• Setup
• Usage
  • Inference
  • Training
• Acknowledgments
• Citations

Zero-Shot Text-to-Image Generation VQGAN+CLIP Dockerized

This is a stripped and minimal dependencies repository for running locally or in production VQGAN+CLIP.

For a Google Colab notebook see the original repository.

Clone this repository and cd inside.

git clone https://github.com/kcosta42/VQGAN-CLIP-Docker.git
cd VQGAN-CLIP-Docker

You can download a pretrained VQGAN model and put it in the ./models folder.

For GPU capability, make sure you have CUDA installed on your system (tested with CUDA 11.1+).

• 6 GB of VRAM is required to generate 256x256 images.
• 11 GB of VRAM is required to generate 512x512 images.
• 24 GB of VRAM is required to generate 1024x1024 images. (Untested)

Install the Python requirements

python3 -m pip install -r requirements.txt

To know if you can run this on your GPU, the following command must return True.

python3 -c "import torch; print(torch.cuda.is_available());"

Make sure you have docker and docker-compose v1.28.0+ installed. nvidia-docker is needed if you want to run this on your GPU through Docker.

A Makefile is provided for ease of use.

make build  # Build the docker image

Two configuration files are provided ./configs/local.json and ./configs/docker.json. They are ready to go, but you may want to edit them to meet your need. Check the Configuration section to understand each field.

By default, the resulting generations can be found in the ./outputs folder.

To run locally:

python3 -m scripts.generate -c ./configs/local.json

To run on docker:

make generate

To run locally:

DEVICE=cpu python3 -m scripts.generate -c ./configs/local.json

To run on docker:

make generate-cpu

These are instructions to train a new VQGAN model. You can also finetunes the pretrained models but you may need to tweak the training script.

Two models configuration files are provided ./configs/models/vqgan_custom.json and ./configs/models/vqgan_custom_docker.json. They are ready to go, but you may want to edit them to meet your need. Check the Model Configuration to understand each field.

By default, the models are saved in the ./models/checkpoints folder.

Put your image in a folder inside the data directory (./data by default).

The dataset must be structured as follow:

./data/
├── class_x/
│   ├── xxx.png
│   ├── xxy.jpg
│   └── ...
│       └── xxz.ppm
└── class_y/
    ├── 123.bmp
    ├── nsdf3.tif
    └── ...
    └── asd932_.webp

To run locally:

python3 -m scripts.train -c ./configs/models/vqgan_custom.json

To run on docker:

make train

To run locally:

DEVICE=cpu python3 -m scripts.train -c ./configs/models/vqgan_custom.json

To run on docker:

make train-cpu

• Let the Generator train without the Discriminator for a few epochs (~3-5 epochs for ImageNet), then enable the Discriminator.
  The variable lossconfig.params.disc_start correspond to the number of global step (ie. batch iterations) before enabling the Discriminator.
• Once enabled, the Discriminator loss will stagnate around ~1.0, this is a normal behaviour. The loss will decrease in later epochs. (It can take a very long time).
• If you've enabled the Discriminator too soon, the Generator will take a lot more time to train.
• Basically there is no rules for the number of epochs. If your dataset is large enough, there is no risk of overfitting. So the more you train, the better.

VQGAN+CLIP

Taming Transformers

CLIP

DALLE-PyTorch

@misc{unpublished2021clip,
    title  = {CLIP: Connecting Text and Images},
    author = {Alec Radford, Ilya Sutskever, Jong Wook Kim, Gretchen Krueger, Sandhini Agarwal},
    year   = {2021}
}

@misc{esser2020taming,
      title={Taming Transformers for High-Resolution Image Synthesis},
      author={Patrick Esser and Robin Rombach and Björn Ommer},
      year={2020},
      eprint={2012.09841},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@misc{ramesh2021zeroshot,
    title   = {Zero-Shot Text-to-Image Generation},
    author  = {Aditya Ramesh and Mikhail Pavlov and Gabriel Goh and Scott Gray and Chelsea Voss and Alec Radford and Mark Chen and Ilya Sutskever},
    year    = {2021},
    eprint  = {2102.12092},
    archivePrefix = {arXiv},
    primaryClass = {cs.CV}
}