Based on Multi-Scale Attention with Dense Encoder for Handwritten Mathematical Expression Recognition.

• Python 3
• PyTorch

All dependencies can be installed with PIP.

pip install -r requirements.txt

If you'd like to use a different installation method or another CUDA version with PyTorch (e.g. CUDA 10) follow the instructions on PyTorch - Getting Started.

CROHME: Competition on Recognition of Online Handwritten Mathematical Expressions has been used. As it is an on-line handwritten dataset, it consists of InkML files, but this architecture is for off-line recognition, which means that images are used as input.

The dataset has been converted to images of size 256x256 and the ground truth has been extracted as well. The converted dataset can be found at Kaggle - crohme-png.

The data needs to be in the data/ directory and a tokens.tsv file defines the available tokens separated by tabs. Training and validation sets are defined in gt_split/train.tsv and gt_split/validation.tsv, where each line is the path to the image and its ground truth.

The training/validation split can be generated by running:

python data_tools/train_validation_split.py -i data/groundtruth_train.tsv -o data/gt_split

Note: The content of the generated images vary greatly in size. As longer expressions are limited to the same width, they will essentially use a smaller font. This makes it much more difficult to correctly predict the sequences, especially since the dataset is quite small. The primary focus was the attention mechanism, to see whether it can handle different sizes. If you want better results, the images need to be normalised.

Training is done with the train.py script:

python train.py --prefix "some-name-" -n 200 -c checkpoints/example-0022.pth

The --prefix option is used to give it a name, otherwise the checkpoints are just numbered without any given name and -c is to resume from the given checkpoint, if not specified it starts fresh.

For all options see python train.py --help:

To evaluate a model use the evaluate.py script with the desired checkpoint and the dataset it should be tested against (can use multiple sets at once):

For example to evaluate the sets 2014 and 2016 with beam width 5:

python evaluate.py -d 2014 2016 --beam-width 5 -c checkpoints/example-0022.pth