Visual instruction tuning towards large language and vision models with GPT-4 level capabilities.
[Project Page] [Demo] [Data] [Model Zoo]
🤝Community Contributions: [llama.cpp] [Colab] [🤗Space]
Improved Baselines with Visual Instruction Tuning [Paper]
Haotian Liu, Chunyuan Li, Yuheng Li, Yong Jae Lee
Visual Instruction Tuning (NeurIPS 2023, Oral) [Paper]
Haotian Liu*, Chunyuan Li*, Qingyang Wu, Yong Jae Lee (*Equal Contribution)
- [10/12] 🔥 Check out the Korean LLaVA (Ko-LLaVA), created by ETRI, who has generously supported our research! [🤗 Demo]
- [10/12] LLaVA is now supported in llama.cpp with 4-bit / 5-bit quantization support!
- [10/11] The training data and scripts of LLaVA-1.5 are released here, and evaluation scripts are released here!
- [10/5] 🔥 LLaVA-1.5 is out! Achieving SoTA on 11 benchmarks, with just simple modifications to the original LLaVA, utilizes all public data, completes training in ~1 day on a single 8-A100 node, and surpasses methods like Qwen-VL-Chat that use billion-scale data. Check out the technical report, and explore the demo! Models are available in Model Zoo.
- [9/26] LLaVA is improved with reinforcement learning from human feedback (RLHF) to improve fact grounding and reduce hallucination. Check out the new SFT and RLHF checkpoints at project [LLavA-RLHF]
- [9/22] LLaVA is accepted by NeurIPS 2023 as oral presentation, and LLaVA-Med is accepted by NeurIPS 2023 Datasets and Benchmarks Track as spotlight presentation.
- [9/20] We summarize our empirical study of training 33B and 65B LLaVA models in a note. Further, if you are interested in the comprehensive review, evolution and trend of multimodal foundation models, please check out our recent survey paper ``Multimodal Foundation Models: From Specialists to General-Purpose Assistants''.
If you are using Windows, do NOT proceed, see instructions here.
- Clone this repository and navigate to LLaVA folder
git clone https://github.com/haotian-liu/LLaVA.git
cd LLaVA- Install Package
conda create -n llava python=3.10 -y
conda activate llava
pip install --upgrade pip # enable PEP 660 support
pip install -e .- Install additional packages for training cases
pip install -e ".[train]"
pip install flash-attn --no-build-isolation
git pull
pip install -e .Please check out our Model Zoo for all public LLaVA checkpoints, and the instructions of how to use the weights.
To run our demo, you need to prepare LLaVA checkpoints locally. Please follow the instructions here to download the checkpoints.
To launch a Gradio demo locally, please run the following commands one by one. If you plan to launch multiple model workers to compare between different checkpoints, you only need to launch the controller and the web server ONCE.
python -m llava.serve.controller --host 0.0.0.0 --port 10000python -m llava.serve.gradio_web_server --controller http://localhost:10000 --model-list-mode reloadYou just launched the Gradio web interface. Now, you can open the web interface with the URL printed on the screen. You may notice that there is no model in the model list. Do not worry, as we have not launched any model worker yet. It will be automatically updated when you launch a model worker.
This is the actual worker that performs the inference on the GPU. Each worker is responsible for a single model specified in --model-path.
python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path liuhaotian/llava-v1.5-13bWait until the process finishes loading the model and you see "Uvicorn running on ...". Now, refresh your Gradio web UI, and you will see the model you just launched in the model list.
You can launch as many workers as you want, and compare between different model checkpoints in the same Gradio interface. Please keep the --controller the same, and modify the --port and --worker to a different port number for each worker.
python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port <different from 40000, say 40001> --worker http://localhost:<change accordingly, i.e. 40001> --model-path <ckpt2>If you are using an Apple device with an M1 or M2 chip, you can specify the mps device by using the --device flag: --device mps.
