yykdev / KoLLaVA

KoLLaVA: Korean Large Language-and-Vision Assistant (feat.LLaVA)

🏔️ KoLLaVA

[Dataset] [Model] [Paper Review]

• Korean Large Language and Vision Assistant (feat. LLaVA)
• 이미지 기반 한국어 대화 가능한 멀티모달 모델

KoLLaVA Logo

Update Logs

• 24.01.02
  • 💥 🤗 KoLLaVA-Instruct-581k 공개 : LLaVA-v1.5 instruction 데이터 필터링 및 번역, 외부 지식 기반 VQA 데이터(AI Hub) 활용해 multi-turn instruction 데이터 추가
• 23.11.30
  • 💥 🤗 KoLLaVA-v1.5-Synatra-7b 공개 : 🤗 Synatra-7b-v0.3-dpo를 KoLLaVA-Instruct-581k으로 Full-finetuning (A100 80G 8개, 약 13시간)
• 23.08.05
  • 💥 🤗 KoLLaVA-LLaMA-v2-7b-qlora-4bit 공개 : 🤗 Llama-2-ko-7b-Chat을 KoLLaVA-Instruct-150k으로 QLoRA 1epoch 학습 (RTX 3090 GPU 4개, 약 10시간)
  • 💥 LLaVA의 최근 업데이트를 반영. LLaMA-2, QLoRA 기반의 코드 및 학습 방법 공유
• 23.07.01

  • 💥 🤗 KoLLaVA-KULLM-13B-8bit 공개 : KULLM을 KoLLaVA-Instruct-150k으로 학습

    → 성능이 기대에 미치지 못해 삭제합니다. 더 나은 13B 모델을 추후에 공개하겠습니다.

  • 💥 클라우드 GPU 대여 비용 문제로 데모를 일시 중지합니다🥲

• 23.06.24
  • 💥 🤗 Ko-Otter-9B-LACR-v0 공개 : Otter를 KoLLaVA_Complex_Resoning_77k 데이터셋으로 학습
• 23.06.18
  • 💥 Gradio를 이용한 데모를 오픈합니다! (RTX 3090 GPU 1개)
• 23.06.12
  • 💥 한국어 Visual Instruction 데이터셋으로 학습한 🤗KoLLaVA-KoVicuna-7B 공개
  • 💥 Colab(Pro) 이용한 inference 예시
• 23.06.09
  • 💥 LLaVA의 학습 데이터셋을 한국어로 번역한 🤗KoLLaVA-Instruct-150K / KoLLaVA-CC3M-Pretrain-595K 공개

Visual Chat Example

Contents

• Install
• Inference
• Training

Install

아래 과정은 Linux 기준으로 작성되었습니다. macOS에서 테스트 하실 경우 여기를 참고하세요.

1. Clone 후 해당 디렉토리로 이동

 git clone https://github.com/tabtoyou/KoLLaVA.git
 cd KoLLaVA

2. Package 설치

 conda create -n kollava python=3.10 -y
 conda activate kollava
 pip install --upgrade pip 
 pip install -e .

3. 학습 진행할 경우 추가 Package 설치

pip install -e ".[train]"
pip install flash-attn --no-build-isolation

Inference

터미널 창에서 아래 명령어를 통해 multi-turn 대화가 가능합니다. 또한 M1/M2 칩이 탑재된 Apple 디바이스를 사용하는 경우 --device flag를 사용하여 mps 디바이스를 지정할 수 있습니다. (--device mps) macOS에서 테스트 하실 경우 여기를 참고하세요.

python -m llava.serve.cli \
    --model-path tabtoyou/KoLLaVA-v1.5-Synatra-7b \
    --image-file "https://llava-vl.github.io/static/images/view.jpg" \

Training

LLaVA/KoLLaVA 학습은 two stage로 진행됩니다: (1) Pretraining(feature alignment stage): CC3M 데이터셋을 필터링한 595K subset을 이용하여, frozen pretrained vision encoder와 frozen LLM을 연결하는 projection layer를 학습합니다.; (2) Finetuning(visual instruction tuning stage): 150K 멀티모달 instruction-following 데이터와 약 academic-oriented tasks 및 AI-Hub에서 얻은 460K VQA 데이터를 이용해 multimodal instruction을 학습합니다.

