Lin Song, Yanwei Li, Rui Yang, Sijie Zhao, Yixiao Ge, Xiu Li, Ying Shan

GPT4Tools is a centralized system that can control multiple visual foundation models. It is based on Vicuna (LLaMA), and 71K self-built instruction data. By analyzing the language content, GPT4Tools is capable of automatically deciding, controlling, and utilizing different visual foundation models, allowing the user to interact with images during a conversation. With this approach, GPT4Tools provides a seamless and efficient solution to fulfill various image-related requirements in a conversation. Different from previous work, we support users teach their own LLM to use tools with simple refinement via self-instruction and LoRA.

• 🔥 We update new code and models adapted to vicuna-v1.5!
• 🔥 Our paper is accpeted by NIPS 2023!
• 🔥 We now release the paper and new demo with LLAVA, OPT, LlaMA and Vicuna.
• 🔥 We released pretrained GPT4Tools models with Vicuna-13B and released the dataset for self-instruction. Check out the blog and demo.

We provide some selected examples using GPT4Tools in this section. More examples can be found in our project page. Feel free to try our onlin demo!

More demos

• gpt4tools_71k.json contains 71K instruction-following data we used for fine-tuning the GPT4Tools model.

• gpt4tools_val_seen.json is the manually cleaned instruction data used for validation, which includes instructions related to tools of gpt4tools_71k.json.

• gpt4tools_test_unseen.json cleaned instruction data used for testing, including instructions related to some tools that are absented in gpt4tools_71k.json.

data.md shows how to generate, format and clean the data.

GTP4Tools mainly contains three parts: LLM for instruction, LoRA for adaptation, and Visual Agent for provided functions. It is a flexible and extensible system that can be easily extended to support more tools and functions. For example, users can replace the existing LLM or tools with their own models, or add new tools to the system. The only things needed are finetuned the LoRA with the provided instruction, which teaches LLM to use the provided tools.

GPT4Tools is based on the Vicuna, we release the LoRA weights of GPT4Tools to comply with the LLaMA model license. You can merge our LoRA weights with the Vicuna weights to obtain the GPT4Tools weights.

git clone https://github.com/AILab-CVC/GPT4Tools
cd GPT4Tools
pip install -r requirements.txt

1. Download vicuna-7b-v1.5 and vicuna-13b-v1.5.

# download to your cache dir
python3 scripts/download.py \
	--model-names "lmsys/vicuna-13b-v1.5" "lmsys/vicuna-7b-v1.5" \
	--cache-dir $your_cache_dir

2. Download gpt4tools LoRA weights following the below links:

Old weights can be found here.

GPT4Tools supports 22 tools. Please check tools.md for more details. When using tools for the first time, the weights of tools need to be downloaded to cache. If you don't like stored them on default cache, please revise the shell environment varibles:

export TRANSFORMERS_CACHE=${your_transformers_cache}
export HUGGINGFACE_HUB_CACHE=${your_diffusers_cache}

Additionally, you can also download weights to the custom cache.

# download huggingface model
python3 scripts/download.py \
	--model-names "Salesforce/blip-image-captioning-base" "Salesforce/blip-vqa-base" "timbrooks/instruct-pix2pix" "runwayml/stable-diffusion-v1-5" "runwayml/stable-diffusion-inpainting" "lllyasviel/ControlNet" "fusing/stable-diffusion-v1-5-controlnet-canny" "fusing/stable-diffusion-v1-5-controlnet-mlsd" "fusing/stable-diffusion-v1-5-controlnet-hed" "fusing/stable-diffusion-v1-5-controlnet-scribble" "fusing/stable-diffusion-v1-5-controlnet-openpose" "fusing/stable-diffusion-v1-5-controlnet-seg" "fusing/stable-diffusion-v1-5-controlnet-depth" "fusing/stable-diffusion-v1-5-controlnet-normal" "sam" "groundingdino" \
	--cache-dir $your_cache_dir

Following scripts/demo.sh or the below code to make a gradio interface on your own devices:

# Advice for 1 GPU
python gpt4tools_demo.py \
	--base_model $path_to_vicuna_with_tokenizer \
	--lora_model $path_to_lora_weights \
	--llm_device "cpu" \ 
	--load "Text2Box_cuda:0,Segmenting_cuda:0,Inpainting_cuda:0,ImageCaptioning_cuda:0" \ 
	--cache-dir $your_cache_dir \
	--server-port 29509 \
	--share

# Advice for 4 GPUs
python gpt4tools_demo.py \
	--base_model $path_to_vicuna_with_tokenizer
	--lora_model $path_to_lora_weights \
	--llm_device "cuda:3" \
	--load "Text2Box_cuda:0,Segmenting_cuda:0,Inpainting_cuda:0,ImageCaptioning_cuda:0,Text2Image_cuda:1,VisualQuestionAnswering_cuda:1,InstructPix2Pix_cuda:2,SegText2Image_cuda:2,Image2Pose_cpu,PoseText2Image_cuda:2" \
	--cache-dir $your_cache_dir \
	--server-port 29509 \
	--share

You can customize the used tools by specifying {tools_name}_{devices} after args --load of gpt4tools_demo.py. tools_name is illustrated in tools.md.

After downloading the gpt4tools_71k.json to ./datasets, you can follow scripts/finetune_lora.sh or run the below code to finetune your model:

deepspeed train.py \
	--base_model $path_to_vicuna_with_tokenizer \
	--data_path $path_to_gpt4tools_71k.json \
	--deepspeed "scripts/zero2.json" \
	--output_dir output/gpt4tools \
	--num_epochs 6 \
	--per_device_train_batch_size 1 \
	--per_device_eval_batch_size 4 \
	--gradient_accumulation_steps 16 \
	--model_max_length 2048 \
	--lora_target_modules '[q_proj,k_proj,v_proj,o_proj]' \
	--lora_r 16 \
	--learning_rate 3e-4 \
	--lazy_preprocess True \
	--cache_dir $your_cache_dir \
	--report_to 'tensorboard' \
	--gradient_checkpointing True

If you want to evaluate the model's successful rate of using tools, please locate here.

• VisualChatGPT: It connects ChatGPT and a series of Visual Foundation Models to enable sending and receiving images during chatting.
• Vicuna: The language ability of Vicuna is fantastic and amazing. And it is open-source!
• Alpaca-LoRA: Instruct-tune LLaMA on consumer hardware.

If you're using our GPT4Tools in your research or applications, please cite:

@misc{gpt4tools,
  title = {GPT4Tools: Teaching LLM to Use Tools via Self-instruction},
  author={Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, Ying Shan},
  journal={arXiv preprint arXiv:2305.18752},
  year={2023}
}