Visual instruction tuning towards building large language and vision models with GPT-4 level capabilities in the biomedicine space.

[Paper, NeurIPS 2023 Datasets and Benchmarks Track (Spotlight)]

LLaVA-Med: Training a Large Language-and-Vision Assistant for Biomedicine in One Day

Chunyuan Li*, Cliff Wong*, Sheng Zhang*, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, Jianfeng Gao (*Equal Contribution)

Generated by GLIGEN using the grounded inpainting mode, with three boxes: white doctor coat, stethoscope, white doctor hat with a red cross sign.

• [Nov 8] LLaVA-Med is open-sourced under the MSR release policy. Huge thanks to commitment of the team, and patience of the community.
• [Sept] LLaVA-Med is accepted in NeurIPS 2023 Datasets and Benchmarks Track, as a spotlight presentation.
• [June 1] 🔥 We released LLaVA-Med: Large Language and Vision Assistant for Biomedicine, a step towards building biomedical domain large language and vision models with GPT-4 level capabilities. Checkout the paper

LLaVA-Med was initialized with the general-domain LLaVA and then continuously trained in a curriculum learning fashion (first biomedical concept alignment then full-blown instruction-tuning). We evaluated LLaVA-Med on standard visual conversation and question answering tasks.

Usage and License Notices: The data, code, and model checkpoints are intended and licensed for research use only. They are also subject to additional restrictions dictated by the Terms of Use: LLaMA, Vicuna and GPT-4 respectively. The data is made available under CC BY NC 4.0. The data, code, and model checkpoints may be used for non-commercial purposes and any models trained using the dataset should be used only for research purposes. It is expressly prohibited for models trained on this data to be used in clinical care or for any clinical decision making purposes.

• Data Download
• Install
• Training
• Model Download
• Serving
• Evaluation
• Model Description

The data statistics of biomedical multimodal instruction-following data: (a,b) The root verb-noun pairs of instruction and responses, where the inner circle of the plot represents the root verb of the output response, and the outer circle represents the direct nouns. (c) The distribution of images and QA pairs on the five domains, one image is shown per domain.

download_images.py is used to download the PMC articles using the above image_urls file and extract the images

To download our langauge-image multimodal instruction-folllowing dataset, please run the following script:

sh download_data.sh

We provide our prompts and few-shot samples for GPT-4 queries, to better facilitate research in this domain. Please check out the llava/instruct/ folder for the instruct data generation and filtering.

To generate medical instruction tuning for 60k samples and with in-text mentions:

Fill in your OpenAI API parameters in the file llava/openai_api.py:

openai.api_type = "azure"
openai.api_key = '...'
openai.api_base = 'https://example-endpoint.openai.azure.com/'
openai.api_version = "2023-03-15-preview"
DEPLOYMENT_ID="deployment-name"

Generate visual instruct tuning conversations using GPT-4

python llava/instruct/instruct_generate.py \
    --input_path data/instruct/llava_med_instruct_fig_captions.json \
    --output_path data/instruct/llava_med_instruct_60k_inline_mentions_gen.jsonl \
    --max-size 60000 \
    --use_inline_mentions True

Postprocessing of GPT-4 generated conversations

python llava/instruct/instruct_postprocess.py \
    --input_path data/instruct/llava_med_instruct_60k_inline_mentions_gen.jsonl \
    --output_path data/instruct/llava_med_instruct_60k_inline_mentions_post.json

The file llava_med_instruct_60k_inline_mentions.json in the download is generated the same way as llava_med_instruct_60k_inline_mentions_post.json output file above.

