💡 Some other multimodal-LLM projects from our team may interest you ✨.
Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding
Hang Zhang, Xin Li, Lidong Bing
VCD: Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding
Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, Lidong Bing
demo_video.webm
- [2024.06.18] 🔥🔥 As of Jun 18, our VideoLLaMA2-7B-16F is the Top-1 ~7B-sized VideoLLM on the VideoMME Leaderboard.
- [2024.06.17] 👋👋 Update technical report with the latest results and the missing references. If you have works closely related to VideoLLaMA 2 but not mentioned in the paper, feel free to let us know.
- [2024.06.14] 🔥🔥 Online Demo is available.
- [2024.06.03] Release training, evaluation, and serving codes of VideoLLaMA 2.
Basic Dependencies:
- Python >= 3.8
- Pytorch >= 2.0.1
- CUDA Version >= 11.7
- transformers >= 4.37.2
[Online Mode] Install required packages (better for development):
git clone https://github.com/DAMO-NLP-SG/VideoLLaMA2
cd VideoLLaMA2
pip install -r requirements.txt
pip install flash-attn --no-build-isolation[Offline Mode] Install VideoLLaMA2 as a Python package (better for direct use):
git clone https://github.com/DAMO-NLP-SG/VideoLLaMA2
cd VideoLLaMA2
pip install --upgrade pip # enable PEP 660 support
pip install -e .
pip install flash-attn --no-build-isolation
| Model Name | Model Type | Visual Encoder | Language Decoder | # Training Frames |
|---|---|---|---|---|
| VideoLLaMA2-7B-Base | Base | clip-vit-large-patch14-336 | Mistral-7B-Instruct-v0.2 | 8 |
| VideoLLaMA2-7B | Chat | clip-vit-large-patch14-336 | Mistral-7B-Instruct-v0.2 | 8 |
| VideoLLaMA2-7B-16F-Base | Base | clip-vit-large-patch14-336 | Mistral-7B-Instruct-v0.2 | 16 |
| VideoLLaMA2-7B-16F | Chat | clip-vit-large-patch14-336 | Mistral-7B-Instruct-v0.2 | 16 |
It is highly recommended to try our online demo first.
To run a video-based LLM (Large Language Model) web demonstration on your device, you will first need to ensure that you have the necessary model checkpoints prepared, followed by adhering to the steps outlined to successfully launch the demo.
- Launch a gradio app directly (VideoLLaMA2-7B is adopted by default):
python videollama2/serve/gradio_web_server_adhoc.py- Launch a global controller
cd /path/to/VideoLLaMA2
python -m videollama2.serve.controller --host 0.0.0.0 --port 10000- Launch a gradio webserver
python -m videollama2.serve.gradio_web_server --controller http:https://localhost:10000 --model-list-mode reload- Launch one or multiple model workers
# export HF_ENDPOINT=https://hf-mirror.com # If you are unable to access Hugging Face, try to uncomment this line.
python -m videollama2.serve.model_worker --host 0.0.0.0 --controller http:https://localhost:10000 --port 40000 --worker http:https://localhost:40000 --model-path /PATH/TO/MODEL1
python -m videollama2.serve.model_worker --host 0.0.0.0 --controller http:https://localhost:10000 --port 40001 --worker http:https://localhost:40001 --model-path /PATH/TO/MODEL2
python -m videollama2.serve.model_worker --host 0.0.0.0 --controller http:https://localhost:10000 --port 40002 --worker http:https://localhost:40002 --model-path /PATH/TO/MODEL3
...To facilitate further development on top of our codebase, we provide a quick-start guide on how to train a customized VideoLLaMA2 with VideoLLaVA dataset and evaluate the trained model on the mainstream video-llm benchmarks.
- Training Data Structure:
VideoLLaMA2
├── datasets
│ ├── videollava_pt
| | ├── llava_image/ # Available at: https://pan.baidu.com/s/17GYcE69FcJjjUM0e4Gad2w?pwd=9ga3
| | ├── valley/ # Available at: https://pan.baidu.com/s/1jluOimE7mmihEBfnpwwCew?pwd=jyjz
| | └── valley_llavaimage.json # Available at: https://drive.google.com/file/d/1zGRyVSUMoczGq6cjQFmT0prH67bu2wXD/view, including 703K video-text and 558K image-text pairs
│ ├── videollava_sft
| | ├── llava_image_tune/ # Available at: https://pan.baidu.com/s/1l-jT6t_DlN5DTklwArsqGw?pwd=o6ko
| | ├── videochatgpt_tune/ # Available at: https://pan.baidu.com/s/10hJ_U7wVmYTUo75YHc_n8g?pwd=g1hf
| | └── videochatgpt_llavaimage_tune.json # Available at: https://drive.google.com/file/d/1zGRyVSUMoczGq6cjQFmT0prH67bu2wXD/view, including 100K video-centric, 625K image-centric and 40K text-only conversations- Command:
# VideoLLaMA2-vllava pretraining
bash scripts/vllava/stc/pretrain.sh
# VideoLLaMA2-vllava finetuning
bash scripts/vllava/stc/finetune.sh- Evaluation Data Structure:
VideoLLaMA2
├── eval
│ ├── egoschema # Official website: https://github.com/egoschema/EgoSchema
| | ├── good_clips_git/ # Available at: https://drive.google.com/drive/folders/1SS0VVz8rML1e5gWq7D7VtP1oxE2UtmhQ
| | └── questions.json # Available at: https://github.com/egoschema/EgoSchema/blob/main/questions.json
│ ├── mvbench # Official website: https://huggingface.co/datasets/OpenGVLab/MVBench
| | ├── video/
| | | ├── clever/
| | | └── ...
| | └── json/
| | | ├── action_antonym.json
| | | └── ...
