An application that maps an image of a LaTeX math equation to LaTeX code.
The problem of image-to-markup generation has been attempted by Deng et al. (2016). They provide the raw and preprocessed versions of im2latex-100K, a dataset consisting of about 100K LaTeX math equation images. Using their dataset, I trained a model that uses ResNet-18 as encoder with 2D positional encoding and a Transformer as decoder with cross-entropy loss. (Similar to the one described in Singh et al. (2021), except that I used ResNet only up to block 3 to reduce computational costs, and I excluded the line number encoding as it doesn't apply to this problem.)
Model architecture. Taken from Singh et al. (2021).
Initially, I used the preprocessed dataset to train my model, because the preprocessed images are downsampled to half of their original sizes for efficiency, and are grouped and padded into similar sizes to facilitate batching. However, this rigid preprocessing turned out to be a huge limitation. Although the model could achieve a reasonable performance on the test set (which was preprocessed the same way as the training set), it did not generalize well to images outside the dataset, most likely because the image quality, padding, and font size are so different from the images in the dataset. This phenomenon has also been observed by others who have attempted the same problem using the same dataset (e.g., this project, this issue and this issue).
To this end, I used the raw dataset and included image augmentation (e.g. random scaling, gaussian noise) in my data processing pipeline to increase the diversity of the samples. Moreover, unlike Deng et al. (2016), I did not group images by size. Rather, I sampled them uniformly and padded them to the size of the largest image in the batch, so that the model must learn how to adapt to different padding sizes.
Additional problems that I faced in the dataset:
- Some latex code produces visually identical outputs (e.g.
\left(
and\right)
look the same as(
and)
), so I normalized them. - Some latex code is used to add space (e.g.
\vspace{2px}
and\hspace{0.3mm}
). However, the length of the space is diffcult to judge even for humans. Also, there are many ways to express the same spacing (e.g. 1 cm = 10 mm). Finally, I don't want the model to generate code on blank images, so I just removed them.
The best run has a character error rate (CER) of 0.17 in test set. Most errors seem to come from unnecessary horizontal spacing, e.g., \;
, \,
and \qquad
. (I wasn't aware of these horizontal spacing commands. I only removed \vspace
and \hspace
during preprocessing.) Also, the model occassionally makes the text bold when it shouldn't have.
Possible improvements include:
- Do a better job cleaning the data (e.g., removing spacing commands)
- Train the model for more epochs (for the sake of time, I only trained the model for 15 epochs, but the validation loss is still going down)
- Use beam search (I only implemented greedy search)
- Use a larger model (e.g., use ResNet-34 instead of ResNet-18)
- Do some hyperparameter tuning
I didn't do any of these, because I had limited computational resources (I was using Google Colab).
Clone the repository to your computer and position your command line inside the repository folder:
git clone https://github.com/kingyiusuen/image-to-latex.git
cd image-to-latex
Then, create a virtual environment named venv
and install required packages:
make venv
make install-dev
Run the following command to download the im2latex-100k dataset and do all the preprocessing. (The image cropping step may take over an hour.)
python scripts/prepare_data.py
An example command to start a training session:
python scripts/run_experiment.py trainer.gpus=1 data.batch_size=32
Configurations can be modified in conf/config.yaml
or in command line. See Hydra's documentation to learn more.
The best model checkpoint will be uploaded to Weights & Biases (W&B) automatically (you will be asked to register or login to W&B before the training starts). Here is an example command to download a trained model checkpoint from W&B:
python scripts/download_checkpoint.py RUN_PATH
Replace RUN_PATH with the path of your run. The run path should be in the format of <entity>/<project>/<run_id>
. To find the run path for a particular experiment run, go to the Overview tab in the dashboard.
For example, you can use the following command to download my best run
python scripts/download_checkpoint.py kingyiusuen/image-to-latex/1w1abmg1
The checkpoint will be downloaded to a folder named artifacts
under the project directory.
The following tools are used to lint the codebase:
isort
: Sorts and formats import statements in Python scripts.
black
: A code formatter that adheres to PEP8.
flake8
: A code linter that reports stylistic problems in Python scripts.
mypy
: Performs static type checking in Python scripts.
Use the following command to run all the checkers and formatters:
make lint
See pyproject.toml
and setup.cfg
at the root directory for their configurations.
Similar checks are done automatically by the pre-commit framework when a commit is made. Check out .pre-commit-config.yaml
for the configurations.
An API is created to make predictions using the trained model. Use the following command to get the server up and running:
make api
You can explore the API via the generated documentation at http://0.0.0.0:8000/docs.
To run the Streamlit app, create a new terminal window and use the following command:
make streamlit
The app should be opened in your browser automatically. You can also open it by visiting http://localhost:8501. For the app to work, you need to download the artifacts of an experiment run (see above) and have the API up and running.
To create a Docker image for the API:
make docker
-
This project is inspired by the project ideas section in the final project guidelines of the course Full Stack Deep Learning at UC Berkely. Some of the code is adopted from its labs.
-
MLOps - Made with ML for introducing Makefile, pre-commit, Github Actions and Python packaging.
-
harvardnlp/im2markup for the im2latex-100k dataset.