1. Clone this repository and navigate to LLaVA-Med folder

https://github.com/microsoft/LLaVA-Med.git
cd LLaVA-Med

2. Install Package: Create conda environment

conda create -n llava-med python=3.10 -y
conda activate llava-med
pip install --upgrade pip  # enable PEP 660 support

3. Install additional packages for training cases

pip uninstall torch torchvision -y
pip install torch==2.0.0+cu117 torchvision==0.15.1+cu117 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu117
pip install openai==0.27.8
pip uninstall transformers -y
pip install git+https://github.com/huggingface/transformers@cae78c46
pip install -e .

pip install einops ninja open-clip-torch
pip install flash-attn --no-build-isolation

To ensure the smooth adaptation in terms of the multimodal chat capability, we initialize model weights from the general-domain LLaVA. The delta weights of LLaVA comply with the LLaMA model license. You can add the delta to the original LLaMA weights to obtain the LLaVA weights.

1. Get the original LLaMA weights in the huggingface format by following the instructions here.
2. Use the following scripts to get LLaVA weights ``LLaVA-7b-v0'' by applying our delta LLaVA-7b-delta-v0). It will automatically download delta weights from our Hugging Face account.

This conversion command needs around 30 GB of CPU RAM.

python3 -m llava.model.apply_delta \
    --base /path/to/llama-7b \
    --target /output/path/to/LLaVA-7b-v0 \
    --delta /huggingface.co/liuhaotian/LLaVA-7b-delta-v0

LLaVA-Med is trained on 8 A100 GPUs with 40GB memory with the following code. To train on fewer GPUs, you can reduce the per_device_train_batch_size and increase the gradient_accumulation_steps accordingly to keep the global batch size the same.

torchrun --nnodes=1 --nproc_per_node=8 --master_port=25001 \
    llava/train/train_mem.py \
    --model_name_or_path ./checkpoints/llava-7b-v0 \
    --data_path /path/to/pubmed_600k.json \
    --image_folder /path/to/pubmed_600k \
    --vision_tower openai/clip-vit-large-patch14 \
    --tune_mm_mlp_adapter True \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end \
    --bf16 True \
    --output_dir ./checkpoints/llava-med-7b-pretrain \
    --num_train_epochs 1 \
    --per_device_train_batch_size 2 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 8 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 2400 \
    --save_total_limit 1 \
    --learning_rate 2e-3 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --lazy_preprocess True \
    --report_to none

You may run this with a single A100 GPU for the debugging purpose. Please note that the per_device_train_batch_size * gradient_accumulation_steps can be reduced to load model checkpoint into GPU memory. But the decreased global batch size increase the total training.

torchrun --nnodes=1 --nproc_per_node=8 --master_port=25001 \
    llava/train/train_mem.py \
    --model_name_or_path /path/to/llama-med-vicuna-7b \
    --data_path /path/to/llava_med_instruct_60k_inline_mention_post.jsonl \
    --image_folder /data/to/llava_med_instruct_images \
    --vision_tower openai/clip-vit-large-patch14 \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end True \
    --bf16 True \
    --output_dir /path/to/checkpoint_llava_med_instruct_60k_inline_mention \
    --num_train_epochs 3 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 8 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 5000 \
    --save_total_limit 3 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --fsdp "full_shard auto_wrap" \
    --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --lazy_preprocess True \
    --report_to wandb

You may directly perform medical instruction tuning on medical instruct data, by skipping Stage 1, and replacing Stage-1 checkpoint with the pretrained LLaVA checkpoint (LLaVA-7b-v0). Please see an example running script at run_training_llava_med.sh

The model weights below are delta weights. The usage of LLaVA-Med checkpoints should comply with the base LLM's model license: LLaMA.

We provide delta weights for LLaVA-Med and 3 LLaVA-Med models each finetuned on the 3 VQA datasets:

Instructions:

1. Download the delta weights above and unzip.
2. Get the original LLaMA weights in the huggingface format by following the instructions here.
3. Use the following scripts to get original LLaVA-Med weights by applying our delta. In the script below, set the --delta argument to the path of the unzipped delta weights directory from step 1.

python3 -m llava.model.apply_delta \
    --base /path/to/llama-7b \
    --target /output/path/to/llava_med_model \
    --delta /path/to/llava_med_delta_weights

python -m llava.serve.controller --host 0.0.0.0 --port 10000

python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path ./checkpoints/LLaVA-Med-7B --multi-modal

Wait until the process finishes loading the model and you see "Uvicorn running on ...".

