/caption2image

PyTorch implementation of GAN-INT-CLS and AttnGAN

Primary LanguagePython

caption2image

PyTorch implementation of GAN-INT-CLS and AttnGAN

Dependencies

  • python 3
  • Pytorch 1.0.0
  • tensorflow, tensorboard (you can train/evaluate your model without this if you do not use tensorboard for logger)

In addition, you may need other packages...

Data

  1. Download preprocessed metadata for COCO filename and COCO text and extract
  2. Download COCO dataset
  3. Download embedding file
  4. Place data as below
data_dir 
  |- COCO
       |- filenames 
            |- train2014 
            |- val2014 
       |- text 
            |- train2014 
            |- val2014 
       |- image 
            |- train2014 
            |- val2014

AttnGAN

Training

  • Train DAMSM models:

    • python DAMSM_main.py
      • you can edit config by directly editting the source code

  • Train AttnGAN models:

    • python main.py
      • you can edit config by passing arguments (see AttnGAN/config.py or python main.py --help)

Evaluation

I prepared notebook for evaluation (AttnGAN/eval.ipynb).
You can evaluate generated images by

  • inception score
  • frechet inception distance
  • R-precision

You can also generate images from your own captions.

Pretrained Model

  1. Download DAMSM image_encoder
  2. Download DAMSM text_encoder
  3. Download AttnGAN Generator and config
  4. Place models as below
AttnGAN
  |- results
       |- DAMSM/COCO/2019_05_04_00_32/model
            |- image_encoder600.pth
            |- text_encoder600.pth
       |- AttnGAN/COCO/2019_05_14_17_08
            |- model
                 |- G_epoch50.pth
            |- config.txt

TODO

  • Paper
    • GAN-INT-CLS
      • survey
      • impl
    • StackGAN
      • survey
      • impl
    • StackGAN++
      • survey
      • impl
    • AttnGAN
      • survey
      • impl
    • MirrorGAN
      • survey
      • impl
  • Dataset
    • Bird
    • MS COCO

Reference