Home

Awesome

TinyGPT-V

<font size='5'>TinyGPT-V: Efficient Multimodal Large Language Model via Small Backbones</font>

Zhengqing Yuan✟, Zhaoxu Li❁, Weiran Huang❋, Yanfang Ye✟, Lichao Sun

✟University of Notre Dame, ❁Lehigh University, ❋Shanghai Jiao Tong University

Zhaoxu is a visiting student in the LAIR lab at Lehigh University.

</a> <a href='https://arxiv.org/abs/2312.16862'><img src='https://img.shields.io/badge/Paper-Arxiv-red'></a> <a href='https://huggingface.co/Tyrannosaurus/TinyGPT-V'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Model-blue'></a> <a href='https://huggingface.co/spaces/llizhx/TinyGPT-V'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue'>

English | 简体中文

</font>

News

[Apr.08 2024] Update our paper v2. We revised some type errors, provided more details and updated TinyGPT-V lastest results.

[Mar.20 2024] Update the Phi-2 weight download link.

[Jan.22 2024] Welcome to Hugging Face online demo to try out our models (for Stage-4 v1)!

[Jan.19 2024] Major Updates! We are officially releasing v1 of TinyGPT-V! After our evaluation, the performance of TinyGPT-V has reached 98% of InstructBLIP's performance and exceeds the performance of other models of the same period!

[Jan.03 2024] Welcome to Hugging Face online demo to try out our models (for Stage-3)!

[Dec.28 2023] Breaking! We release the code of our TinyGPT-V.

TinyGPT-V Model Structure

Whole Model Structure

Model

Language Model Structure

Model

TinyGPT-V Traning Process

Traning_Process

TinyGPT-V Results

Radar Chart

Results

Performance and Efficiency

Results

Getting Started

Installation

1. Prepare the code and the environment

Git clone our repository, creating a python environment and activate it via the following command

git clone https://github.com/DLYuanGod/TinyGPT-V.git
cd TinyGPT-V
conda env create -f environment.yml
conda activate tinygptv

2. Prepare the pretrained LLM weights

TinyGPT-V is based on Phi-2. Download the corresponding LLM weights from the following huggingface space via clone the repository using git-lfs.

Phi-2 2.7B: Download

Then, set the variable phi_model in the model config file to the LLM weight path.

3. Prepare the pretrained model checkpoints

Download the pretrained model checkpoints

After stage-1After stage-2After stage-3After stage-4
DownloadDownloadDownloadDownload

For TinyGPT-V, set the path to the pretrained checkpoint in the evaluation config file in tinygptv_stage1_2_3_eval.yaml at Line 8 for Stage 1, 2 and 3 version or tinygptv_stage4_eval.yaml for Stage 4 version.

4. Update the Phi-2 Modeling for transformers lib.

Linux system:

cp modeling_phi.py /root/miniconda3/envs/tinygptv/lib/python3.9/site-packages/transformers/models/phi/

Windows system

Find your conda yourself: conda_sit/envs/tinygptv/lib/python3.9/site-packages/transformers/models/phi/ Replace modeling_phi.py in that directory with the one in TinyGPT-V/modeling_phi.py.

Launching Demo Locally

For Stage 4, run

python demo_v2.py --cfg-path eval_configs/tinygptv_stage4_eval.yaml  --gpu-id 0

Note: Stage 4 will have some Grounding abilities. But the performance is not very good, we are working on this!

For Stage 1, 2 and 3, run

python demo.py --cfg-path eval_configs/tinygptv_stage1_2_3_eval.yaml  --gpu-id 0

To perfer more powerful model, LLMs loads as 16 bit by default. This configuration requires about 8G GPU memory. To more save GPU memory, you can run the model in 8 bit below 8G device by setting low_resource to True in the relevant config file:

Training

First you need to adjust all the updated weights in the LLM to be calculated with full precision:Here. Remove the comments from the following lines:

                layer.self_attn.q_layernorm.weight.data = layer.self_attn.q_layernorm.weight.data.float()
                layer.self_attn.k_layernorm.weight.data = layer.self_attn.k_layernorm.weight.data.float()
                layer.post_layernorm.weight.data = layer.post_layernorm.weight.data.float()
                layer.input_layernorm.weight.data = layer.input_layernorm.weight.data.float()

                # Perform a similar operation for the bias item
                if layer.self_attn.q_layernorm.bias is not None:
                    layer.self_attn.q_layernorm.bias.data = layer.self_attn.q_layernorm.bias.data.float()
                if layer.self_attn.k_layernorm.bias is not None:
                    layer.self_attn.k_layernorm.bias.data = layer.self_attn.k_layernorm.bias.data.float()
                if layer.input_layernorm.bias is not None:
                    layer.input_layernorm.bias.data = layer.input_layernorm.bias.data.float()


            llama_model.model.model.final_layernorm.weight.requires_grad = True
            llama_model.model.model.final_layernorm.weight.data = llama_model.model.model.final_layernorm.weight.data.float()
            if llama_model.model.model.final_layernorm.bias is not None:
                llama_model.model.model.final_layernorm.bias.data = llama_model.model.model.final_layernorm.bias.float()

Stage 1 and 2:

torchrun --nproc-per-node NUM_GPU train.py --cfg-path train_configs/tinygptv_stage1.yaml

You need to execute the above code 17 times to complete the first stage of training.

torchrun --nproc-per-node NUM_GPU train.py --cfg-path train_configs/tinygptv_stage2.yaml

Stage 3:

torchrun --nproc-per-node NUM_GPU train.py --cfg-path train_configs/tinygptv_stage3.yaml

Stage 4:

torchrun --nproc-per-node NUM_GPU train.py --cfg-path train_configs/tinygptv_stage4.yaml

Evaluation

For eval. details of TinyGPT-V, check here

Star History

Star History Chart

Acknowledgement

If you're using TinyGPT-V in your research or applications, please cite using this BibTeX:


@misc{yuan2024tinygptv,
      title={TinyGPT-V: Efficient Multimodal Large Language Model via Small Backbones}, 
      author={Zhengqing Yuan and Zhaoxu Li and Weiran Huang and Yanfang Ye and Lichao Sun},
      year={2024},
      eprint={2312.16862},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

License

This repository is under BSD 3-Clause License. Many codes are based on Lavis with BSD 3-Clause License here.