ANIME-FACE-GAN

Implement

1. SA-GAN

https://arxiv.org/pdf/1805.08318.pdf

Implemented with Attention, Conv2DTranspose, hinge loss and spectral norm.

The SAGAN was trained in batchsize=64 and cost only 3GB GPU memory. It needs about 50000 steps for training.

2. BIG-GAN

https://arxiv.org/abs/1809.11096?context=cs.LG

The BIGGAN was trained in batchsize=64 and cost 16GB GPU memory (batchsize=32 cost 10GB GPU memory for 1080Ti). It needs only 10000 steps for training.

results

state-of-art BIGGAN 12600 steps

GIF

SAGAN 61600 steps (without the residual structure)

Open sourced dataset

We decide to opensource our used datasets.

The datasets are handlely cleaned and labeled, enjoy your own Playground with Gan! LOL

url:https://pan.baidu.com/s/1xXPeqr6SDnQkaNZcVHCZ7Q

extracting-code:u3bi

Loss

Why the generator loss is crippled in 15k steps?

Model Records

SAGAN_V2: SAGAN + deconv

SAGAN_V3: SAGAN + deconv + bs=64 + truncated_normal

SAGAN_V4: SAGAN + upsample + bs=128 + truncated_normal

SAGAN_V5: SAGAN + deconv + bs=64 + lr_decay after 50k steps + ema for genrator

SAGAN_V6: SAGAN + deconv + bs=64 + ema for genrator

SResNetGAN_V0: SResNetGAN + pixelshuffler (failed)

SResNetGAN_V1: SResNetGAN + deconv (failed)

BIGGAN_V0: BIGGAN + generator 512

BIGGAN_V1: BIGGAN + generator 1024 (best now!)

Experience

• Use truncated norm (std=0.5, truncated from -1 to 1) instead of uniform and Gaussian normal can help convergence.
• Binomial distribution works badly.
• Use Conv2DTranspose instead of Upsampling can improve the quality of images, and Upsampling also loses some diversities.
• Bigger batch size (128, 256,...) dosen't achieve better performance in this project (not sure).
• Ensure enough steps to train (at least 50k in SAGAN).
• Add ExponentialMovingAverage to the generator can improve the stability of generated images.
• It is important to remain close parameter size for both discriminator and generator.
• The residual structure and increasing the parameters scale of both discriminator and generator can improve the image details for generated results.

Questions

• Pixelshuffle works bad (pool diversity).
• The hinge loss of discriminator usually equals 0 during the second half of training.
• The quality of the generated images rapidly falling after several steps (70k in SAGAN, 14K in BIGGAN).

TODO

• [x] Add ExponentialMovingAverage to the generator
• [x] Learning rate exponentially decease after 50000 iterations of training (failed, not sure).
• [x] Add labels from illustration2vec.