license: cc-by-nc-sa-4.0 datasets:

• lmms-lab/LLaVA-Video-178K language:
• en metrics:
• accuracy base_model:
• lmms-lab/LLaVA-Video-7B-Qwen2 pipeline_tag: video-text-to-text library_name: transformers tags:
• 动作识别
• 视频理解
• 多模态问答
• 多模态
• 视觉语言模型
• LLaVAction框架
• 多模态大语言模型 model-index:
• name: LLaVAction-7B results:
  • task: type: 多模态理解 dataset: name: EgoSchema type: 自我中心视频数据集 metrics:
    • type: 准确率 value: 59 name: accuracy verified: true
  • task: type: 多模态理解 dataset: name: MVBench type: 多模态视频基准测试 metrics:
    • type: 准确率 value: 61.1 name: accuracy verified: true
  • task: type: 多模态理解 dataset: name: NextQA type: 视频问答数据集 metrics:
    • type: 准确率 value: 82.8 name: accuracy verified: true
  • task: type: 多模态理解 dataset: name: PercepTest type: 感知测试集 metrics:
    • type: 准确率 value: 70.2 name: accuracy verified: true
  • task: type: 多模态理解 dataset: name: LongVideoBench type: 长视频基准测试 metrics:
    • type: 准确率 value: 58.6 name: accuracy verified: true
  • task: type: 多模态理解 dataset: name: VideoMME type: 视频多模态评估 metrics:
    • type: 准确率 value: 63.9 name: accuracy verified: true
  • task: type: 多模态理解 dataset: name: VideoMME (带字幕) type: 视频多模态评估 metrics:
    • type: 准确率 value: 71.4 name: accuracy verified: true

LLaVAction-7B

模型概述

LLaVAction-7B模型基于Qwen2语言模型架构，上下文窗口达32K tokens，在EPIC-KITCHENS-100-MQA数据集上训练完成。该模型最多支持处理64帧视频输入。

• 项目主页: https://mmathislab.github.io/llavaction/
• 论文: 详见我们的论文
• 代码库: https://github.com/AdaptiveMotorControlLab/LLaVAction
• 联系人: 麦肯齐·马西斯
• 支持语言: 英语

使用说明

适用场景

本模型在EPIC-KITCHENS-100-MQA[待发布数据集]和LLaVA-Video-178K数据集上训练，显著提升了从第一人称视角视频理解人类动作的能力。

生成示例

我们提供基础调用示例，更多细节请参考GitHub仓库。

!pip install llavaction

from llavaction.model.builder import load_pretrained_model
from llavaction.mm_utils import get_model_name_from_path, process_images, tokenizer_image_token
from llavaction.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, IGNORE_INDEX
from llavaction.conversation import conv_templates, SeparatorStyle
from PIL import Image
import requests
import copy
import torch
import sys
import warnings
from decord import VideoReader, cpu
import numpy as np
warnings.filterwarnings("ignore")

#输入视频路径（需第一人称视角）
video_path = "XXXX"

#推荐提示词模板（支持自定义）：
perspective_prompt = "当前视频为第一人称视角，你即是画面中的主角。你的双手正在与物体交互，请描述你正在执行什么动作？"
task_prompt = "请详细描述视频画面内容。"

def load_video(video_path, max_frames_num,fps=1,force_sample=False):
    if max_frames_num == 0:
        return np.zeros((1, 336, 336, 3))
    vr = VideoReader(video_path, ctx=cpu(0),num_threads=1)
    total_frame_num = len(vr)
    video_time = total_frame_num / vr.get_avg_fps()
    fps = round(vr.get_avg_fps()/fps)
    frame_idx = [i for i in range(0, len(vr), fps)]
    if len(frame_idx) > max_frames_num or force_sample:
        sample_fps = max_frames_num
        uniform_sampled_frames = np.linspace(0, total_frame_num - 1, sample_fps, dtype=int)
        frame_idx = uniform_sampled_frames.tolist()
        frame_time = [i/vr.get_avg_fps() for i in frame_idx]
    spare_frames = vr.get_batch(frame_idx).asnumpy()
    return spare_frames,frame_time,video_time

pretrained = "MLAdaptiveIntelligence/LLaVAction-7B"
model_name = "llava_qwen"
device = "cuda"
device_map = "auto"
tokenizer, model, image_processor, max_length = load_pretrained_model(pretrained, None, model_name, torch_dtype="bfloat16", device_map=device_map)  # 可添加其他模型参数
model.eval()
max_frames_num = 64
video,frame_time,video_time = load_video(video_path, max_frames_num, 1, force_sample=True)
video = image_processor.preprocess(video, return_tensors="pt")["pixel_values"].cuda().to(torch.bfloat16)
video = [video]
conv_template = "qwen_1_5"  # 注意根据模型选择对应对话模板
time_instruction = f"视频总时长{video_time:.2f}秒，从中均匀采样{len(video[0])}帧。"
question = DEFAULT_IMAGE_TOKEN + f"\n{time_instruction}\n{perspective_prompt} {task_prompt}"

conv = copy.deepcopy(conv_templates[conv_template])
conv.append_message(conv.roles[0], question)
conv.append_message(conv.roles[1], None)
prompt_question = conv.get_prompt()
input_ids = tokenizer_image_token(prompt_question, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(device)

cont = model.generate(
    input_ids,
    images=video,
    modalities= ["video"],
    do_sample=False,
    temperature=0,
    max_new_tokens=4096,
)
text_outputs = tokenizer.batch_decode(cont, skip_special_tokens=True)[0].strip()
print(text_outputs)

训练细节

详见叶等人2025年论文：arxiv.org/abs/2503.18712

模型架构

• 基础架构: SO400M + Qwen2
• 初始化模型: lmms-lab/LLaVA-Video-7B-Qwen2
• 训练数据: LLaVA-178K与EPIC-KITCHENS-100-MQA混合数据集，2个epoch，全参数训练
• 计算精度: bfloat16

硬件环境

GPU配置: 32块Nvidia GH-200（完整模型训练） 训练框架: HuggingFace Trainer 神经网络库: PyTorch

引用文献

arXiv: arxiv.org/abs/2503.18712

@article{YeQi2025llavaction,
  title={LLaVAction：面向动作识别的多模态大语言模型评估与训练},
  author={叶绍凯 and 齐浩哲 and 亚历山大·马西斯 and 麦肯齐·W·马西斯},
  journal={arXiv预印本},
  year={2025}
}