Text-2-Strokes

An attempt to implement handwriting synthesis basing the paper 'Generating Sequences with Recurrent Neural Networks' by Alex Graves. GUI is built with electron.js.

Demo

Dependencies

Python libraries required:

numpy==1.19.5
matplotlib==3.2.2
torch==1.7.1

npm dependencies:

electron.js

How to use

Prerequisites : A working installation of python3 and npm.

After cloning repo, install python packages using pip install -r requirements.txt (virtual environment recommended)

And similarly install node-modules using npm install

1. Download dataset

Download IAM On-Line Handwriting Database. Place the extracted folder original-xml-part under text-2-strokes/data/.

2. Preprocess dataset

python extract_data.py

This scipt searches original-xml-part directory for xml files with handwriting data > converts coordinates to offsets > saves the output as ./data/strokes.npy and ./data/sentences.txt.

3. Train model

python train.py --n_epochs 100 --model synthesis --batch_size 32 --text_req 

A number of arguments can be set for training if you wish to experiment with the parameters. The default values are in train.py

  --hidden_size   #hidden states for LSTM layer
  --n_layers      #LSTM layer
  --batch_size    size of training batch
  --step_size     step size for learning rate decay
  --n_epochs      #Training epochs
  --lr            learning rate
  --patience      patience for early stopping
  --model_type    train model type
  --data_path     path to processed training data
  --save_path     path where training weights are stored
  --text_req      flag indicating to fetch text data also
  --data_aug      flag to whether data augmentation required
  --seed          random seed

4. Generate handwriting

python generate.py --char_seq "input text for handwriting synthesis" --save_img --style 4

Similarly a number of arguments can be set for generation also, if you wish to experiment with the parameters. The default values are in generate.py

  --model        type of model
  --model_path   path to trained weights
  --save_path    output path
  --seq_len      length of input sequence
  --bias         bias term
  --char_seq     input text
  --text_req     flag indicating to fetch text data also
  --seed         random seed
  --data_path    path to processed training data
  --style        style number [0,4]
  --save_img     save output as .png
  --save_gif     save output as .gif

To run GUI, execute:

npm start

Examples

python generate.py --char_seq "A sample of generated text" --save_gif --style 1

python generate.py --char_seq "I am Rahul" --save_gif --style 4

References

[1] Alex Graves, Generating Sequences With Recurrent Neural Networks, 2014

[2] IAM On-Line Handwriting Database, Universität Bern, 2005

[3] Swechha Choudhary's implemetation of the paper

[4] Grzegorz Opoka's implementation of the paper

Any feedback is much appreciated 😃