If your the VRAM of your GPU is less than 24GB (e.g., RTX 3090, RTX 4090, etc.), you may try running it with multiple GPUs.

python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path ./checkpoints/LLaVA-Med-7B --multi-modal --num-gpus 2

Wait until the process finishes loading the model and you see "Uvicorn running on ...".

python -m llava.serve.test_message --model-name LLaVA-Med-7B --controller http://localhost:10000

python -m llava.serve.gradio_web_server --controller http://localhost:10000

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.

1. Generate LLaVA-Med responses

python model_vqa.py \
    --model-name ./checkpoints/LLaVA-7B-v0 \
    --question-file data/eval/llava_med_eval_qa50_qa.jsonl \
    --image-folder data/images/ \
    --answers-file /path/to/answer-file.jsonl

2. Evaluate the generated responses. In our case, llava_med_eval_qa50_qa.jsonl contains the questions, context (captions and inline-mentions) and responses generated by text-only GPT-4 (0314), which we treat as ground truth.

python llava/eval/eval_multimodal_chat_gpt_score.py \
    --question_input_path data/eval/llava_med_eval_qa50_qa.jsonl \
    --input_path /path/to/answer-file.jsonl \
    --output_path /path/to/save/gpt4-eval-for-individual-answers.jsonl

3. Summarize the evaluation results

python summarize_gpt_review.py

Three Medical VQA datasets are considered in our experiments, including VQA-Rad, SLAKE, Pathology-VQA. We use VQA-Rad as the running example to illustrate how LLaVA-Med is applied to a downstream scenario.

1. Please see VQA-Rad repo for setting up the dataset.
2. Generate VQA-Rad dataset for LLaVA-Med conversation-style format (the same format with instruct tuning). For each dataset, we process it into three components: train.json, test.json, images.

To achieve the higher performance for given a downstream dataset, the same full-model tuning script with instruct tuning is used to continue train LLaVA-Med.

torchrun --nnodes=1 --nproc_per_node=8 --master_port=25001 \
    llava/train/train_mem.py \
    --model_name_or_path /path/to/checkpoint_llava_med_instruct_60k_inline_mention \
    --data_path /path/to/eval/vqa_rad/train.json \
    --image_folder /path/to/eval/vqa_rad/images \
    --vision_tower openai/clip-vit-large-patch14 \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end True \
    --bf16 True \
    --output_dir /path/to/checkpoint_llava_med_instruct_60k_inline_mention/eval/fine_tuned/vqa_rad \
    --num_train_epochs 3 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 8 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 5000 \
    --save_total_limit 3 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --fsdp "full_shard auto_wrap" \
    --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --lazy_preprocess True \
    --report_to wandb

Depending on which checkpoint is employed in evaluation, zero-shot performance is reported on medical instruct tuned checkpoint (eg, LLaVA-Med-7B), and fine-tuned performance is reported on checkpoint that has been further tuned on training set of the downstream datasets (eg, LLaVA-Med-7B-VQA-Rad ).

(a) Generate LLaVA responses on ScienceQA dataset

(a.1). [Option 1] Multiple-GPU inference You may evaluate this with multiple GPUs, and concatenate the generated jsonl files. Please refer to our script for batch evaluation.

python llava/eval/run_med_datasets_eval_batch.py --num-chunks 8  --model-name /path/to/checkpoint_llava_med_instruct_60k_inline_mention/eval/fine_tuned/vqa_rad \
    --question-file path/to/eval/vqa_rad/test.json \
    --image-folder path/to/eval/vqa_rad/images \
    --answers-file /path/to/checkpoint_llava_med_instruct_60k_inline_mention/eval/fine_tuned/vqa_rad/test-answer-file.jsonl