KoLLaVA-v1.5는 8 A100 GPUs (80GB)로 학습했으며, 더 적은 GPU로 학습할 경우 per_device_train_batch_size를 줄이고 그 수에 맞게 gradient_accumulation_steps를 늘리면 됩니다. 재현을 위해서는 global batch size(per_device_train_batch_size x gradient_accumulation_steps x num_gpus)를 아래 Hyperparameters에 맞게 유지하세요.

Hyperparameters

1. Pretraining

Hyperparameter	Global Batch Size	Learning rate	Epochs	Max length	Weight decay
KoLLaVA-v1.5-Synatra-7B	256	1e-3	1	2048	0

2. Finetuning

Hyperparameter	Global Batch Size	Learning rate	Epochs	Max length	Weight decay
KoLLaVA-v1.5-Synatra-7B	128	2e-5	1	2048	0

Download Synatra checkpoints (automatically)

Base LLM 모델인 Synatra-7b의 weights은 주어진 training scripts를 실행하면 자동으로 다운로드 됩니다.

Pretrain (feature alignment)

Pretrain 과정에는 8 A100 GPUs (80GB) 기준 약 4시간이 소요됐습니다.

Prepare Pretraining Dataset

🤗 KoLLaVA-CC3M-Pretrain-595K : LLaVA Pretrain 데이터셋의 index에 맞춰 Ko-CC3M 한국어 caption 추출

Data	English	Korean	Size
CC3M Concept-balanced 595K	chat.json	ko_chat.json	211 MB / 229 MB

Details

     사전학습 데이터셋은 CC3M을 필터링해 생성했으며, 595K개의 데이터로 이루어져 있습니다. 데이터셋 구조와 영어 버전 다운로드 방법에 대한 자세한 설명은 여기를, 한국어 데이터셋은 여기를 참고하세요. (주의 : DeepL로 번역한 결과가 아니며, 품질이 조금 떨어질 수 있습니다.)

License: CC-3M 준수

Image Dataset

images.zip - LLaVA의 저자들은 사전학습에 사용한 이미지 파일도 공유했습니다. 이 이미지 파일은 연구 외에 다른 용도로 사용해서는 안 되며, 이미지의 사용은 CC3M의 라이선스를 준수해야 합니다. 원본 CC3M 데이터셋 소유자 혹은 참조된 이미지의 소유자가 요청할 경우 언제든지 해당 이미지는 삭제될 수 있습니다.

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.

Run

sh scripts/v1_5/pretrain.sh

Visual Instruction Tuning

1. Prepare data

Instruction tuning data : 🤗 KoLLaVA-Instruct-581k

• COCO: train2017
• GQA: images
• VisualGenome: part1, part2
• EKVQA: ekvqa

위의 데이터를 모두 다운받은 뒤, /workspace/data 디렉토리를 아래와 같이 구성하세요. 이때 workspace는 각자의 이미지 데이터를 저장하는 디렉토리 이름입니다.

• 주의 : COCO,GQA,VG 데이터셋은 모두 academic-oriented tasks인 영어 데이터셋이며, 이를 DeepL로 번역했습니다. 번역 과정에서 오류가 있을 수 있으며, VG의 경우 영어 단어 OCR 및 Bounding Box에 대한 정보도 포함합니다. EKVQA는 AI-Hub에서 제공하는 외부 지식 기반 멀티모달 질의응답 데이터이며 상식적인 지식이나 배경지식을 바탕으로 이미지에 관련한 질문에 대해 답을 하는 task로, Instruction-following data 형식으로 재구성했습니다. 데이터셋에 대한 저작권은 각 데이터셋의 license 규정을 따릅니다.

