ManiGAN

Pytorch implementation for ManiGAN: Text-Guided Image Manipulation. The goal is to semantically edit parts of an image according to the given text while preserving text-irrelevant contents.

Overview

ManiGAN: Text-Guided Image Manipulation.
Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, Philip H. S. Torr.
University of Oxford
CVPR 2020

Data

Download the preprocessed metadata for bird and coco, and save both into data/
Download bird dataset and extract the images to data/birds/
Download coco dataset and extract the images to data/coco/

Training

All code was developed and tested on CentOS 7 with Python 3.7 (Anaconda) and PyTorch 1.1.

DAMSM model includes text encoder and image encoder

Pre-train DAMSM model for bird dataset:

python pretrain_DAMSM.py --cfg cfg/DAMSM/bird.yml --gpu 0

Pre-train DAMSM model for coco dataset:

python pretrain_DAMSM.py --cfg cfg/DAMSM/coco.yml --gpu 1

ManiGAN, main module

Train main module for bird dataset:

python main.py --cfg cfg/train_bird.yml --gpu 2

Train main module for coco dataset:

python main.py --cfg cfg/train_coco.yml --gpu 3

*.yml files include configuration for training and testing.

ManiGAN, detail correction module (DCM)

Save trained main module to models/

Train DCM for bird dataset:

python DCM.py --cfg cfg/train_bird.yml --gpu 2

Train DCM for coco dataset:

python DCM.py --cfg cfg/train_coco.yml --gpu 3

Pretrained DAMSM Model

DAMSM for bird. Download and save it to DAMSMencoders/
DAMSM for coco. Download and save it to DAMSMencoders/

Pretrained ManiGAN Model

Main module for bird. Download and save it to models/
Main module for coco. Download and save it to models/
DCM for bird. Download and save it to models/
To verify the IS value on the bird dataset, please use this link. Download and save it to models/
DCM for coco. Download and save it to models/

Testing

Test ManiGAN model for bird dataset:

python main.py --cfg cfg/eval_bird.yml --gpu 4

Test ManiGAN model for coco dataset:

python main.py --cfg cfg/eval_coco.yml --gpu 5

Evaluation

To generate images for all captions in the testing dataset, change B_VALIDATION to True in the eval_*.yml.
Inception Score for bird dataset: StackGAN-inception-model.
Inception Score for coco dataset: improved-gan/inception_score.

Code Structure

code/main.py: the entry point for training the main module and testing ManiGAN.
code/DCM.py: the entry point for training the DCM.
code/trainer.py: creates the main module networks, harnesses and reports the progress of training.
code/trainerDCM.py: creates the DCM networks, harnesses and reports the progress of training.
code/model.py: defines the architecture of ManiGAN.
code/attention.py: defines the spatial and channel-wise attentions.
code/VGGFeatureLoss.py: defines the architecture of the VGG-16.
code/datasets.py: defines the class for loading images and captions.
code/pretrain_DAMSM.py: trains the text and image encoders, harnesses and reports the progress of training.
code/miscc/losses.py: defines and computes the losses for the main module.
code/miscc/lossesDCM.py: defines and computes the losses for DCM.
code/miscc/config.py: creates the option list.
code/miscc/utils.py: additional functions.

Citation

If you find this useful for your research, please use the following.

@inproceedings{li2020manigan,
  title={Manigan: Text-guided image manipulation},
  author={Li, Bowen and Qi, Xiaojuan and Lukasiewicz, Thomas and Torr, Philip HS},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={7880--7889},
  year={2020}
}

Acknowledgements

This code borrows heavily from ControlGAN repository. Many thanks.

mrlibw/ManiGAN