Internvl 2 5 HiCo R16

🚀 📕InternVL2.5_HiCo_R16⚡

InternVL2.5_HiCo_R16 是基於 InternVideo2.5 構建的視頻多模態大語言模型（MLLM），增強了長而豐富上下文（LRC）建模能力。它通過直接偏好優化（TPO）進行密集視覺任務註釋，以及通過自適應分層令牌壓縮（HiCo）實現緊湊的時空表示，顯著提升了現有 MLLMs 感知細粒度細節和捕捉長形式時間結構的能力。該模型是 InternVideo2.5 消融實驗的一個變體，僅基於 HiCo 技術構建（R16 表示每幀 16 個令牌）。

🚀 快速開始

安裝依賴

首先，你需要安裝 flash attention2 和其他一些模塊。以下是一個簡單的安裝示例：

pip install transformers==4.40.1
pip install av
pip install imageio
pip install decord
pip install opencv-python
pip install flash-attn --no-build-isolation

使用模型

import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer


# model setting
model_path = 'OpenGVLab/InternVL_2_5_HiCo_R16'

tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModel.from_pretrained(model_path, trust_remote_code=True).half().cuda()

IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)

def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([T.Lambda(lambda img: img.convert("RGB") if img.mode != "RGB" else img), T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC), T.ToTensor(), T.Normalize(mean=MEAN, std=STD)])
    return transform


def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float("inf")
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio


def dynamic_preprocess(image, min_num=1, max_num=6, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # calculate the existing image aspect ratio
    target_ratios = set((i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    # find the closest aspect ratio to the target
    target_aspect_ratio = find_closest_aspect_ratio(aspect_ratio, target_ratios, orig_width, orig_height, image_size)

    # calculate the target width and height
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    # resize the image
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = ((i % (target_width // image_size)) * image_size, (i // (target_width // image_size)) * image_size, ((i % (target_width // image_size)) + 1) * image_size, ((i // (target_width // image_size)) + 1) * image_size)
        # split the image
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images


def load_image(image, input_size=448, max_num=6):
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values


def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
    if bound:
        start, end = bound[0], bound[1]
    else:
        start, end = -100000, 100000
    start_idx = max(first_idx, round(start * fps))
    end_idx = min(round(end * fps), max_frame)
    seg_size = float(end_idx - start_idx) / num_segments
    frame_indices = np.array([int(start_idx + (seg_size / 2) + np.round(seg_size * idx)) for idx in range(num_segments)])
    return frame_indices

def get_num_frames_by_duration(duration):
        local_num_frames = 4        
        num_segments = int(duration // local_num_frames)
        if num_segments == 0:
            num_frames = local_num_frames
        else:
            num_frames = local_num_frames * num_segments
        
        num_frames = min(512, num_frames)
        num_frames = max(128, num_frames)

        return num_frames

def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32, get_frame_by_duration = False):
    vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
    max_frame = len(vr) - 1
    fps = float(vr.get_avg_fps())

    pixel_values_list, num_patches_list = [], []
    transform = build_transform(input_size=input_size)
    if get_frame_by_duration:
        duration = max_frame / fps
        num_segments = get_num_frames_by_duration(duration)
    frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
    for frame_index in frame_indices:
        img = Image.fromarray(vr[frame_index].asnumpy()).convert("RGB")
        img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
        pixel_values = [transform(tile) for tile in img]
        pixel_values = torch.stack(pixel_values)
        num_patches_list.append(pixel_values.shape[0])
        pixel_values_list.append(pixel_values)
    pixel_values = torch.cat(pixel_values_list)
    return pixel_values, num_patches_list

# evaluation setting
max_num_frames = 512
generation_config = dict(
    do_sample=False,
    temperature=0.0,
    max_new_tokens=1024,
    top_p=0.1,
    num_beams=1
)
video_path = "your_video.mp4"
num_segments=128


with torch.no_grad():
  
  pixel_values, num_patches_list = load_video(video_path, num_segments=num_segments, max_num=1, get_frame_by_duration=False)
  pixel_values = pixel_values.to(torch.bfloat16).to(model.device)
  video_prefix = "".join([f"Frame{i+1}: <image>\n" for i in range(len(num_patches_list))])
  # single-turn conversation
  question1 = "Describe this video in detail."
  question = video_prefix + question1
  output1, chat_history = model.chat(tokenizer, pixel_values, question, generation_config, num_patches_list=num_patches_list, history=None, return_history=True)
  print(output1)
  
  # multi-turn conversation
  question2 = "How many people appear in the video?"
  output2, chat_history = model.chat(tokenizer, pixel_values, question, generation_config, num_patches_list=num_patches_list, history=chat_history, return_history=True)
  
  print(output2)

✨ 主要特性

• 長上下文建模：增強了長而豐富上下文（LRC）建模能力，能夠更好地處理長視頻內容。
• 細粒度感知：通過直接偏好優化（TPO）進行密集視覺任務註釋，提升了感知細粒度細節的能力。
• 緊湊時空表示：利用自適應分層令牌壓縮（HiCo）實現緊湊的時空表示，有助於捕捉長形式時間結構。

📈 性能表現

✏️ 引用信息

@article{wang2025internvideo,
  title={InternVideo2.5: Empowering Video MLLMs with Long and Rich Context Modeling},
  author={Wang, Yi and Li, Xinhao and Yan, Ziang and He, Yinan and Yu, Jiashuo and Zeng, Xiangyu and Wang, Chenting and Ma, Changlian and Huang, Haian and Gao, Jianfei and Dou, Min and Chen, Kai and Wang, Wenhai and Qiao, Yu and Wang, Yali and Wang, Limin},
  journal={arXiv preprint arXiv:2501.12386},
  year={2025}
}


@article{li2024videochatflash,
  title={VideoChat-Flash: Hierarchical Compression for Long-Context Video Modeling},
  author={Li, Xinhao and Wang, Yi and Yu, Jiashuo and Zeng, Xiangyu and Zhu, Yuhan and Huang, Haian and Gao, Jianfei and Li, Kunchang and He, Yinan and Wang, Chenting and others},
  journal={arXiv preprint arXiv:2501.00574},
  year={2024}
}

📄 許可證

本項目採用 Apache-2.0 許可證。