Pytorch-InstaGAN

Unofficial re-implementation of InstaGAN (https://openreview.net/pdf?id=ryxwJhC9YX) with PyTorch

Libraries

Assumed

PyTorch==1.3.1

Required

pycocotools (https://github.com/cocodataset/cocoapi/tree/master/PythonAPI/pycocotools)

Dataset preparation

Learning InstaGAN requires annotation information on a per-instance basis. In this repository, The format of the annotations should follow the COCO dataset.

Download COCO dataset

To get COCO dataset, run get_dataset.py
CAUTION: By this script, you will download a large amount of image data (> 18GB).

$ python get_dataset.py

After execution, a checkpoint directory with the following structure will be created.

├── data
│   ├── instances_train2017.json
│   └── train2017
│          ├── 000000000009.jpg
│          ├── 000000000025.jpg
│          ├── ...

Train

run train.py

$ python train.py params.yaml

After execution, a checkpoint directory with the following structure will be created to store the learning results.

├── result
│   └── yymmdd_HHMM_[domain X]2[domain Y]
│          ├── params.json
│          ├── weights
│          └── samples

Train from checkpoint

$ python resume_train.py [path to checkpoint directory]