If the VRAM of your GPU is less than 24GB (e.g., RTX 3090, RTX 4090, etc.), you may try running it with multiple GPUs. Our latest code base will automatically try to use multiple GPUs if you have more than one GPU. You can specify which GPUs to use with CUDA_VISIBLE_DEVICES. Below is an example of running with the first two GPUs.
CUDA_VISIBLE_DEVICES=0,1 python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path liuhaotian/llava-v1.5-13bYou can launch the model worker with quantized bits (4-bit, 8-bit), which allows you to run the inference with reduced GPU memory footprint, potentially allowing you to run on a GPU with as few as 12GB VRAM. Note that inference with quantized bits may not be as accurate as the full-precision model. Simply append --load-4bit or --load-8bit to the model worker command that you are executing. Below is an example of running with 4-bit quantization.
python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path liuhaotian/llava-v1.5-13b --load-4bitYou can launch the model worker with LoRA weights, without merging them with the base checkpoint, to save disk space. There will be additional loading time, while the inference speed is the same as the merged checkpoints. Unmerged LoRA checkpoints do not have lora-merge in the model name, and are usually much smaller (less than 1GB) than the merged checkpoints (13G for 7B, and 25G for 13B).
To load unmerged LoRA weights, you simply need to pass an additional argument --model-base, which is the base LLM that is used to train the LoRA weights. You can check the base LLM of each LoRA weights in the model zoo.
python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path liuhaotian/llava-v1-0719-336px-lora-vicuna-13b-v1.3 --model-base lmsys/vicuna-13b-v1.3Chat about images using LLaVA without the need of Gradio interface. It also supports multiple GPUs, 4-bit and 8-bit quantized inference. With 4-bit quantization, for our LLaVA-1.5-7B, it uses less than 8GB VRAM on a single GPU.
python -m llava.serve.cli \
--model-path liuhaotian/llava-v1.5-7b \
--image-file "https://llava-vl.github.io/static/images/view.jpg" \
--load-4bit
Below is the latest training configuration for LLaVA v1.5. For legacy models, please refer to README of this version for now. We'll add them in a separate doc later.
LLaVA training consists of two stages: (1) feature alignment stage: use our 558K subset of the LAION-CC-SBU dataset to connect a frozen pretrained vision encoder to a frozen LLM; (2) visual instruction tuning stage: use 150K GPT-generated multimodal instruction-following data, plus around 515K VQA data from academic-oriented tasks, to teach the model to follow multimodal instructions.
LLaVA is trained on 8 A100 GPUs with 80GB memory. To train on fewer GPUs, you can reduce the per_device_train_batch_size and increase the gradient_accumulation_steps accordingly. Always keep the global batch size the same: per_device_train_batch_size x gradient_accumulation_steps x num_gpus.
We use a similar set of hyperparameters as Vicuna in finetuning. Both hyperparameters used in pretraining and finetuning are provided below.
- Pretraining
| Hyperparameter | Global Batch Size | Learning rate | Epochs | Max length | Weight decay |
|---|---|---|---|---|---|
| LLaVA-v1.5-13B | 256 | 1e-3 | 1 | 2048 | 0 |
- Finetuning
| Hyperparameter | Global Batch Size | Learning rate | Epochs | Max length | Weight decay |
|---|---|---|---|---|---|
| LLaVA-v1.5-13B | 128 | 2e-5 | 1 | 2048 | 0 |
Our base model Vicuna v1.5, which is an instruction-tuned chatbot, will be downloaded automatically when you run our provided training scripts. No action is needed.
Please download the 558K subset of the LAION-CC-SBU dataset with BLIP captions we use in the paper here.
Pretrain takes around 5.5 hours for LLaVA-v1.5-13B on 8x A100 (80G), due to the increased resolution to 336px. It takes around 3.5 hours for LLaVA-v1.5-7B.
Training script with DeepSpeed ZeRO-2: pretrain.sh.
--mm_projector_type mlp2x_gelu: the two-layer MLP vision-language connector.--vision_tower openai/clip-vit-large-patch14-336: CLIP ViT-L/14 336px.
