Longva 7B TPO

🚀 LongVA-7B-TPO

本仓库包含论文 Temporal Preference Optimization for Long-form Video Understanding 中描述的模型。

论文 Temporal Preference Optimization for Long-form Video Understanding 提出的 LongVA-7B-TPO 模型，基于 LongVA-7B 进行了时间偏好优化。LongVA-7B-TPO 模型在一系列基准测试中取得了最先进的性能，与 LongVA-7B 相比，平均性能提升了 2%。

📦 模型信息

📊 评估结果

🚀 快速开始

使用以下代码开始使用该模型。更多信息请参考我们的 GitHub 仓库。

基础用法

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 = "ruili0/LongVA-TPO"
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)

📄 许可证

本项目使用了一些数据集和检查点，这些数据集和检查点受其各自的原始许可证约束。用户必须遵守这些原始许可证的所有条款和条件，包括但不限于数据集的 OpenAI 使用条款和基础语言模型的特定许可证（如 Qwen2 许可证）。本项目不会在原始许可证规定的约束之外施加任何额外的限制。此外，提醒用户确保其对数据集和检查点的使用符合所有适用的法律法规。

📚 引用

BibTeX 引用

@article{li2025temporal,
      title={Temporal Preference Optimization for Long-Form Video Understanding},
      author={Li, Rui and Wang, Xiaohan and Zhang, Yuhui and Wang, Zeyu and Yeung-Levy, Serena},
      journal={arXiv preprint arXiv:2501.13919},
      year={2025}
    }

参考文献

[1]. Zhang, P., Zhang, K., Li, B., Zeng, G., Yang, J., Zhang, Y., ... & Liu, Z. (2024). Long context transfer from language to vision. arXiv preprint arXiv:2406.16852.