owenanalytics / scrabble-gan

Adversarial Generation of Handwritten Text Images

ScrabbleGAN (TensorFlow 2/ Python3)

Implementation of ScrabbleGAN for adversarial generation of handwritten text images in TensorFlow 2.1.

More details can be found here

Implementational notes

• the BigGAN-part/ structure of ScrabbleGAN is based on google/compare_gan rewritten in TF2
• all dependencies (batch size, epochs etc.) are handled using google gin-config

Setup (only tested on MacOS)

1. Download and extract IAM-Handwriting DB and random word list as described in src/dinterface/dinterface.py

   wget --user <user-id> --password <pw> http://www.fki.inf.unibe.ch/DBs/iamDB/data/words/words.tgz 
   wget --user <user-id> --password <pw> http://www.fki.inf.unibe.ch/DBs/iamDB/data/ascii/words.txt
   wget -O random_words.txt https://www.dropbox.com/s/o60ubzh3m163j0x/random_words.txt?dl=0
   tar zxvf words.tgz
   rm -rf words.tgz       
   

2. Move folders into correct position (see File Structure); update file paths in scrabble_gan.gin + main.py

3. Install dependencies (e.g. AWS P3)

Usage

Train model

python3 main.py

File Structure

 doc
     ├── installation_aws_p3.txt                    # how to install dependencies on AWS P3 instance
     ├── ...                                        # addtional README resources
 res                               
     ├── data/                                      # various word datasets (e.g. RIMES, IAM, ...)
         ├── iamDB/                                              
             ├── words/                                          
             ├── words-Reading/ 
             ├── words.txt                                            
         ├── random_words.txt/                                  
     ├── out/                                       # location where to store generated output files   
         ├── big_ac_gan/                            # output of ScrabbleGAN (here denoted as big_ac_gan)
             ├── images/                            
             ├── model/                             
             ├── training_checkpoints/              
         ├── ...                                    # potentially output of more models   
 src
     ├── bigacgan/                                  # ScrabbleGAN implementation
         ├── ...                                           
     ├── dinterface/                                # purpose: bring various datasets into equal format
         ├── ...                                         
     ├── main.py                                    # core of this repo - train ScrabbleGAN
     ├── scrabble_gan.gin                           # all dependencies are handled here (gin config file)

License

License