cGAN experiments

Implementation of:

• Pix2pix paper: https://arxiv.org/abs/1611.07004
• cGAN (For MNIST)
• GAN (For MNIST)

Pix2Pix CGAN (Unet + PatchGAN: GAN Loss + L1)

Facades 100 epoches (400 images)

Maps 40 epoches (1.3k images)

CityScapes 20 epoches (3k images)

Parameters:

D_architecture: PatchGAN 70x70 D_lr = G_lr: 0.0002 D_betas = G_betas: (0.5, 0.999) Optimizers: Adam + Adam EPOCHS: 100 G_architecture: Unet init_mean: 0 init_std: 0.02 lambda: 100 Batch size: 1 Augmentations: Resize to (286, 286) then RandomCrop, then Random Flip

PatchGAN 70x70 is applied to the image of size of (256x256) it downsample to (30x30), each pixel of the new picture represents predicted probability of overlaping patch of size (70x70) to be from real distribution.

CGAN (MLP + MLP) on MNIST

As you can observe from the picture, two MLP is enough to get ahold of such primitive dataset.

Parameters:

1. 50 Epoches
2. Discriminator and Generator are both of MLP architecture
3. Optimizers: Adam + Adam, lr = 3e-4, betas=(0.9, 0.999)

Embeddings: Learnable torch.nn.Embedding

Pix2Pix CGAN (Unet: only L1 Loss)

References:

1. https://arxiv.org/abs/1611.07004
2. https://arxiv.org/abs/1411.1784
3. https://machinelearningmastery.com/how-to-develop-a-conditional-generative-adversarial-network-from-scratch/
4. https://d2l.ai/chapter_generative-adversarial-networks/index.html
5. https://www.youtube.com/watch?v=banZhpreS2Y&list=PLEwK9wdS5g0onnKgvKxuUJN1Ojchl9Q9P&index=22
6. https://github.com/soumith/ganhacks
7. https://developers.google.com/machine-learning/gan/problems
8. https://www.kaggle.com/code/kmldas/mnist-generative-adverserial-networks-in-pytorch
9. http://efrosgans.eecs.berkeley.edu/pix2pix/datasets/