├── coco
│   └── train2017
├── gqa
│   └── images
├── vg
│   ├── VG_100K
│   └── VG_100K_2
└── ekvqa

KoLLaVA-v1 Visual Instruction Dataset

Visual Instruction Dataset

🤗 KoLLaVA-Instruct-150K : LLaVA의 instruction-following 데이터셋을 DeepL로 번역

English	Korean
llava_instruct_150k.json	ko_llava_instruct_150k.json
conversation_58k.json	ko_conversation_58k.json
detail_23k.json	ko_detail_23k.json
complex_reasoning_77k.json	ko_complex_reasoning_77k.json

Details

     Visual instruction tuning에 사용하는 instruction-following 데이터는 GPT-4로 생성된 데이터입니다. 이때 GPT-4의 인풋으로는 텍스트만 넣어줍니다(이미지 X). 구체적으로는 image-text pair 데이터셋인 COCO의 텍스트 정보(caption, bounding box)만을 이용해 instruction-following 형식의 데이터를 생성한 것입니다. 이러한 데이터 생성 파이프라인이 궁금하신 분은 블로그를 참고해주세요.

License: Attribution-NonCommercial 4.0 International | OpenAI policy 준수

Image Dataset

Finetuning에 사용되는 이미지 데이터셋은 COCO-train2014 입니다.

wget http://images.cocodataset.org/zips/train2014.zip

2. Start training!

Pretrain을 통해 projection layer를 생성하거나, 저희가 미리 pretrain한 KoLLaVA-v1.5-mlp2x-336px-pretrain-Synatra-7b를 다운로드 받으세요.

Visual instruction tuning은 8x A100 (80G)에서 7B 기준 대략 13시간 학습했습니다.

Training script with DeepSpeed ZeRO-3: finetune.sh.

Run

sh scripts/v1_5/finetune.sh

GPU 메모리가 충분하지 않을 경우:

• LoRA: finetune_lora.sh. global batch size(per_device_train_batch_size x gradient_accumulation_steps x num_gpus) 위에서 주어진 scripts와 동일하게 유지하세요.

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.

KoLLaVA-v1 Pretrain

     클라우드 GPU 대여 서비스인 vast.ai를 이용해 학습을 진행했습니다. KoLLaVA-KoVicuna-7b 모델 학습 시 4개의 A100(80GB) GPU를 대여했으며 Disk Space는 200GB 이상을 추천드립니다(시간 당 약 $7.44). 인스턴스 생성 시 Docker image로 pytorch/pytorch:2.0.1-cuda11.7-cudnn8-devel 를 사용했습니다.

+추가 : KoLLaVA-LLaMA-v2-7b-qlora 모델 학습에는 4개의 RTX 3090(24G) GPU를 사용했습니다.

To-do

• Finetuning 데이터셋 한국어 번역 (LLaVA-Instruct-150K)
• Pretraining 데이터셋 한국어 번역 (LLaVA-CC3M-Pretrain-595K)
• LLaVA 모델에서 Vicuna -> KoVicuna-7B 대체 후 학습
• Ko-Otter 모델 학습 및 허깅페이스 공개
• KoLLaVA-13B 모델 학습 및 허깅페이스 공개
• QLoRA 이용해 low GPU memory에서도 학습할 수 있도록 (RTX 3090 등)
• LLaVA-v1.5 코드 반영 및 모델 공개
• KoLLaVA의 linear layer를 Q-former로 업데이트(InstructBLIP)

Team

KoLLaVA-v1 프로젝트는 딥러닝 스터디 구성원들과 함께 진행했습니다.

팀원 : Jeonghyeon, Seongyeon, Seonghwan, Seungwoo, Seonghun, Taebaek

KoLLaVA-v1.5 프로젝트는 복지이십사의 지원을 받아 진행되었습니다.

---

🌋 LLaVA: Large Language and Vision Assistant

Visual instruction tuning towards large language and vision models with GPT-4 level capabilities.

[Project Page] [Paper] [Demo] [Data] [Model]

Visual Instruction Tuning
Haotian Liu*, Chunyuan Li*, Qingyang Wu, Yong Jae Lee (*Equal Contribution)

Usage and License Notices: The data, code and checkpoint is intended and licensed for research use only. They are also restricted to uses that follow the license agreement of LLaMA, Vicuna and GPT-4. The dataset is CC BY NC 4.0 (allowing only non-commercial use) and models trained using the dataset should not be used outside of research purposes.

