Characters Are Like Faces

This is the source code for ICLR 2023 TinyPaper Characters Are Like Faces

File explanation

character_argumentation.py: part of argumentation approach mentioned in the paper

dataset.py: the dataset class used to load training data

evaluate.py: evaluation through our website(we didn't utilize python with pytorch to evaluate simply because it's too slow. We established a local http server to do this, code can be found in the web directory)

export_model.py: to export our trained model to onnx format, making it convenience to deploy on our web server

generate_database.py: generate the final database using font files in eval_fonts directory with trained model.

modules.py: the definition of our network, including MobileFaceNetV3 and Arcface Loss

test_database.py: reconginze the test.png file and give out the possible result using python with pytorch, it's super slow!

train.py: training script

training_data_generation.py: generate our training data using font files in the fonts directory

Usage

To train your own network, you might first run traning_data_generation.py to generate the training data

python traning_data_generation.py

Then, run train.py to train the network

python train.py

After training, run generate_data_base.py to generate the database for query

python generate_database.py

At last, run test_database.py to scognize the file test.png

python test_database.py

note that if you want to evalueate the accuracy on a certain font, you may modify evaluate.py. This file recognize a character via a web API (we establish it on a local server). The code for the server is in web directory. You should first put the database into the resource directory since it's a little big that we did not upload to the repository.