license: bsd-3-clause pipeline_tag: video-text-to-text

VideoMind-7B

VideoMind是一个多模态智能体框架，通过模拟人类思维过程（如任务分解、时刻定位与验证和答案合成）来增强视频推理能力。

🔖 模型详情

• 模型类型: 多模态大语言模型
• 支持语言: 英语
• 许可证: BSD-3-Clause

🚀 快速开始

环境安装

1. 从GitHub克隆仓库。

git clone git@github.com:yeliudev/VideoMind.git
cd VideoMind

2. 初始化conda环境。

conda create -n videomind python=3.11 -y
conda activate videomind

3. 安装依赖项。

pip install -r requirements.txt

NPU用户请修改requirements.txt文件的18-25行。

快速推理演示

以下脚本展示了如何使用VideoMind的不同角色进行推理。更多模型细节请参考我们的GitHub仓库。

import torch

from videomind.constants import GROUNDER_PROMPT, PLANNER_PROMPT, VERIFIER_PROMPT
from videomind.dataset.utils import process_vision_info
from videomind.model.builder import build_model
from videomind.utils.io import get_duration
from videomind.utils.parser import parse_span

MODEL_PATH = 'yeliudev/VideoMind-7B'

video_path = '<视频路径>'
question = '<问题>'

# 初始化*定位器*角色
model, processor = build_model(MODEL_PATH)
device = next(model.parameters()).device

# 初始化*规划器*角色
model.load_adapter(f'{MODEL_PATH}/planner', adapter_name='planner')

# 初始化*验证器*角色
model.load_adapter(f'{MODEL_PATH}/verifier', adapter_name='verifier')

# ==================== 规划器 ====================

messages = [{
    'role':
    'user',
    'content': [{
        'type': 'video',
        'video': video_path,
        'min_pixels': 36 * 28 * 28,
        'max_pixels': 64 * 28 * 28,
        'max_frames': 100,
        'fps': 1.0
    }, {
        'type': 'text',
        'text': PLANNER_PROMPT.format(question)
    }]
}]

# 预处理输入
text = processor.apply_chat_template(messages, add_generation_prompt=True)
images, videos = process_vision_info(messages)
data = processor(text=[text], images=images, videos=videos, return_tensors='pt').to(device)

# 切换到*规划器*适配器
model.base_model.disable_adapter_layers()
model.base_model.enable_adapter_layers()
model.set_adapter('planner')

# 运行推理
output_ids = model.generate(**data, do_sample=False, temperature=None, top_p=None, top_k=None, max_new_tokens=256)

# 解码输出
output_ids = output_ids[0, data.input_ids.size(1):-1]
response = processor.decode(output_ids, clean_up_tokenization_spaces=False)

print(f'规划器响应: {response}')

# ==================== 定位器 ====================

messages = [{
    'role':
    'user',
    'content': [{
        'type': 'video',
        'video': video_path,
        'min_pixels': 36 * 28 * 28,
        'max_pixels': 64 * 28 * 28,
        'max_frames': 150,
        'fps': 1.0
    }, {
        'type': 'text',
        'text': GROUNDER_PROMPT.format(question)
    }]
}]

# 预处理输入
text = processor.apply_chat_template(messages, add_generation_prompt=True)
images, videos = process_vision_info(messages)
data = processor(text=[text], images=images, videos=videos, return_tensors='pt').to(device)

# 切换到*定位器*适配器
model.base_model.disable_adapter_layers()
model.base_model.enable_adapter_layers()
model.set_adapter('grounder')

# 运行推理
output_ids = model.generate(**data, do_sample=False, temperature=None, top_p=None, top_k=None, max_new_tokens=256)

# 解码输出
output_ids = output_ids[0, data.input_ids.size(1):-1]
response = processor.decode(output_ids, clean_up_tokenization_spaces=False)

print(f'定位器响应: {response}')

duration = get_duration(video_path)

# 1. 提取时间戳和置信度
blob = model.reg[0].cpu().float()
pred, conf = blob[:, :2] * duration, blob[:, -1].tolist()

# 2. 限制时间戳范围
pred = pred.clamp(min=0, max=duration)

# 3. 排序时间戳
inds = (pred[:, 1] - pred[:, 0] < 0).nonzero()[:, 0]
pred[inds] = pred[inds].roll(1)

# 4. 转换时间戳为列表
pred = pred.tolist()

print(f'定位器预测时间戳: {pred}')