Contents

• Data Download
• Install
• LLaVA Weights
• Serving
• Evaluation
• Fine-tuning

Data Download

Data file name	Size
llava_instruct_150k.json	229 MB
llava_instruct_80k.json	229 MB
conversation_58k.json	126 MB
detail_23k.json	20.5 MB
complex_reasoning_77k.json	79.6 MB

To download our langauge-image multimodal instruction-folllowing dataset LLaVA-Instruct-150K, please run the following script:

sh download_data.sh

Pretraining Dataset

The pretraining dataset used in this release is a subset of CC-3M dataset, filtered with a more balanced concept coverage distribution. Please see here for a detailed description on the dataset structure and how to download the images.

If you already have CC-3M dataset on your disk, the image names follow this format: GCC_train_000000000.jpg. You may edit the image field correspondingly if necessary.

Data	Chat File	Meta Data	Size
CC-3M Concept-balanced 595K	chat.json	metadata.json	211 MB
LAION/CC/SBU BLIP-Caption Concept-balanced 558K	blip_laion_cc_sbu_558k.json	metadata.json	181 MB

GPT-4 Prompts

We provide our prompts and few-shot samples for GPT-4 queries, to better facilitate research in this domain. Please check out the prompts folder for three kinds of questions: conversation, detail description, and complex reasoning.

They are organized in a format of system_message.txt for system message, pairs of abc_caps.txt for few-shot sample user input, and abc_conv.txt for few-shot sample reference output.

Note that you may find them in different format. For example, conversation is in jsonl, and detail description is answer-only. The selected format in our preliminary experiments work slightly better than a limited set of alternatives that we tried: jsonl, more natural format, answer-only. If interested, you may try other variants or conduct more careful study in this. Contributions are welcomed!

Install

1. Clone this repository and navigate to LLaVA folder

git clone https://github.com/haotian-liu/LLaVA.git
cd LLaVA

2. Install Package

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

NOTE: [Update 4/30/23] We have successfully moved LLaVA framework to this repo, without the need of a special transformers modified by us. If you install our repo before 4/30/23, please reinstall transformers following the instructions here.

3. Install additional packages for training cases

pip install ninja
pip install flash-attn==1.0.2

Upgrade to v0.1

NOTE: If you install our package before 4/30/23, please make sure to execute the command below to correctly upgrade to v0.1. You may try a clean install as well.

git pull
pip uninstall transformers
pip install git+https://github.com/huggingface/transformers@cae78c46
pip install -e .

LLaVA Weights

We release LLaVA weights as delta weights to comply with the LLaMA model license. You can add our delta to the original LLaMA weights to obtain the LLaVA weights.

Instructions:

1. Get the original LLaMA weights in the huggingface format by following the instructions here.
2. Use the following scripts to get LLaVA weights by applying our delta (13b-v0, 7b-v0, lightning-7B-v1-1). It will automatically download delta weights from our Hugging Face account.

LLaVA-13B

This conversion command needs around 60 GB of CPU RAM.

python3 -m llava.model.apply_delta \
    --base /path/to/llama-13b \
    --target /output/path/to/LLaVA-13B-v0 \
    --delta liuhaotian/LLaVA-13b-delta-v0

LLaVA-7B

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 liuhaotian/LLaVA-7b-delta-v0

LLaVA pretrained projector weights

The initial release is pretrained on LLaVA-filtered CC3M 595K with 1 epoch. The pretrained weights are released here.

You may perform instruction tuning on our pretrained checkpoints, by using our visual instruction tuning data following the instructions here.

Serving

Web UI

Launch a controller

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

Launch a model worker

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-13B-v0 --multi-modal

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

Launch a model worker (Multiple GPUs, when GPU VRAM <= 24GB)

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-13B-v0 --multi-modal --num-gpus 2

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

Launch a gradio web server.

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

You can open your browser and chat with a model now.

