callifrey / animefacegeneration

This is a anime faces generation demo based on DCGAN(Deep Convolutional GAN) by Pyrotch.

AnimeFaceGeneration

Description

This is a anime faces generation demo based on DCGAN(Deep Convolutional GAN), the framework including:

• Generator: Consist of one linear layer and 4 transpose Conv layers, output [3,64,64]
• Discriminator: 5 layers strides convolution to output score for input

Notice:

• Generator and Discriminator use Batch Normalization and ReLU activation function in all Linear / Conv/ Transpose Conv blocks.
• Use Tanh for last layer of Generator, instead of Sigmoid for Discriminator

Dataset

The datasets of this project is offer by Kaggle, you can download from here

There are 21.6k training images totally, which resolution (64,64)

Training

You can train the model by run:

python train.py

Maybe you need to change some hyper-parameters and data path by modifying config.py

Testing

You can also testing the model by offered pre-trained weight in ./data/out/checkpoints, running:

python test.py

Result

Here are some generate anime faces under different epochs in training:

Epoch 0	Epoch 10
Epoch 50	Epoch 100
Epoch 200	Epoch 400

Next, the test result for final model(epoch = 400)

Training methodology

• Using small learning rate for both G and D
• Using label smooth to introduce random
• Control the update step rate when you see training procedure become strange

Acknowlegement

Thanks to Google Colab offering free GPU for accelerating computing.