VDLM

<h1 align="center"> Visually Descriptive Language Model for Vector Graphics Reasoning </h1> <p align="center"> <a href="https://mikewangwzhl.github.io/VDLM">🌐 Homepage</a> • <a href="https://arxiv.org/abs/2404.06479">📃 Paper</a> • <a href="https://huggingface.co/datasets/mikewang/PVD-160K" >🤗 Data (PVD-160k)</a> • <a href="https://huggingface.co/mikewang/PVD-160k-Mistral-7b" >🤗 Model (PVD-160k-Mistral-7b)</a> • <a href="https://github.com/MikeWangWZHL/VDLM" >💻 Code</a> </p>

We observe that current large multimodal models (LMMs) still struggle with seemingly straightforward reasoning tasks that require precise perception of low-level visual details, such as identifying spatial relations or solving simple mazes. In particular, this failure mode persists in question-answering tasks about vector graphics—images composed purely of 2D objects and shapes.

To solve this challenge, we propose Visually Descriptive Language Model (VDLM), a visual reasoning framework that operates with intermediate text- based visual descriptions—SVG representations and learned Primal Visual Description, which can be directly integrated into existing LLMs and LMMs. We demonstrate that VDLM outperforms state-of-the-art large multimodal models, such as GPT-4V, across various multimodal reasoning tasks involving vector graphics. See our paper for more details.

💻 Environment Setup

• Minimum requirements:

  conda env create -f environment.yml
  conda activate vdlm
  

• (Optional) For llava inference:

  cd third_party
  git clone https://github.com/haotian-liu/LLaVA.git
  cd LLaVA
  pip install -e .
  

• (Optional) For ViperGPT inference:

  cd third_party
  git clone https://github.com/MikeWangWZHL/viper.git
  

  Set up the environment for ViperGPT following the instructions.

🚀 Quick Start (Inference Demo)

• Download the pretrained SVG-to-PVD model from here. It is an LLM finetuned from Mistral-7B-v0.1. Make sure it is stored at data/ckpts/PVD-160k-Mistral-7b

  mkdir -p data/ckpts
  cd data/ckpts
  git lfs install
  git clone https://huggingface.co/mikewang/PVD-160k-Mistral-7b
  

• Serve the model with vllm:

  CUDA_VISIBLE_DEVICES=0 ./vllm_serve_model.sh
  

• A detailed inference demo 🚀 can be found here.

📊 Downstream Task Evaluation

Downstream Task Data Download

You can download the data for downstream tasks from here. Unzip the file and place the downstream_tasks folder under data/datasets/.

Run VDLM Perception: Image -> SVG -> PVD (in JSON format)

bash scripts/perception/eval_perception.sh    

Run Reasoning: PVD + question -> answer

• VDLM-mm:

  • GPT-4o:

    bash scripts/reasoning/vdlm_mm_gpt4o_pvd.sh
    

  • GPT-4V:

    bash scripts/reasoning/vdlm_mm_gpt4v_pvd.sh
    

• VDLM-txt:

  • GPT-4 Chat API without Code Interpreter:

    bash scripts/reasoning/vdlm_txt_gpt4_pvd.sh
    

  • GPT-4 Assistant API with Code Interpreter:

    bash scripts/reasoning/vdlm_txt_gpt4_assistant_pvd.sh
    

• Image-based Baselines:

  • GPT-4o + Image input:

    bash scripts/reasoning/gpt4o_image.sh
    

  • GPT-4v + Image input:

    bash scripts/reasoning/gpt4v_image.sh
    

  • Llava-v1.5 + Image input:

    # 7b
    bash scripts/reasoning/llava_1.5_7b_image.sh
    # 13b
    bash scripts/reasoning/llava_1.5_13b_image.sh
    

  • ViperGPT w/ GPT-4 + Image input:

    bash scripts/reasoning/vipergpt_inference.sh
    

📂 SVG-to-PVD Model Data

PVD-160k Dataset

The dataset used for training our SVG-to-PVD model can be downloaded from here, which contains the preprocessed instruction-tuning data instances for training the SVG-to-PVD model. The format of each line is as follows:

{
    "id": "XXX",
    "conversations": [
        {"role": "system", "content": "XXX"},
        {"role": "user", "content": "XXX"},
        {"role": "assistant", "content": "XXX"}
        // ...
    ]
}

Additioanlly, the raw PNGs, SVGs and PVD annotations generated by our data generator can be downloaded from here.

<!-- By default, the dataset is stored in `data/datasets/pretraining_data/pvd_160k.jsonl`. -->

Generating custom PVD data

pvd_data_generator/generate_pvd_img_svg.py provides the procedural data generator we used for generating the 160K Image/SVG/PVD pairs.

Example usage: bash pvd_data_generator/gen_dataset_pvd_160K.sh

To specify custom configurations, one can modify the main() function in pvd_data_generator/generate_pvd_img_svg.py.

Once generated the SVGs and PVD annotations, one can use the pvd_data_generator/get_instruction_pair.py to construct instruction-tuning data instances in vicuna or openai/mistral format. Modify the #TODO parts in the script with the generated custom dataset information. Then run: python pvd_data_generator/get_instruction_pair.py

📘 SVG-to-PVD Model Training

We finetune a Mistral-7B model using Megatron-LLM on the PVD-160K dataset. We follow https://github.com/xingyaoww/code-act/blob/main/docs/MODEL_TRAINING.md for doing the preprocessing and postprocessing on the model and data. We train the model on a SLURM cluster with 4 NVIDIA-A100-40GB GPUs.

Example usage:

• clone the code-act repo:

  cd third_party
  git clone https://github.com/xingyaoww/code-act.git
  

• Follow the instructions in https://github.com/xingyaoww/code-act/blob/main/docs/MODEL_TRAINING.md#environment-setup; for environmental setup, model preprocessing, data conversion.

• Modify the TODO: items in scripts/training/finetune_4xA100_4tp_mistral__pvd_3ep.slurm and scripts/training/finetune_4xA100_4tp_mistral__pvd_3ep.sh

• Copy scripts/training/finetune_4xA100_4tp_mistral__pvd_3ep.slurm into code-act/scripts/slurm/configs; Copy scripts/training/finetune_4xA100_4tp_mistral__pvd_3ep.sh into code-act/scripts/models/megatron.

• Run training by:

  cd third_party/code-act
  sbatch scripts/slurm/configs/finetune_4xA100_4tp_mistral__pvd_3ep.slurm scripts/models/megatron/finetune_4xA100_4tp_mistral__pvd_3ep.sh
  

• Follow https://github.com/xingyaoww/code-act/blob/main/docs/MODEL_TRAINING.md#convert-back-to-huggingface-format to convert the trained model back to Huggingface format. The converted model can be served with vllm for inference.

📚 Citation

@article{wang2024vdlm,
  title={Visually Descriptive Language Model for Vector Graphics Reasoning},
  author={Wang, Zhenhailong and Hsu, Joy and Wang, Xingyao and Huang, Kuan-Hao and Li, Manling and Wu, Jiajun and Ji, Heng},
  journal={arXiv preprint arXiv:2404.06479},
  year={2024}
}

Website License

This website's template is based on the Nerfies website.

<a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/"><img alt="Creative Commons License" style="border-width:0" src="https://i.creativecommons.org/l/by-sa/4.0/88x31.png" /></a><br />The website is licensed under a <a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/">Creative Commons Attribution-ShareAlike 4.0 International License</a>.