(a.2). [Option 2] Single-GPU inference

python llava/eval/model_vqa_med.py --model-name /path/to/checkpoint_llava_med_instruct_60k_inline_mention/eval/fine_tuned/vqa_rad \
    --question-file path/to/eval/vqa_rad/test.json \
    --image-folder path/to/eval/vqa_rad/images \
    --answers-file /path/to/checkpoint_llava_med_instruct_60k_inline_mention/eval/fine_tuned/vqa_rad/test-answer-file.jsonl

(b) Evaluate the generated responses

(b.1). [Option 1] Evaluation for all three VQA datasets

python llava/eval/run_eval_batch.py \
    --pred_file_parent_path /path/to/llava-med \
    --target_test_type test-answer-file

It collects the decoding results of all predictions files under the project path, computes the corresponding evaluation metrics, and outputs the results in "eval_results_med_datasets.jsonl". To analyze the score, we provdie ipython notebook run_eval_metrics.ipynb.

(b.2). [Option 2] Evaluation for on one specific VQA dataset

python llava/eval/run_eval.py \
    --gt /path/to/eval/vqa_rad/test.json \
    --pred /path/to/checkpoint_llava_med_instruct_60k_inline_mention/eval/fine_tuned/vqa_rad/test-answer-file.jsonl

Please find the LLaVA-Med performance in llava_med_performance.md or in the paper.

Large Language and Vision Assistant for bioMedicine (i.e., “LLaVA-Med”) is a large language and vision model trained using a curriculum learning method for adapting LLaVA to the biomedical domain. It is an open-source release intended for research use only to facilitate reproducibility of the corresponding paper which claims improved performance for open-ended biomedical questions answering tasks, including common visual question answering (VQA) benchmark datasets such as PathVQA and VQA-RAD.

The data, code, and model checkpoints are intended to be used solely for (I) future research on visual-language processing and (II) reproducibility of the experimental results reported in the reference paper. The data, code, and model checkpoints are not intended to be used in clinical care or for any clinical decision making purposes.

The primary intended use is to support AI researchers reproducing and building on top of this work. LLaVA-Med and its associated models should be helpful for exploring various biomedical vision-language processing (VLP ) and vision question answering (VQA) research questions.

Any deployed use case of the model --- commercial or otherwise --- is out of scope. Although we evaluated the models using a broad set of publicly-available research benchmarks, the models and evaluations are intended for research use only and not intended for deployed use cases. Please refer to the associated paper for more details.

This model builds upon PMC-15M dataset, which is a large-scale parallel image-text dataset for biomedical vision-language processing. It contains 15 million figure-caption pairs extracted from biomedical research articles in PubMed Central. It covers a diverse range of biomedical image types, such as microscopy, radiography, histology, and more.

This model was developed using English corpora, and thus may be considered English-only. This model is evaluated on a narrow set of biomedical benchmark tasks, described in LLaVA-Med paper. As such, it is not suitable for use in any clinical setting. Under some conditions, the model may make inaccurate predictions and display limitations, which may require additional mitigation strategies. In particular, this model is likely to carry many of the limitations of the model from which it is derived, LLaVA.

Further, this model was developed in part using the PMC-15M dataset. The figure-caption pairs that make up this dataset may contain biases reflecting the current practice of academic publication. For example, the corresponding papers may be enriched for positive findings, contain examples of extreme cases, and otherwise reflect distributions that are not representative of other sources of biomedical data.

• Our project is built upon LLaVA and Vicuna: They provide our base models with the amazing multimodal and langauge capabilities, respectively!

If you find LLaVA-Med useful for your your research and applications, please cite using this BibTeX:

@article{li2023llavamed,
  title={Llava-med: Training a large language-and-vision assistant for biomedicine in one day},
  author={Li, Chunyuan and Wong, Cliff and Zhang, Sheng and Usuyama, Naoto and Liu, Haotian and Yang, Jianwei and Naumann, Tristan and Poon, Hoifung and Gao, Jianfeng},
  journal={arXiv preprint arXiv:2306.00890},
  year={2023}
}

• LLaVA
• BioMed CLIP
• Instruction Tuning with GPT-4