ScrabbleGAN - Handwritten Text Generation

A PyTorch implementation of the ScrabbleGAN: Semi-Supervised Varying Length Handwritten Text Generation paper.

The main code was taken from amzn/convolutional-handwriting-gan and arshjot/ScrabbleGAN.

Quick setup and start

• Nvidia drivers >= 470, CUDA >= 11.4
• Docker, nvidia-docker

The provided Dockerfile is supplied to build an image with CUDA support and cuDNN.

Preparations

• Clone the repo.
• Download and extract dataset to the
  data/
  folder.
• sudo make all
  to build a docker image and create a container. Or
  sudo make all GPUS=device=0 CPUS=10
  if you want to specify gpu devices and limit CPU-resources.

If you don't want to use Docker, you can install dependencies via requirements.txt

Preprocess data

The dataset should contain a folder with text images and csv file with annotations. The csv file should consist of two columns: "filename" with the relative path to the images (folder-name/image-name.png), and "text"-column with the image transcription.

To preprocess data and to create pickle:

You can pass data_csv_path argument multuple times if you have a lot of csv files.

Train

• unlabeled_pkl_path - path to unlabeled data, if you need to train model with images without transcriptions (these images will be used to train discriminator only).
• lexicon_path - the path to the txt-file with text in each line.

Generate images

• char_map_path - path to labeled pickle data, that contains the alphabet of the data (char_map).
• lexicon_path - the path to the txt-file with text in each line. It may differ from the lexicon used for training the model.