│ ├── perception_test_mcqa # Official website: https://huggingface.co/datasets/OpenGVLab/MVBench
| | ├── videos/ # Available at: https://storage.googleapis.com/dm-perception-test/zip_data/test_videos.zip
| | └── mc_question_test.json # Download from https://storage.googleapis.com/dm-perception-test/zip_data/mc_question_test_annotations.zip
│ ├── videomme # Official website: https://video-mme.github.io/home_page.html#leaderboard
| | ├── test-00000-of-00001.parquet
| | ├── videos/
| | └── subtitles/
│ ├── Activitynet_Zero_Shot_QA # Official website: https://github.com/MILVLG/activitynet-qa
| | ├── all_test/ # Available at: https://mbzuaiac-my.sharepoint.com/:u:/g/personal/hanoona_bangalath_mbzuai_ac_ae/EatOpE7j68tLm2XAd0u6b8ABGGdVAwLMN6rqlDGM_DwhVA?e=90WIuW
| | ├── test_q.json # Available at: https://github.com/MILVLG/activitynet-qa/tree/master/dataset
| | └── test_a.json # Available at: https://github.com/MILVLG/activitynet-qa/tree/master/dataset
│ ├── MSVD_Zero_Shot_QA # Official website: https://github.com/xudejing/video-question-answering
| | ├── videos/
| | ├── test_q.json
| | └── test_a.json
│ ├── videochatgpt_gen # Official website: https://github.com/mbzuai-oryx/Video-ChatGPT/tree/main/quantitative_evaluation
| | ├── Test_Videos/ # Available at: https://mbzuaiac-my.sharepoint.com/:u:/g/personal/hanoona_bangalath_mbzuai_ac_ae/EatOpE7j68tLm2XAd0u6b8ABGGdVAwLMN6rqlDGM_DwhVA?e=90WIuW
| | ├── Test_Human_Annotated_Captions/ # Available at: https://mbzuaiac-my.sharepoint.com/personal/hanoona_bangalath_mbzuai_ac_ae/_layouts/15/onedrive.aspx?id=%2Fpersonal%2Fhanoona%5Fbangalath%5Fmbzuai%5Fac%5Fae%2FDocuments%2FVideo%2DChatGPT%2FData%5FCode%5FModel%5FRelease%2FQuantitative%5FEvaluation%2Fbenchamarking%2FTest%5FHuman%5FAnnotated%5FCaptions%2Ezip&parent=%2Fpersonal%2Fhanoona%5Fbangalath%5Fmbzuai%5Fac%5Fae%2FDocuments%2FVideo%2DChatGPT%2FData%5FCode%5FModel%5FRelease%2FQuantitative%5FEvaluation%2Fbenchamarking&ga=1
| | ├── generic_qa.json # These three json files available at: https://mbzuaiac-my.sharepoint.com/personal/hanoona_bangalath_mbzuai_ac_ae/_layouts/15/onedrive.aspx?id=%2Fpersonal%2Fhanoona%5Fbangalath%5Fmbzuai%5Fac%5Fae%2FDocuments%2FVideo%2DChatGPT%2FData%5FCode%5FModel%5FRelease%2FQuantitative%5FEvaluation%2Fbenchamarking%2FBenchmarking%5FQA&ga=1
| | ├── temporal_qa.json
| | └── consistency_qa.json- Command:
# mvbench evaluation
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash scripts/eval/eval_video_qa_mvbench.sh
# activitynet-qa evaluation (need to set azure openai key/endpoint/deployname)
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash scripts/eval/eval_video_qa_mvbench.shIf you want to train a video-llm on your data, you need to follow the procedures below to prepare the video/image sft data:
- Suppose your data structure is like:
VideoLLaMA2
├── datasets
│ ├── custom_sft
│ | ├── images
│ | ├── videos
| | └── custom.json- Then you should re-organize the annotated video/image sft data according to the following format:
[
{
"id": 0,
"video": "images/xxx.jpg",
"conversations": [
{
"from": "human",
"value": "<image>\nWhat are the colors of the bus in the image?"
},
{
"from": "gpt",
"value": "The bus in the image is white and red."
},
...
],
}
{
"id": 1,
"video": "videos/xxx.mp4",
"conversations": [
{
"from": "human",
"value": "<video>\nWhat are the main activities that take place in the video?"
},
{
"from": "gpt",
"value": "The main activities that take place in the video are the preparation of camera equipment by a man, a group of men riding a helicopter, and a man sailing a boat through the water."
},
...
],
},
...