- Prepare data
Please download the annotation of the final mixture our instruction tuning data llava_v1_5_mix665k.json, and download the images from constituting datasets:
- COCO: train2017
- GQA: images
- OCR-VQA: download script, we save all files as
.jpg - TextVQA: train_val_images
- VisualGenome: part1, part2
After downloading all of them, organize the data as follows in ./playground/data,
├── coco
│ └── train2017
├── gqa
│ └── images
├── ocr_vqa
│ └── images
├── textvqa
│ └── train_images
└── vg
├── VG_100K
└── VG_100K_2
- Start training!
You may download our pretrained projectors in Model Zoo. It is not recommended to use legacy projectors, as they may be trained with a different version of the codebase, and if any option is off, the model will not function/train as we expected.
Visual instruction tuning takes around 20 hours for LLaVA-v1.5-13B on 8x A100 (80G), due to the increased resolution to 336px. It takes around 10 hours for LLaVA-v1.5-7B on 8x A100 (40G).
Training script with DeepSpeed ZeRO-3: finetune.sh.
New options to note:
--mm_projector_type mlp2x_gelu: the two-layer MLP vision-language connector.--vision_tower openai/clip-vit-large-patch14-336: CLIP ViT-L/14 336px.--image_aspect_ratio pad: this pads the non-square images to square, instead of cropping them; it slightly reduces hallucination.--group_by_modality_length True: this should only be used when your instruction tuning dataset contains both language (e.g. ShareGPT) and multimodal (e.g. LLaVA-Instruct). It makes the training sampler only sample a single modality (either image or language) during training, which we observe to speed up training by ~25%, and does not affect the final outcome.
In LLaVA-1.5, we evaluate models on a diverse set of 12 benchmarks. To ensure the reproducibility, we evaluate the models with greedy decoding. We do not evaluate using beam search to make the inference process consistent with the chat demo of real-time outputs.
See Evaluation.md.
Our GPT-assisted evaluation pipeline for multimodal modeling is provided for a comprehensive understanding of the capabilities of vision-language models. Please see our paper for more details.
- Generate LLaVA responses
python model_vqa.py \
--model-path ./checkpoints/LLaVA-13B-v0 \
--question-file \
playground/data/coco2014_val_qa_eval/qa90_questions.jsonl \
--image-folder \
/path/to/coco2014_val \
--answers-file \
/path/to/answer-file-our.jsonl- Evaluate the generated responses. In our case,
answer-file-ref.jsonlis the response generated by text-only GPT-4 (0314), with the context captions/boxes provided.
OPENAI_API_KEY="sk-***********************************" python llava/eval/eval_gpt_review_visual.py \
--question playground/data/coco2014_val_qa_eval/qa90_questions.jsonl \
--context llava/eval/table/caps_boxes_coco2014_val_80.jsonl \
--answer-list \
/path/to/answer-file-ref.jsonl \
/path/to/answer-file-our.jsonl \
--rule llava/eval/table/rule.json \
--output /path/to/review.json- Summarize the evaluation results
python summarize_gpt_review.pyIf you find LLaVA useful for your research and applications, please cite using this BibTeX:
@misc{liu2023improvedllava,
title={Improved Baselines with Visual Instruction Tuning},
author={Liu, Haotian and Li, Chunyuan and Li, Yuheng and Lee, Yong Jae},
publisher={arXiv:2310.03744},
year={2023},
}
@misc{liu2023llava,
title={Visual Instruction Tuning},
author={Liu, Haotian and Li, Chunyuan and Wu, Qingyang and Lee, Yong Jae},
publisher={arXiv:2304.08485},
year={2023},
}- Vicuna: the codebase we built upon, and our base model Vicuna-13B that has the amazing language capabilities!
- Instruction Tuning with GPT-4
- LLaVA-Med: Training a Large Language-and-Vision Assistant for Biomedicine in One Day
- Otter: In-Context Multi-Modal Instruction Tuning
For future project ideas, please check out:
- SEEM: Segment Everything Everywhere All at Once
- Grounded-Segment-Anything to detect, segment, and generate anything by marrying Grounding DINO and Segment-Anything.