im2latex
This model is a base VisionEncoderDecoderModel fine-tuned on a dataset for generating LaTeX formulas from images.
Model Details
- Encoder: Swin Transformer
- Decoder: GPT-2
- Framework: PyTorch
- DDP (Distributed Data Parallel): Used for training
Training Data
The data is taken from OleehyO/latex-formulas. The data was divided into 80:10:10 for train, val and test. The splits were made as follows:
dataset = load_dataset(OleehyO/latex-formulas, cleaned_formulas)
train_val_split = dataset["train"].train_test_split(test_size=0.2, seed=42)
train_ds = train_val_split["train"]
val_test_split = train_val_split["test"].train_test_split(test_size=0.5, seed=42)
val_ds = val_test_split["train"]
test_ds = val_test_split["test"]
Evaluation Metrics
The model was evaluated on a test set with the following results:
- Test Loss: 0.10
- Test BLEU Score: 0.67
Usage
You can use the model directly with the transformers library:
from transformers import VisionEncoderDecoderModel, AutoTokenizer, AutoFeatureExtractor
import torch
from PIL import Image
# load model, tokenizer, and feature extractor
model = VisionEncoderDecoderModel.from_pretrained("DGurgurov/im2latex")
tokenizer = AutoTokenizer.from_pretrained("DGurgurov/im2latex")
feature_extractor = AutoFeatureExtractor.from_pretrained("microsoft/swin-base-patch4-window7-224-in22k") # using the original feature extractor for now
# prepare an image
image = Image.open("path/to/your/image.png")
pixel_values = feature_extractor(images=image, return_tensors="pt").pixel_values
# generate LaTeX formula
generated_ids = model.generate(pixel_values)
generated_texts = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
print("Generated LaTeX formula:", generated_texts[0])
Training Script
The training script for this model can be found in the following repository: GitHub
Citation:
- If you use this work in your research, please cite our paper:
@misc{gurgurov2024imagetolatexconvertermathematicalformulas,
title={Image-to-LaTeX Converter for Mathematical Formulas and Text},
author={Daniil Gurgurov and Aleksey Morshnev},
year={2024},
eprint={2408.04015},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2408.04015},
}
License [MIT]