Skip to content

Latest commit

 

History

History
89 lines (65 loc) · 2.69 KB

aotgan.md

File metadata and controls

89 lines (65 loc) · 2.69 KB
Raw

AOT GAN

1 Principle

The Aggregated COntextual-Transformation GAN (AOT-GAN) is for high-resolution image inpainting.The AOT blocks aggregate contextual transformations from various receptive fields, allowing to capture both informative distant image contexts and rich patterns of interest for context reasoning.

Paper: Aggregated Contextual Transformations for High-Resolution Image Inpainting

Official Repo: https://github.com/megvii-research/NAFNet

2 How to use

2.1 Prediction

Download pretrained generator weights from: (https://paddlegan.bj.bcebos.com/models/AotGan_g.pdparams)

python applications/tools/aotgan.py \
	--input_image_path data/aotgan/armani1.jpg \
	--input_mask_path data/aotgan/armani1.png \
	--weight_path test/aotgan/g.pdparams \
	--output_path output_dir/armani_pred.jpg \
	--config-file configs/aotgan.yaml

Parameters:

  • input_image_path：input image
  • input_mask_path：input mask
  • weight_path：pretrained generator weights
  • output_path：predicted image
  • config-file：yaml file,same with the training process

AI Studio Project:(https://aistudio.baidu.com/aistudio/datasetdetail/165081)

2.2 Train

Data Preparation:

The pretained model uses 'Place365Standard' and 'NVIDIA Irregular Mask' as its training datasets. You can download then from (Place365Standard) and (NVIDIA Irregular Mask Dataset).

└─data
    └─aotgan
        ├─train_img
        ├─train_mask
        ├─val_img
        └─val_mask

Train(Single Card):

python -u tools/main.py --config-file configs/aotgan.yaml

Train(Mult-Card):

!python -m paddle.distributed.launch \
    tools/main.py \
    --config-file configs/photopen.yaml \
    -o dataset.train.batch_size=6

Train(continue):

python -u tools/main.py \
	--config-file configs/aotgan.yaml \
	--resume  output_dir/[path_to_checkpoint]/iter_[iternumber]_checkpoint.pdparams

Results

On Places365-Val Dataset

mask PSNR SSIM download
20-30% 26.04001 0.89011 download

References

@inproceedings{yan2021agg, author = {Zeng, Yanhong and Fu, Jianlong and Chao, Hongyang and Guo, Baining}, title = {Aggregated Contextual Transformations for High-Resolution Image Inpainting}, booktitle = {Arxiv}, pages={-}, year = {2020} }