nogu-atsu / small-dataset-image-generation

Image Generation from Small Datasets via Batch Statistics Adaptation

The author's official minimal implementation of Image Generation from Small Datasets via Batch Statistics Adaptation.

Requirements

chainer>=5.0.0
opencv_python
numpy
scipy
Pillow
PyYAML

Dataset preparation

data_path
├── dataset1
├── dataset2
...

• Place all training sample in the same directory for each dataset.
• Specify the root path to data_path in configs/default.yml.
• Specify the dataset directory name to dataset in configs/default.yml.

Run

• Download SNGAN-128 or BigGAN-256 generator pre-trained on ImageNet.
  • SNGAN generator can be downloaded from https://drive.google.com/drive/folders/1m04Db3HbN10Tz5XHqiPpIg8kw23fkbSi (ResNetGenerator_850000.npz)
  • BigGAN generator can be saved by https://github.com/nogu-atsu/chainer-BIGGAN
• Specify paths for them to snapshot in configs/default.yml
• Specify gan_type in configs/default.yml

For single GPU training, run

python ./train.py --config_path configs/default.yml

For Multiple GPU training, run

mpirun python ./train.py --config_path configs/default.yml

Multiple GPU training is supported only for BigGAN. For BigGAN ,we used 4 GPUs for training.

Inference

Initialize the generator and load pretrained weights.

e.g. biggan based model

gen = AdaBIGGAN(config, datasize, comm=comm)
chainer.serializers.load_npz("your_pretrained_model.h5", gen.gen)
gen.to_gpu(device) # send to gpu if necessary
gen.gen.to_gpu(device) # send to gpu if necessary

Random sampling

e.g. randomly sample 5 images with temperature=0.5 without truncation

random_imgs = gen.random(tmp=0.5, n=5, truncate=False)

Interpolation

e.g. interpolation between 0th and 1st image

interpolated_imgs = gen.interpolate(self, source=0, dest=1, num=5)

Acknowledgement

Pytorch re-implementation from Satoshi Tsutsui and Minjun Li.