CLI Inference

A starting script for inference with LLaVA without the need of Gradio interface. The current implementation only supports for a single-turn Q-A session, and the interactive CLI is WIP. This also serves as an example for users to build customized inference scripts.

python -m llava.eval.run_llava \
    --model-name /path/to/LLaVA-13B-v0 \
    --image-file "https://llava-vl.github.io/static/images/view.jpg" \
    --query "What are the things I should be cautious about when I visit here?"

Example output (varies in different runs):

When visiting this picturesque location with a serene lake and a wooden pier extending over the water, one should be cautious about various safety aspects. Some important considerations include:

1. Ensuring that the pier is structurally sound andstable, as old or weakened pier structures might not support the weight of visitors.
2. Being aware of the water depth around the pier and lake, as sudden drop-offs or strong currents may pose a risk to swimmers, boaters, or those who venture too close to the edge.
3. Staying vigilant about the presence of wildlife in the area, such as slippery, stealthy fish or other animals that might cause harm or inconvenience.
4. Maintaining a safe distance from the water's edge, particularly for children, elderly individuals, or those who are not strong swimmers.
5. Following any posted signs or guidelines related to safety and the use of the pier and surrounding areas.

By considering these safety precautions, visitors can enjoy the natural beauty of the location while minimizing risks and ensuring a safe and pleasant experience.

Evaluation

GPT-assisted Evaluation

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 responses

python model_vqa.py \
    --model-name ./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.jsonl

2. Evaluate the generated responses. In our case, answer-file-1.jsonl is the response generated by text-only GPT-4 (0314), with the context captions/boxes provided.

OPENAI_API_KEY="sk-***********************************" python eval_gpt_review_visual.py \
    --question playground/data/coco2014_val_qa_eval/qa90_questions.jsonl \
    --context table/caps_boxes_coco2014_val_80.jsonl \
    --answer-list \
    /path/to/answer-file-1.jsonl \
    /path/to/answer-file-2.jsonl \
    --rule table/rule.json \
    --output /path/to/review.json

3. Summarize the evaluation results

python summarize_gpt_review.py

ScienceQA

Prepare Data

1. Please see ScienceQA repo for setting up the dataset.
2. Generate ScienceQA dataset for LLaVA conversation-style format.

python scripts/convert_sqa_to_llava \
    convert_to_llava \
    --base-dir /path/to/ScienceQA/data/scienceqa \
    --split {train,val,minival,test,minitest}

Evaluation

1. Download our pretrained LLaVA-13B (delta) weights for ScienceQA dataset here. Convert the delta weights to actual weights.

python -m llava.model.apply_delta \
    --base /path/to/llama-13b \
    --target /path/to/LLaVA-13b-v0-science_qa \
    --delta liuhaotian/LLaVA-13b-delta-v0-science_qa

2. [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 and results gathering.

3. [Option 2] Single-GPU inference

(a) Generate LLaVA responses on ScienceQA dataset

python -m llava.eval.model_vqa_science \
    --model-name /path/to/LLaVA-13b-v0-science_qa \
    --question-file /path/to/ScienceQA/data/scienceqa/llava_test.json \
    --image-folder /path/to/ScienceQA/data/scienceqa/images/test \
    --answers-file vqa/results/ScienceQA/test_llava-13b.jsonl \
    --answer-prompter
    --conv-mode simple

(b) Evaluate the generated responses

python eval_science_qa.py \
    --base-dir /path/to/ScienceQA/data/scienceqa \
    --result-file vqa/results/ScienceQA/test_llava-13b.jsonl \
    --output-file vqa/results/ScienceQA/test_llava-13b_output.json \
    --output-result vqa/results/ScienceQA/test_llava-13b_result.json \

For reference, we attach our prediction file test_llava-13b_result.json here for comparison when reproducing our results, as well as for further analysis in detail.

Fine-tuning

Data

The current version of LLaVA is fine-tuned from a Vicuna-13B model. We use approximately 600K filtered CC3M in feature alignment pretraining and 150K GPT-generated multimodal instruction-following data in finetuning. For detailed description of the data generation pipeline, please refer see our paper.

We are working on a more capable model that is pretrained with the data at a larger scale. Stay tuned!

We release all three types of multimodal instruction-following data. The use of these data is subject to OpenAI TOS.

