand the property and features of CNN model. What does CNN actually learn?
Train Time
Using one node on Bluewater(UIUC Supercomputer) cost: ~45 hours for GAN_discriminator_With_Generator.py, acc on cifar10, 83.44% in epoch 200, max acc: 85.027% in epoch 199 ~2 hours for Discriminator_Without_Generator.py, acc on cifar10, 88.0600%'s It's normal to have a GAN discriminator result in lower acc than discriminator without generator
cifar10 real images
Here is example of cifar10 dataset's real images
GAN Result
1st epochs example
50th epochs example
100 epochs example
150 epochs example
200 epochs example
Perturb Real Image
A batch of real images, a batch of the gradients from an alternate class for these images, and the modified images the discriminator incorrectly classifies.
Real Image
Here is example of cifar10 dataset's real images
Gradient image
modified image, jittered
Synthetic Images Maximizing Classification Output
Synthetic image without generator
Synthetic image with generator
Synthetic Features Maximizing Features at Various Layers
From this part. Synthetic images maximizing a particular layer of features. Do this for at least two different layers. Here I did layer 2 and layer 7 for both discriminator
For discriminator without Generator
Layer2
Layer7
For discriminator with Generator
Layer2
Layer7
