The Chinese Visual Embeddings

The Chinese Visual Embeddings are trained for the purpose of visual representations for the Chinese written signs. It is part of a larger project Graphein by Qianxun Chen

Training Chinese Visual Embeddings

STEP 0 Setup the environment

• Download 'NotoSansCJKsc-Regular' and 'NotoSansCJKtc-Regular' from Google Noto Fonts and save them to folder 'fonts'
• Create a new virtual environment with python 3 virtualenv --python=python3.6 venv
• Activate the environment activate source venv/bin/activate
• Install all the packages pip3 install -r requirements.txt

STEP 1 Prepare the data

python preprocess.py -img

STEP 2 Train embeddings

• An earlier version of the embeddings was trained with python CNN.py
• The latest version of the embeddings can be trained with python CNN_cuda_multiLabel.py

STEP3 Generate supporting files

• Generate tsv files to preview embeddings in Embedding Projector python embeddings/VC/generateTSV.py
• Generate word2vec format embeddings python embeddings/VC/tsvs2txt.py

API

Acknowledgement

dict.csv is created based on the dictionary.txt file from the project, makemehanzi.