Code and Hyperparameters

We fine-tune the model using the code from FastChat. We use a similar set of hyperparameters as Vicuna in finetuning. Both hyperparameters used in pretraining and finetuning are provided below.

1. Pretraining

Hyperparameter	Global Batch Size	Learning rate	Epochs	Max length	Weight decay
LLaVA-13B	128	2e-3	1	2048	0

2. Finetuning

Hyperparameter	Global Batch Size	Learning rate	Epochs	Max length	Weight decay
LLaVA-13B	32	2e-5	3	2048	0

Fine-tuning with Local GPUs

LLaVA is trained on 8 A100 GPUs with 80GB 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.

1. Pretraining

Pretrain: LLaVA-13B, 8x A100 (80G). Time: ~4 hours.

torchrun --nnodes=1 --nproc_per_node=8 --master_port=25001 \
    llava/train/train_mem.py \
    --model_name_or_path ./checkpoints/llama-vicuna-13b \
    --data_path /path/to/cc3m_595k.json \
    --image_folder /path/to/cc3m_595k \
    --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-13b-pretrain \
    --num_train_epochs 1 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --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 wandb

You may run this with a single A100 GPU with the following code. Please note that the per_device_train_batch_size * gradient_accumulation_steps should be equal to 128 to keep the global batch size the same.

Pretrain: LLaVA-13B, 1x A100 (80G). Time: ~33 hours.

python llava/train/train_mem.py \
    --model_name_or_path ./checkpoints/llama-vicuna-13b \
    --data_path /path/to/cc3m_595k.json \
    --image_folder /path/to/cc3m_595k \
    --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-13b-pretrain \
    --num_train_epochs 1 \
    --per_device_train_batch_size 16 \
    --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 wandb

Pretrain: LLaVA-7B, 1x A100 (80G/40G). Time: ~19 hours.

python llava/train/train_mem.py \
    --model_name_or_path ./checkpoints/llama-vicuna-7b \
    --data_path /path/to/cc3m_595k.json \
    --image_folder /path/to/cc3m_595k \
    --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-7b-pretrain \
    --num_train_epochs 1 \
    --per_device_train_batch_size 16 \
    --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 wandb

Experimental: use FSDP to save memory in pretraining

Learn more

Currently, PyTorch and Huggingface does not yet have stable/native support for FSDP on parameter efficient tuning (part of the parameters are frozen). However, the feature is being developed in PyTorch nightly and shall be shipped in the next release. We provide an experimental script to enable FSDP in pretraining. To use it, please create a new enviroment (to be safe), install PyTorch nightly (MUST), and LLaVA package following the instructions below.

1. Prepare environment

conda create -n llava_beta python=3.10 -y
conda activate llava_beta
pip install --upgrade pip
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu117
pip install -e .
pip install einops ninja
pip install flash-attn

2. Run pretraining with FSDP (experimental)

torchrun --nnodes=1 --nproc_per_node=8 --master_port=25001 \
    llava/train/train_mem.py \
    --model_name_or_path ./checkpoints/llama-vicuna-13b \
    --data_path /path/to/cc3m_595k.json \
    --image_folder /path/to/cc3m_595k \
    --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-13b-pretrain_fsdp \
    --num_train_epochs 1 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --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 \
    --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

2. Extract projector features

python scripts/extract_mm_projector.py \
  --model_name_or_path ./checkpoints/llava-13b-pretrain \
  --output ./checkpoints/mm_projector/llava-13b-pretrain.bin

3. Finetuning

torchrun --nnodes=1 --nproc_per_node=8 --master_port=25001 \
    llava/train/train_mem.py \
    --model_name_or_path /path/to/llama-vicuna-13b \
    --data_path /path/to/llava_instruct_150k.json \
    --image_folder /Data/haotian/coco/train2014 \
    --vision_tower openai/clip-vit-large-patch14 \
    --pretrain_mm_mlp_adapter ./checkpoints/mm_projector/llava-13b-pretrain.bin \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end True \
    --bf16 True \
    --output_dir ./checkpoints \
    --num_train_epochs 3 \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --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

Train LLaVA Lightning

LLaVA-Lightning can be trained on 8x A100 GPUs in just 3 hours, including both pretraining and finetuning. When using spot instances, it costs just ~$40. We are working on SkyPilot tutorial to make spot instance training even easier, stay tuned!

