Small Language Model Meets with Reinforced Vision Vocabulary

Haoran Wei*, Lingyu Kong*, Jinyue Chen, Liang Zhao, Zheng Ge, En Yu, Jianjian Sun, Chunrui Han, Xiangyu Zhang

The Young's First ``Large'' Vision Language Model

Release

• [2024/9/03] 🔥🔥🔥 We release a very strong and comprehensive OCR model GOT-OCR2.0.
• [2024/7/21] 🎉🎉🎉 OneChart is accepted by ACM'MM 2024 Oral! (3.97%)
• [2024/7/2] 🔥🔥🔥 Vary is accepted by ECCV2024. To thank everyone for their attention, I will release a model that performs on par with the Vary-document soon.
• [2024/5/27] 🔥🔥🔥 We present a document understanding benchmark in Fox .
• [2024/5/24] 🔥🔥🔥 We propose a multi-page document understanding work -- Fox, which supports 8-page pdf-image input !!!
• [2024/4/21] 🔥🔥🔥 For OneChart, we have released the web demo in Project Page. Have fun!!
• [2024/4/21] 🔥🔥🔥 We present a Vary-tiny LAVIS codebase (for training from scratch) and the Vary-600k dataset (300K English and 300K Chinese pages) here !!!
• [2024/4/15]🔥🔥🔥We release a chart parsing model OneChart here.
• [2024/4/12]🔥🔥🔥We will release a chart parsing model based on Vary-tiny next week. The model supports both English and Chinese charts.
• [2024/3/16]🔥🔥🔥I found many friends very interested in Vary-tiny(OPT-125M), so I opened source it here, a PDF-dense OCR and object detection version.
• [2024/1/23] 🔥Eval codes will be available soon.
• [2024/1/23] 🔥🔥🔥You only need a single 1080Ti to experience all features of current LVLMs.

Usage and License Notices: The data, code, and checkpoint are intended and licensed for research use only. They are also restricted to use that follow the license agreement of LLaMA, Vicuna, GPT-4, Qwen, and LLaVA.

Contents

• Install
• Vary-toy Weights
• Demo
• Train

Note

If you have built the original Vary, please rebuild this repo !!!

Install

1. Clone this repository and navigate to the Vary folder

git clone https://github.com/Ucas-HaoranWei/Vary-toy.git
cd /path/to/vary-toy

2. Install Package

conda create -n vary python=3.10 -y
conda activate vary
pip install e .

3. Install Flash-Attention

pip install ninja
pip install flash-attn --no-build-isolation

Vary Weights

• Download the Vary-toy weights here.
• Download the CLIP-VIT-L here.

Demo

1. Update the CLIP-VIT path in the codes (/cache/vit-large-patch14/) to your path.

python vary/demo/run_qwen_vary.py  --model-name  /vary/model/path/ --image-file /an/image/file.png

Train

deepspeed   Vary/train/train_qwen_vary.py  --deepspeed /Vary/zero_config/zero2.json
            --model_name_or_path /Vary-toy/path/
            --vision_tower /vit-large-patch14/path/
            --freeze_vision_tower True
            --freeze_lm_model False
            --vision_select_layer  -2
            --use_im_start_end True
            --bf16 True
            --per_device_eval_batch_size 4
            --gradient_accumulation_steps 1
            --evaluation_strategy "no"
            --save_strategy "steps"
            --save_steps 5000
            --save_total_limit 1
            --weight_decay 0.
            --warmup_ratio 0.03
            --lr_scheduler_type "cosine"
            --logging_steps 1 --tf32 True
            --model_max_length 4096
            --gradient_checkpointing True
            --dataloader_num_workers 4
            --report_to none
            --per_device_train_batch_size 4
            --num_train_epochs 1
            --learning_rate 5e-5
            --datasets  data_name1+data_name2+data_name3
            --output_dir /path/to/output/

We encourage you to extract the new vision vocabulary weights for your new base language model !!!

Contact

If you have any questions about the code or the paper, please email ([email protected]).

Discussion

Vary-toy is not a toy, and we have designed two excellent models based on it, one is Vary-document (specifically for document/pdf processing), and the other is Vary-plot for chart analysis. You can see their amazing performance here Vary-family.

Citation

If you find our work useful in your research, please consider citing Vary:

@article{wei2023vary,
  title={Vary: Scaling up the Vision Vocabulary for Large Vision-Language Models},
  author={Wei, Haoran and Kong, Lingyu and Chen, Jinyue and Zhao, Liang and Ge, Zheng and Yang, Jinrong and Sun, Jianjian and Han, Chunrui and Zhang, Xiangyu},
  journal={arXiv preprint arXiv:2312.06109},
  year={2023}
}

@article{wei2024small,
  title={Small Language Model Meets with Reinforced Vision Vocabulary},
  author={Wei, Haoran and Kong, Lingyu and Chen, Jinyue and Zhao, Liang and Ge, Zheng and Yu, En and Sun, Jianjian and Han, Chunrui and Zhang, Xiangyu},
  journal={arXiv preprint arXiv:2401.12503},
  year={2024}
}