UniChart: A Universal Vision-language Pretrained Model for Chart Comprehension and Reasoning

• Authors: Ahmed Masry, Parsa Kavehzadeh, Do Long, Shafiq Joty, Enamul Hoque (*equal contribution)
• Paper Link: UniChart
• [NEW] If you are looking for more advanced Chart Models, explore our latest models for chart understanding:
  • ChartInstruct
    • Our advanced Chart Large Language Model based on LLaVA, supporting LLama2 (7B) and Flan-T5-XL (3B). Perfect for a wide range of chart-related tasks.
  • ChartGemma
    • The state-of-the-art Chart LLM built on PaliGemma (3B), optimized for visual reasoning tasks.
  • Both models are user-friendly and can be run with just a few lines of code. Public web demos are available! Check out their GitHub repositories for more details.

UniChart Pretraining Dataset

Our pretraining dataset is divided into two primary components:

1. A zip file encompassing all the images. You can access the images through this huggingface dataset: Images
2. A Huggingface dataset containing the input/output pairs utilized for model pretraining. You can find the dataset here: Huggingface Dataset

UniChart Model Checkpoints

We release the checkpoints for our pretrained models as well as the finetuned checkpoints on the different downstream tasks

Task	Checkpoint Path
Pretrained	unichart-base-960
ChartQA	unichart-chartqa-960
Chart2Text-Statista	unichart-chart2text-statista-960
Chart2Text-Pew	unichart-chart2text-pew-960
OpenCQA	unichart-opencqa-960

Web Demo

If you wish to quickly try our models, you can access our public web demoes hosted on the Hugging Face Spaces platform with a friendly interface!

Tasks	Web Demo
Base Model (Best for Chart Summarization and Data Table Generation)	UniChart-Base
Chart Question Answering	UniChart-ChartQA

The input prompt for Chart summarization is <summarize_chart> and Data Table Generation is <extract_data_table>

Requirements

transformers==4.28.1
pytorch-lightning==1.8.5
datasets
sentencepiece

Please make sure to use the exact same version of the Transformers library. We have noticed that there might be a drop in performance when using different versions of the library!

Inference

You can easily use our models for inference with the huggingface library! You just need to do the following:

1. Change model_name to your prefered checkpoint.
2. Chage the imag_path to your chart example image path on your system
3. Write the input_prompt based on your prefered task as shown in the table below.

Task	Input Prompt
Chart Question Answering	<chartqa> question <s_answer>
Open Chart Question Answering	<opencqa> question <s_answer>
Chart Summarization	<summarize_chart> <s_answer>
Data Table Extraction	<extract_data_table> <s_answer>

from transformers import DonutProcessor, VisionEncoderDecoderModel
from PIL import Image
import torch, os, re

torch.hub.download_url_to_file('https://raw.githubusercontent.com/vis-nlp/ChartQA/main/ChartQA%20Dataset/val/png/multi_col_1229.png', 'chart_example_1.png')

model_name = "ahmed-masry/unichart-chartqa-960"
image_path = "/content/chart_example_1.png"
input_prompt = "<chartqa> What is the lowest value in blue bar? <s_answer>"

model = VisionEncoderDecoderModel.from_pretrained(model_name)
processor = DonutProcessor.from_pretrained(model_name)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

image = Image.open(image_path).convert("RGB")
decoder_input_ids = processor.tokenizer(input_prompt, add_special_tokens=False, return_tensors="pt").input_ids
pixel_values = processor(image, return_tensors="pt").pixel_values

outputs = model.generate(
    pixel_values.to(device),
    decoder_input_ids=decoder_input_ids.to(device),
    max_length=model.decoder.config.max_position_embeddings,
    early_stopping=True,
    pad_token_id=processor.tokenizer.pad_token_id,
    eos_token_id=processor.tokenizer.eos_token_id,
    use_cache=True,
    num_beams=4,
    bad_words_ids=[[processor.tokenizer.unk_token_id]],
    return_dict_in_generate=True,
)
sequence = processor.batch_decode(outputs.sequences)[0]
sequence = sequence.replace(processor.tokenizer.eos_token, "").replace(processor.tokenizer.pad_token, "")
sequence = sequence.split("<s_answer>")[1].strip()
print(sequence)

Finetuning

In order to finetune the model on the ChartQA dataset, you can edit and run the following command:

python finetune_chartqa.py --data-path "ahmed-masry/chartqa_without_images" --train-images '/content/ChartQA/ChartQA Dataset/train/png/' \
    --valid-images '/content/ChartQA/ChartQA Dataset/val/png' --max-steps 40000 --batch-size 8 --valid-batch-size 1 --num-workers 12 --lr 5e-5 \
    --check-val-every-n-epoch 1 --warmup-steps 100 --checkpoint-steps 7000 --checkpoint-path "ahmed-masry/unichart-base-960"

Contact

If you have any questions about this work, please contact Ahmed Masry using the following email addresses: [email protected] or [email protected].

Reference

Please cite our paper if you use our models or dataset in your research.

@misc{masry2023unichart,
      title={UniChart: A Universal Vision-language Pretrained Model for Chart Comprehension and Reasoning}, 
      author={Ahmed Masry and Parsa Kavehzadeh and Xuan Long Do and Enamul Hoque and Shafiq Joty},
      year={2023},
      eprint={2305.14761},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}