]- Modify the
scripts/custom/finetune.sh:
...
--data_path datasets/custom_sft/custom.json
--data_folder datasets/custom_sft/
--pretrain_mm_mlp_adapter CONNECTOR_DOWNLOAD_PATH (e.g., DAMO-NLP-SG/VideoLLaMA2-7B-Base)
...Video/Image Inference:
import torch
import transformers
import sys
sys.path.append('./')
from videollama2.conversation import conv_templates
from videollama2.constants import DEFAULT_MMODAL_TOKEN, MMODAL_TOKEN_INDEX
from videollama2.mm_utils import get_model_name_from_path, tokenizer_MMODAL_token, process_video, process_image
from videollama2.model.builder import load_pretrained_model
def inference():
# Video Inference
paths = ['assets/cat_and_chicken.mp4']
questions = ['What animals are in the video, what are they doing, and how does the video feel?']
# Reply:
# The video features a kitten and a baby chick playing together. The kitten is seen laying on the floor while the baby chick hops around. The two animals interact playfully with each other, and the video has a cute and heartwarming feel to it.
modal_list = ['video']
# Image Inference
paths = ['assets/sora.png']
questions = ['What is the woman wearing, what is she doing, and how does the image feel?']
# Reply:
# The woman in the image is wearing a black coat and sunglasses, and she is walking down a rain-soaked city street. The image feels vibrant and lively, with the bright city lights reflecting off the wet pavement, creating a visually appealing atmosphere. The woman's presence adds a sense of style and confidence to the scene, as she navigates the bustling urban environment.
modal_list = ['image']
# 1. Initialize the model.
model_path = 'DAMO-NLP-SG/VideoLLaMA2-7B'
# Base model inference (only need to replace model_path)
# model_path = 'DAMO-NLP-SG/VideoLLaMA2-7B-Base'
model_name = get_model_name_from_path(model_path)
tokenizer, model, processor, context_len = load_pretrained_model(model_path, None, model_name)
model = model.to('cuda:0')
conv_mode = 'llama_2'
# 2. Visual preprocess (load & transform image or video).
if modal_list[0] == 'video':
tensor = process_video(paths[0], processor, model.config.image_aspect_ratio).to(dtype=torch.float16, device='cuda', non_blocking=True)
default_mm_token = DEFAULT_MMODAL_TOKEN["VIDEO"]
modal_token_index = MMODAL_TOKEN_INDEX["VIDEO"]
else:
tensor = process_image(paths[0], processor, model.config.image_aspect_ratio)[0].to(dtype=torch.float16, device='cuda', non_blocking=True)
default_mm_token = DEFAULT_MMODAL_TOKEN["IMAGE"]
modal_token_index = MMODAL_TOKEN_INDEX["IMAGE"]
tensor = [tensor]
# 3. text preprocess (tag process & generate prompt).
question = default_mm_token + "\n" + questions[0]
conv = conv_templates[conv_mode].copy()
conv.append_message(conv.roles[0], question)
conv.append_message(conv.roles[1], None)
prompt = conv.get_prompt()
input_ids = tokenizer_MMODAL_token(prompt, tokenizer, modal_token_index, return_tensors='pt').unsqueeze(0).to('cuda:0')
with torch.inference_mode():
output_ids = model.generate(
input_ids,
images_or_videos=tensor,
modal_list=modal_list,
do_sample=True,
temperature=0.2,
max_new_tokens=1024,
use_cache=True,
)
outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
print(outputs[0])
if __name__ == "__main__":
inference()If you find VideoLLaMA useful for your research and applications, please cite using this BibTeX:
@article{damonlpsg2024videollama2,
title={VideoLLaMA 2: Advancing Spatial-Temporal Modeling and Audio Understanding in Video-LLMs},
author={Cheng, Zesen and Leng, Sicong and Zhang, Hang and Xin, Yifei and Li, Xin and Chen, Guanzheng and Zhu, Yongxin and Zhang, Wenqi and Luo, Ziyang and Zhao, Deli and Bing, Lidong},
journal={arXiv preprint arXiv:2406.07476},
year={2024},
url = {https://arxiv.org/abs/2406.07476}
}
@article{damonlpsg2023videollama,
title = {Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding},
author = {Zhang, Hang and Li, Xin and Bing, Lidong},
journal = {arXiv preprint arXiv:2306.02858},
year = {2023},
url = {https://arxiv.org/abs/2306.02858}
}The codebase of VideoLLaMA 2 is adapted from LLaVA 1.5 and FastChat. We are also grateful for the following projects our VideoLLaMA 2 arise from:
- LLaMA 2, Mistral-7B, OpenAI CLIP, Honeybee.
- Video-ChatGPT, Video-LLaVA.
- WebVid, Panda-70M, LanguageBind, InternVid.
- VideoChat2, Valley, VTimeLLM, ShareGPT4V.
This project is released under the Apache 2.0 license as found in the LICENSE file. The service is a research preview intended for non-commercial use ONLY, subject to the model Licenses of LLaMA and Mistral, Terms of Use of the data generated by OpenAI, and Privacy Practices of ShareGPT. Please get in touch with us if you find any potential violations.