Please make sure to: (1) install or upgrade to the latest code base, and (2) pass the correct model version identifier v0/v1 to ensure the correct conversation template is loaded.

bash ./scripts/train_lightning.sh {v0,v1}

Hyperparameters

1. Pretraining

Hyperparameter	Global Batch Size	Learning rate	Epochs	Max length	Weight decay
LLaVA-Lightning-7B	128	2e-3	1	2048	0

2. Finetuning

Hyperparameter	Global Batch Size	Learning rate	Epochs	Max length	Weight decay
LLaVA-Lightning-7B	128	2e-5	1	2048	0

LLaVA-MPT-7b

Thanks to LLaVA-Lightning, we are able to train a checkpoint based on MPT-7b-Chat on 8x A100 GPUs in just 3 hours, including both pretraining and finetuning.

NOTE: This is a research preview of the LLaVA-Lightning based on MPT-7B-chat checkpoint. The usage of the model should comply with MPT-7B-chat license and agreements.

NOTE: Unlike other LLaVA models, this model should be used directly without delta weights conversion!

NOTE: You need to upgrade to our latest code base to use LLaVA-MPT-7b!

1. Usage

You do not need to download our checkpoint, it will directly load from our Hugging Face model: liuhaotian/LLaVA-Lightning-MPT-7B-preview.

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 liuhaotian/LLaVA-Lightning-MPT-7B-preview
python -m llava.serve.gradio_web_server --controller http://localhost:10000

2. Training

We use the same set of training dataset, and the hyperparameters as other Lightning checkpoints.

bash ./scripts/train_lightning_mpt.sh

Fine-tuning on ScienceQA

NOTE: Due to that ScienceQA experiments were done earlier, the current checkpoints are trained without <im_start> and <im_end> tokens. Checkpoints with these tokens will be updated later. Here we provide our training scripts for the current checkpoints.

1. Pretraining

torchrun --nnodes=1 --nproc_per_node=8 --master_port=25001 \
    llava/train/train_mem.py \
    --model_name_or_path ./checkpoints/llama-vicuna-13b \
    --data_path /path/to/cc3m_595k.json \
    --image_folder /path/to/cc3m_595k \
    --vision_tower openai/clip-vit-large-patch14 \
    --tune_mm_mlp_adapter True \
    --mm_vision_select_layer -2 \
    --bf16 True \
    --output_dir ./checkpoints/llava-13b-pretrain-no_im_start_end_token \
    --num_train_epochs 1 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --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 wandb

2. Extract projector features

python scripts/extract_mm_projector.py \
  --model_name_or_path ./checkpoints/llava-13b-pretrain-no_im_start_end_token \
  --output ./checkpoints/mm_projector/llava-13b-pretrain-no_im_start_end_token.bin

3. Finetuning

You may download our pretrained llava-13b-pretrain-no_im_start_end_token.bin here.

torchrun --nnodes=1 --nproc_per_node=8 --master_port=25001 \
    llava/train/train_mem.py \
    --model_name_or_path /path/to/llama-vicuna-13b \
    --data_path /path/to/scienceqa/llava_train_QCM-LEPA.json \
    --image_folder /path/to/scienceqa/images/train \
    --vision_tower openai/clip-vit-large-patch14 \
    --pretrain_mm_mlp_adapter ./checkpoints/mm_projector/llava-13b-pretrain-no_im_start_end_token.bin \
    --mm_vision_select_layer -2 \
    --bf16 True \
    --output_dir ./checkpoints/llava-13b-pretrain-no_im_start_end_token-finetune_scienceqa \
    --num_train_epochs 12 \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --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

Acknowledgement

• Vicuna: the codebase we built upon, and our base model Vicuna-13B that has the amazing language capabilities!

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

@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},
}

Related Projects

• Instruction Tuning with GPT-4

For future project ideas, pleae 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.