Fontastic

Font classifier

This repository contains experiments for classfying fonts from images. Currently it consists of scrapers to pull the data.

Installation Instructions

• Requirments

  • python3
  • pip3
  • virtualenv

• How to install

  • virtualenv -p /usr/bin/python3 venv
  • source venv/bin/activate
  • pip install -r requirements.txt
  • export PYTHONPATH=.

Working with the Fontastic package

The fontastic package contains the following folder structure:

fontastic -- 
            scrape
            utils

How to generate font images?

1. Download fonts and untar it in the location <project_root>. So every font will be list as below

   project_root/data/src/fonts/roboto
   project_root/data/src/fonts/roboto/roboto-regular.ttf
   project_root/data/src/fonts/roboto/roboto-bold.ttf
   project_root/data/src/fonts/roboto/roboto-italic.ttf
   
   project_root/data/src/fonts/open-sans
   project_root/data/src/fonts/open-sans/open-sans-regular.ttf
   project_root/data/src/fonts/open-sans/open-sans-regular.ttf
   project_root/data/src/fonts/open-sans/open-sans-regular.ttf
   ...
   ...
   ...
   

2. now run python fontastic/data_generation/local_image_gen.py

3. the program will create 4k images for every random text and ttf combination. The output image location is project_root/data/dst/roboto/<output_images>

4. Next run python fontastic/data_generation/training_data_gen.py. This program will create crops of the image and perform data transforms that can be used for training the network.

Training and Evaluation

The thought process behind running training and evaluation is to run each off it as an experiment with certain artifacts associated to the experiment. This gives us the following benefits:

• Restart experiments at various stagesired rated during the experiment run
• Version control of experiments

Splitting dataset into training and test

Please look at the TRAIN_TEST_SPLIT section in the config.ini to understand the configurations required to start an experiment and generate test and train data.

The command to use is :

python fontastic/data_generation/train_test_split.py --config fontastic/data_generation/config.ini

This will generate an experiments folder with the experiment_id as sub folder to store the artifacts. In case the experiment_id is left empty in the config during the first run, this will be populated and written back to the config.

You should see a csv for train and test respectively in the experiments path folder.

List of Fonts

Please use this zip to get the 70 fonts to train the model

References

• Image Data Generator using PIL
• Grad Cam visualizations
• Learning Rate Finder