Ucas-HaoranWei%2FVary | Trendshift

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

• [2024年12月24日] 🔥🔥🔥 My new work on system-2 perception is released slow-perception.
• [2024年9月03日] 🔥🔥🔥 We release a very strong and comprehensive OCR model GOT-OCR2.0.
• [2024年7月16日] 🎉🎉🎉 OneChart is accepted by ACM'MM 2024 oral (3.97%)!
• [2024年7月2日] 🔥🔥🔥 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年5月27日] 🔥🔥🔥 We present a document understanding benchmark in Fox .
• [2024年5月24日] 🔥🔥🔥 We propose a multi-page document understanding work -- Fox, which supports 8-page pdf-image input !!!
• [2024年4月21日] 🔥🔥🔥 For OneChart, we have released the web demo in Project Page. Have fun!!
• [2024年4月21日] 🔥🔥🔥 We present a Vary-tiny LAVIS codebase (for training from scratch) and the Vary-600k dataset (300K English and 300K Chinese pages) here !!!
• [2024年4月15日]🔥🔥🔥We release a chart parsing model OneChart here.
• [2024年4月12日]🔥🔥🔥We will release a chart parsing model based on Vary-tiny next week. The model supports both English and Chinese charts.
• [2024年3月16日]🔥🔥🔥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.
• [2023年1月23日]🔥🔥🔥We release the Vary-toy here. Besides, we show the super good Vary-family results here.
• [2023年12月29日]🔥🔥🔥We will release a new model (a small-size Vary, about 2B) at the beginning of next month and introduce a new feature (object detection). Our online demo will be temporarily closed to prepare for the deployment of the new model.
• [2023年12月11日] We released the online demo, have fun!
• [2023年12月11日] We released the codes of Vary (train and inference)!

Code License Data License 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.

• Install
• Vary Weights
• Demo
• Train

1. Clone this repository and navigate to the Vary folder

git clone https://github.com/Ucas-HaoranWei/Vary.git
cd Vary

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

• If you are in urgent need of weights for your research recently, please contact me by email.
• Download the CLIP-VIT-L in Hugging Face
• Here for Vary-toy weights.

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

• We currently do not plan to open source the weights of the intermediate.
• However, we release the train codes. So you can train on your own dataset. If you want to do this, you can try this:

1. For Vary-base (one machine, if you have multiple machines you need to prepare your host file)

deepspeed Vary/train/train_qwen_vary.py --deepspeed /Vary/zero_config/zero2.json
 --model_name_or_path /Qwen-7B/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/

2. For Vary-tiny

deepspeed Vary/train/train_opt.py --deepspeed /Vary/zero_config/zero2.json
 --model_name_or_path /opt125m/path/
 --conversation_version opt
 --freeze_vision_tower False
 --freeze_lm_model False
 --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 16
 --num_train_epochs 1
 --learning_rate 5e-5
 --datasets data_name1+data_name2+data_name3
 --output_dir /path/to/output/

If you have any questions related to the code or the paper, feel free to email (weihaoran18@mails.ucas.ac.cn).

• LLaVA: the codebase we built upon!
• Qwen: the LLM base model of Vary, which is good at both English and Chinese!

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}
}