Skip to content

Latest commit

 

History

History
98 lines (75 loc) · 4.8 KB

README.md

File metadata and controls

98 lines (75 loc) · 4.8 KB
Raw

HWGAN: Handwritten Text Generator

Made with Python

This project was executed as a school assignment at the University of Twente. HWGAN is our own implementation of a handwriting generator build with a GAN and OCR neural network. The neural networks in this project have been build with tensorflow.

Project Overview

  • School: University of Twente
  • Course: Machine Learning II
  • Assignment Type: Open Project
  • Group Size: 4

Setup

  1. Use Python 3.6-3.8
  2. Execute the following command to install required packages:
pip install -r ./helper/requirements.txt
  1. For GPU support we recommend to also install CUDA Toolkit 11.0, cuDNN 8.0.4 and NVIDIA GPU Driver 450 or higher (NVIDIA website)(TensorFlow guide)

Usage

  • The training of the OCR and GAN models will by default use the EMNIST ByMerge dataset, follow our DATA_GUIDE for the setup of custom datasets. For training the neural networks use the following command:
python train.py -data emnist
  • For creating a word use the following command (where example_word is the word you want to create):
python run.py -data emnist -text example_word

Available characters: 0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz_
For more specific options look below at arguments

Arguments

Training (train.py)

  • -data: Specify which dataset to use (default = emnist)
  • -sample: Gives an example of how data is splitted. A number is given to indicate how many example images you want to retrieve (default = 0)
  • -text: Specify how you want the data to be splitted, options: chars, words, lines (default = chars)
    • chars is single character images
    • words is images of multiple words but only on a single line
    • lines is multiple words and lines in a single image
    • Currently it is not possible to have a combination of these options, so your dataset needs to adhere to one of these options
  • -ocr: Specify if you want to train the OCR (recognizer) model, options: True, False (default = False)

Create Handwriting (run.py)

  • -data: Specify which dataset to use (default = emnist)
  • -text: Specify the word you want to create (e.g. = example_word)

File Structure

 dataset
     ├── DATA_GUIDE.md                      # How to add custom datasets
     ├── ...                                # Location to add custom datasets
 helper
     ├── coversion.py                       # Custom images are split into characters
     ├── requirements.txt                   # Configuration file with all dependencies to install
     ├── split_data.py                      # Split the dataset into letter specific data
     ├── userinput.py                       # Handle user arguments
 models
     ├── gan_model/                         # Holds all files related to the GAN model
         ├── gifs/
             ├── ...                        # .gif files of the training process
         ├── graphs/
             ├── ...                        # Loss graphs of the training process
         ├── saved_models/
             ├── ...                        # Trained discriminator and generator models
         ├── GAN.py                         # GAN model
     ├── ocr_model/                         # Holds all files related to the OCR model
         ├── ocr_model.h5                   # Trained OCR model
         ├── OCR.py                         # OCR model
         ├── ...                            # Stats on the performance
 out
     ├── ...                                # Output of the run.py executable
 run.py                                     # Main executable - Generate given word
 train.py                                   # Main executable - Train all models
 ...                                        # Extra project management files

Acknowledgments

The neural network setup has already been executed by ScrabbleGAN, which is a more elaborate implementation of this principal with pytoch.