LongVA & V-NIAH

🌐 Blog | 📃 Paper | 🤗 Hugging Face | 🎥 Demo

Long context capability can zero-shot transfer from language to vision.

LongVA can process 2000 frames or over 200K visual tokens. It achieves state-of-the-art performance on Video-MME among 7B models.

Installation

This codebase is tested on CUDA 11.8 and A100-SXM-80G.

conda create -n longva python=3.10 -y && conda activate longva
pip install torch==2.1.2 torchvision --index-url https://download.pytorch.org/whl/cu118
pip install -e "longva/.[train]"
pip install packaging &&  pip install ninja && pip install flash-attn --no-build-isolation --no-cache-dir
pip install -r requirements.txt

Local Demo

# For CLI inference
python local_demo/longva_backend.py --video_path local_demo/assets/dc_demo.mp4 --question "What does this video show?" --num_sampled_frames 32 --device auto
python local_demo/longva_backend.py --image_path local_demo/assets/lmms-eval.png --question "What is inside the image?"

# For multimodal chat demo with gradio UI
python local_demo/multimodal_chat.py

Quick Start With HuggingFace

Example Code

from longva.model.builder import load_pretrained_model
from longva.mm_utils import tokenizer_image_token, process_images
from longva.constants import IMAGE_TOKEN_INDEX
from PIL import Image
from decord import VideoReader, cpu
import torch
import numpy as np
# fix seed
torch.manual_seed(0)

model_path = "lmms-lab/LongVA-7B-DPO"
image_path = "local_demo/assets/lmms-eval.png"
video_path = "local_demo/assets/dc_demo.mp4"
max_frames_num = 16 # you can change this to several thousands so long you GPU memory can handle it :)
gen_kwargs = {"do_sample": True, "temperature": 0.5, "top_p": None, "num_beams": 1, "use_cache": True, "max_new_tokens": 1024}
# you can also set the device map to auto to accomodate more frames
tokenizer, model, image_processor, _ = load_pretrained_model(model_path, None, "llava_qwen", device_map="cuda:0")


#image input
prompt = "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<image>\nDescribe the image in details.<|im_end|>\n<|im_start|>assistant\n"
input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(model.device)
image = Image.open(image_path).convert("RGB")
images_tensor = process_images(image, image_processor, model.config).to(model.device, dtype=torch.float16)
with torch.inference_mode():
    output_ids = model.generate(input_ids, images=images_tensor, image_sizes=[image.size], modalities=["image"], **gen_kwargs)
outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
print(outputs)
print("-"*50)

#video input
prompt = "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<image>\nGive a detailed caption of the video as if I am blind.<|im_end|>\n<|im_start|>assistant\n"
input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(model.device)
vr = VideoReader(video_path, ctx=cpu(0))
total_frame_num = len(vr)
uniform_sampled_frames = np.linspace(0, total_frame_num - 1, max_frames_num, dtype=int)
frame_idx = uniform_sampled_frames.tolist()
frames = vr.get_batch(frame_idx).asnumpy()
video_tensor = image_processor.preprocess(frames, return_tensors="pt")["pixel_values"].to(model.device, dtype=torch.float16)
with torch.inference_mode():
    output_ids = model.generate(input_ids, images=[video_tensor],  modalities=["video"], **gen_kwargs)
outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
print(outputs)

V-NIAH Evaluation

To begin, download a video longer than one hour to use as the haystack video and save it at vision_niah/data/long_video.mp4. We cannot provide the video ourselves as we use an actual movie in our evaluation.

You can view all needle questions at lmms-lab/v_niah_needles.

huggingface-cli download lmms-lab/LongVA-7B --local-dir vision_niah/model_weights/LongVA-7B
sh vision_niah/eval.sh

Results will be saved to vision_niah/niah_output. We run on V-NIAH using PPL-based evaluation. If you want to use generation-based evaluation, check out a demo at vision_niah/eval_vision_niah_sampling.py. Please refer to Section 4 of our paper for more details.

LMMs-Eval Evaluation

We provide both our video and image evaluation pipeline using lmms-eval. After installing lmms-eval and longva, you can use the following script to evaluate on both image and video tasks

Image evaluation command

accelerate launch --num_processes 8 --main_process_port 12345 -m lmms_eval \
    --model longva \
    --model_args pretrained=lmms-lab/LongVA-7B,conv_template=qwen_1_5,model_name=llava_qwen \
    --tasks mme \
    --batch_size 1 \
    --log_samples \
    --log_samples_suffix mme_longva \
    --output_path ./logs/

Video evaluation command

accelerate launch --num_processes 8 --main_process_port 12345 -m lmms_eval \
    --model longva \
    --model_args pretrained=lmms-lab/LongVA-7B,conv_template=qwen_1_5,video_decode_backend=decord,max_frames_num=32,model_name=llava_qwen \
    --tasks videomme \
    --batch_size 1 \
    --log_samples \
    --log_samples_suffix videomme_longva \
    --output_path ./logs/ 

Long Text Training

sh text_extend/extend_qwen2.sh

It takes around 2 days to train the model on 8 A100 GPUs. You can also download our long-context-pretrained model from huggingface:

huggingface-cli download lmms-lab/Qwen2-7B-Instrcuct-224K --local-dir text_extend/training_output/Qwen2-7B-Instrcuct-224K

You can evaluate the text-niah performance with this command:

sh text_extend/eval.sh

The results will be saved to text_extend/niah_output.

Vision Text Alignment

Coming soon...

Citation

If you find this work useful, please consider citing our paper:

@article{zhang2024longva,
  title={Long Context Transfer from Language to Vision},
  author={Peiyuan Zhang and Kaichen Zhang and Bo Li and Guangtao Zeng and Jingkang Yang and Yuanhan Zhang and Ziyue Wang and Haoran Tan and Chunyuan Li and Ziwei Liu},
  journal={arXiv preprint arXiv:2406.16852},
  year={2024},
  url = {https://arxiv.org/abs/2406.16852}
}

Acknowledgement

• LLaVA: the codebase we built upon.
• LMMs-Eval: the codebase we used for evaluation.