# ==================== 验证器 ====================

# 使用前5个预测
probs = []
for cand in pred[:5]:
    s0, e0 = parse_span(cand, duration, 2)
    offset = (e0 - s0) / 2
    s1, e1 = parse_span([s0 - offset, e0 + offset], duration)

    # 计算s0,e0在s1,e1中的百分比
    s = (s0 - s1) / (e1 - s1)
    e = (e0 - s1) / (e1 - s1)

    messages = [{
        'role':
        'user',
        'content': [{
            'type': 'video',
            'video': video_path,
            'video_start': s1,
            'video_end': e1,
            'min_pixels': 36 * 28 * 28,
            'max_pixels': 64 * 28 * 28,
            'max_frames': 64,
            'fps': 2.0
        }, {
            'type': 'text',
            'text': VERIFIER_PROMPT.format(question)
        }]
    }]

    text = processor.apply_chat_template(messages, add_generation_prompt=True)
    images, videos = process_vision_info(messages)
    data = processor(text=[text], images=images, videos=videos, return_tensors='pt')

    # ===== 插入片段起始/结束标记 =====
    video_grid_thw = data['video_grid_thw'][0]
    num_frames, window = int(video_grid_thw[0]), int(video_grid_thw[1] * video_grid_thw[2] / 4)
    assert num_frames * window * 4 == data['pixel_values_videos'].size(0)

    pos_s, pos_e = round(s * num_frames), round(e * num_frames)
    pos_s, pos_e = min(max(0, pos_s), num_frames), min(max(0, pos_e), num_frames)
    assert pos_s <= pos_e, (num_frames, s, e)

    base_idx = torch.nonzero(data['input_ids'][0] == model.config.vision_start_token_id).item()
    pos_s, pos_e = pos_s * window + base_idx + 1, pos_e * window + base_idx + 2

    input_ids = data['input_ids'][0].tolist()
    input_ids.insert(pos_s, model.config.seg_s_token_id)
    input_ids.insert(pos_e, model.config.seg_e_token_id)
    data['input_ids'] = torch.LongTensor([input_ids])
    data['attention_mask'] = torch.ones_like(data['input_ids'])
    # ===========================================

    data = data.to(device)

    # 切换到*验证器*适配器
    model.base_model.disable_adapter_layers()
    model.base_model.enable_adapter_layers()
    model.set_adapter('verifier')

    # 运行推理
    with torch.inference_mode():
        logits = model(**data).logits[0, -1].softmax(dim=-1)

    # 注意：此处为魔法数字
    # Qwen2-VL词表中：9454 -> 是，2753 -> 否
    score = (logits[9454] - logits[2753]).sigmoid().item()
    probs.append(score)

# 按验证器置信度排序预测
ranks = torch.Tensor(probs).argsort(descending=True).tolist()

pred = [pred[idx] for idx in ranks]
conf = [conf[idx] for idx in ranks]

print(f'验证器重排时间戳: {pred}')

# ==================== 回答器 ====================

# 选择最佳候选时刻
s, e = parse_span(pred[0], duration, 32)

messages = [{
    'role':
    'user',
    'content': [{
        'type': 'video',
        'video': video_path,
        'video_start': s,
        'video_end': e,
        'min_pixels': 128 * 28 * 28,
        'max_pixels': 256 * 28 * 28,
        'max_frames': 32,
        'fps': 2.0
    }, {
        'type': 'text',
        'text': question
    }]
}]

text = processor.apply_chat_template(messages, add_generation_prompt=True)
images, videos = process_vision_info(messages)
data = processor(text=[text], images=images, videos=videos, return_tensors='pt').to(device)

# 移除所有适配器，使用基础模型作为*回答器*
with model.disable_adapter():
    output_ids = model.generate(**data, do_sample=False, temperature=None, top_p=None, top_k=None, max_new_tokens=256)

# 解码输出
output_ids = output_ids[0, data.input_ids.size(1):-1]
response = processor.decode(output_ids, clean_up_tokenization_spaces=False)

print(f'回答器响应: {response}')

📖 引用

如果您觉得本项目有帮助，请引用我们的论文。

@article{liu2025videomind,
  title={VideoMind: A Chain-of-LoRA Agent for Long Video Reasoning},
  author={Liu, Ye and Lin, Kevin Qinghong and Chen, Chang Wen and Shou, Mike Zheng},
  journal={arXiv preprint arXiv:2503.13444},
  year={2025}
}