base_model:

• Qwen/Qwen2.5-7B-Instruct datasets:
• THUdyh/Oryx-SFT-Data language:
• en
• zh license: apache-2.0 pipeline_tag: video-text-to-text library_name: oryx

Oryx-1.5-7B

模型概述

Oryx-1.5系列模型是基于Qwen2.5语言模型开发的7B/32B参数模型，训练数据来自Oryx-SFT-Data，支持32K tokens的上下文窗口。

Oryx提供按需解决方案，可高效处理任意空间尺寸和时长的视觉输入。

• 代码仓库: https://github.com/Oryx-mllm/Oryx
• 项目主页: https://oryx-mllm.github.io
• 支持语言: 英语、中文
• 论文地址: https://arxiv.org/abs/2409.12961

使用方式

我们提供基础调用示例，更多细节请参考Github仓库

from oryx.model.builder import load_pretrained_model
from oryx.mm_utils import get_model_name_from_path, process_images, tokenizer_image_token
from oryx.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, IGNORE_INDEX
from oryx.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

def load_video(self, 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)]
    frame_time = [i/fps for i in frame_idx]
    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]
    frame_time = ",".join([f"{i:.2f}s" for i in frame_time])
    spare_frames = vr.get_batch(frame_idx).asnumpy()
    return spare_frames,frame_time,video_time
pretrained = "THUdyh/Oryx-7B"
model_name = "oryx_qwen"
device = "cuda"
device_map = "auto"
tokenizer, model, image_processor, max_length = load_pretrained_model(pretrained, None, model_name, device_map=device_map)
model.eval()
video_path = ""
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().bfloat16()
video = [video]
video_data = (video, video)
input_data = (video_data, (384, 384), "video")
conv_template = "qwen_1_5"
question = DEFAULT_IMAGE_TOKEN + "\n请详细描述这个视频内容。"
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)
output_ids = model.generate(
    inputs=input_ids,
    images=input_data[0][0],
    images_highres=input_data[0][1],
    modalities=video_data[2],
    do_sample=False,
    temperature=0,
    max_new_tokens=128,
    use_cache=True,
)

text_outputs = tokenizer.batch_decode(cont, skip_special_tokens=True)
print(text_outputs)

性能表现

通用视频基准测试

长视频理解

通用图像基准

3D空间理解

模型架构

• 架构: 预训练Oryx-ViT + Qwen2.5-7B
• 数据: 120万图像/视频混合数据
• 精度: BFloat16

硬件与软件

• 硬件: 64张NVIDIA Tesla A100
• 训练框架: HuggingFace Trainer
• 代码库: Pytorch

引用文献

@article{liu2024oryx,
title={Oryx MLLM: On-Demand Spatial-Temporal Understanding at Arbitrary Resolution},
author={Liu, Zuyan and Dong, Yuhao and Liu, Ziwei and Hu, Winston and Lu, Jiwen and Rao, Yongming},
journal={arXiv preprint arXiv:2409.12961},
year={2024}
}