VProChart: Answering Chart Question through Visual Perception Alignment Agent and Programmatic Solution Reasoning

🎉 News

• 2024.12.09: Our paper has been accepted by AAAI-25.

🚀 Quick Start

📥 Model Download

Click the badge below to grab the pretrained VPAgent model from ModelScope:

---

🛠️ Dependencies

Install the exact versions for full compatibility:

pip install transformers==4.28.1 \
            pytorch-lightning==1.8.5 \
            datasets \
            sentencepiece

---

🎯 Usage (Inference)

For a quick example of loading and querying VProChart, see our test script:

Programmatic Solution Reasoning is still being organized and will be released soon.

🔧 Finetuning

To adapt VProChart on your own data, consult the finetuning script:

Example CLI

python finetune_chartqa.py \
  --data-path "your_hf_dataset_name_or_local_path" \
  --train-images "/path/to/train/images/" \
  --valid-images "/path/to/val/images/" \
  --output-dir "./finetuned_model_output/" \
  --max-steps 40000 \
  --batch-size 8 \
  --valid-batch-size 1 \
  --max-length 512 \
  --num-workers 12 \
  --lr 5e-5 \
  --check-val-every-n-epoch 1 \
  --log-every-n-steps 50 \
  --warmup-steps 100 \
  --checkpoint-steps 7000 \
  --gradient-clip-val 1.0 \
  --accumulate-grad-batches 1 \
  --gpus-num 1 \
  --nodes-num 1 \
  --checkpoint-path "/path/to/vprochart_pretrained/"

📖 Citation

If you find VProChart useful in your research, please cite:

@misc{huang2024vprochartansweringchartquestion,
  title = {VProChart: Answering Chart Question through Visual Perception Alignment Agent and Programmatic Solution Reasoning},
  author = {Muye Huang and Lingling Zhang and Lai Han and Wenjun Wu and Xinyu Zhang and Jun Liu},
  year = {2024},
  eprint = {2409.01667},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url = {https://arxiv.org/abs/2409.01667},
}

---

🔗 Resources

• Paper: VProChart: Answering Chart Question through Visual Perception Alignment Agent and Programmatic Solution Reasoning
• ModelScope Hub: VProChart-VPAgent

---

🙏 Acknowledgements

This work is partially based on the UniChart project: vis-nlp/UniChart