dsanno / chainer-stack-gan

Stack GAN

Implementation of StackGAN except text to image synthesis.

Usage

Convert images to pickle file

$ python src/convert_dataset.py image_dir out_path [-n max_image_num]

Parameters:

• image_dir: (Required) Image file directory path
• out_path: (Required) Output pickle file path
• -n max_image_num: (Optional) The maximum number of images to be stored (default: 1000000)

Example:

$ python src/convert_dataset.py ../images dataset/images.pkl

Train Stage-I model

Example:

$ python src/train.py -d dataset/images.pkl -o model/stage1 -g 0 --out-image-dir image/stage1 --clip-rect 25,55,128,128

Train Stage-II model

Example:

$ python src/train.py model/stage1_050 -d dataset/images.pkl -o model/stage2 -g 0 --out-image-dir image/stage2 --clip-rect 25,55,128,128

Generate image

Example:

$ python src/generate2.py model/stage1_050.gen.model model\stage2_050.gen.model image\generated -g 0

Differences from original paper

• Text to image synthesis is not implemented.
  Input of generator is 100-dimension latent vector.
• Use Energy-based GAN instead of normal DCGAN.
• Output image size of Stage-II GAN is only 128 x 128.

License

MIT License