zh/models/yolo12/

[giscus[bot]]

zh/models/yolo12/

了解 YOLO12，它采用开创性的以注意力为中心的架构，能够以无与伦比的准确性和效率进行最先进的物体检测。

https://docs.ultralytics.com/zh/models/yolo12/

[UltralyticsAssistant]

👋 Hello, thank you for your interest in Ultralytics 🚀! We encourage you to explore the Documentation for all things YOLO, and particularly the Models section where detailed information about the latest versions like YOLO12 is provided 🌟. For your provided link in Chinese, you might find the localized YOLO12 documentation helpful.

If this is a 🐛 Bug Report, we kindly request you include a minimum reproducible example to assist us in addressing the issue more effectively.

If you have a specific ❓ Question or need clarification on YOLO12, please provide additional context to help us better understand your inquiry. This could include model configurations, use cases, or any challenges you're encountering.

To stay connected with the Ultralytics community:

• Chat live on Discord 🎧.
• Join in-depth discussions on Discourse.
• Share insights or get inspired on our Subreddit.

Upgrade

To ensure you're working with the latest and greatest features, upgrade your ultralytics package with the following command in a Python >=3.8 environment:

pip install -U ultralytics

Verified Environments

You can run YOLO in any of the following verified environments:

• Notebooks with free GPU:
• Cloud platforms like Google Cloud (Quickstart Guide) or AWS (Quickstart Guide).
• Prebuilt Docker Images: See our Docker Quickstart Guide

Status

This automated response is here to assist you 🤖, but rest assured, an Ultralytics engineer will review your discussion and follow up soon! 🚀

[1029459794]

Hello, I would like to inquire about the release schedule for the pre-trained models of YOLOv12 for image classification.

[glenn-jocher]

@1029459794 yOLO12 classification models (YOLO12-cls) are available now and can be accessed through our Python package or CLI. For implementation details, see the Classification documentation. However, we generally recommend using YOLO11 models for most applications due to YOLO12's higher computational demands and training instability.

[1029459794]

Thank you for your response. I am using the following code to train the model:

from ultralytics import YOLO

model = YOLO('yolov12-cls.pt')

# Train the model
results = model.train(data="/home/Sushi-50", epochs=100, imgsz=640)

However, I am encountering the following error:

Traceback (most recent call last):
  File "/home/liangqingqi/yolov12/50.py", line 8, in <module>
    model = YOLO('yolov12-cls.pt') 
            ^^^^^^^^^^^^^^^^^^^^^^
  File "/home/liangqingqi/yolov12/ultralytics/models/yolo/model.py", line 23, in __init__
    super().__init__(model=model, task=task, verbose=verbose)
  File "/home/liangqingqi/yolov12/ultralytics/engine/model.py", line 148, in __init__
    self._load(model, task=task)
  File "/home/liangqingqi/yolov12/ultralytics/engine/model.py", line 291, in _load
    self.model, self.ckpt = attempt_load_one_weight(weights)
                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/liangqingqi/yolov12/ultralytics/nn/tasks.py", line 905, in attempt_load_one_weight
    ckpt, weight = torch_safe_load(weight)  # load ckpt
                   ^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/liangqingqi/yolov12/ultralytics/nn/tasks.py", line 832, in torch_safe_load
    ckpt = torch.load(file, map_location="cpu")
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/liangqingqi/yolov12/ultralytics/utils/patches.py", line 86, in torch_load
    return _torch_load(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/liangqingqi/anaconda3/envs/yolov12/lib/python3.11/site-packages/torch/serialization.py", line 1425, in load
    with _open_file_like(f, "rb") as opened_file:
         ^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/liangqingqi/anaconda3/envs/yolov12/lib/python3.11/site-packages/torch/serialization.py", line 751, in _open_file_like
    return _open_file(name_or_buffer, mode)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/liangqingqi/anaconda3/envs/yolov12/lib/python3.11/site-packages/torch/serialization.py", line 732, in __init__
    super().__init__(open(name, mode))
                     ^^^^^^^^^^^^^^^^
FileNotFoundError: [Errno 2] No such file or directory: 'yolov12-cls.pt'

How can I correctly train the model?

[glenn-jocher]

The error occurs because the model file 'yolov12-cls.pt' isn't found in your path. YOLO12 models are community contributions and not officially distributed by Ultralytics. To resolve this, download the YOLO12-cls weights from the original repository at sunsmarterjie/yolov12 and ensure the file path is correct. For production use, we recommend YOLO11 models due to their stability and optimized performance across tasks.

[ChengxiLii]

Hi all,

Is there a pre-trained model of yolo12_obb available for download? thanks.

[glenn-jocher]

@ChengxiLii thank you for your question! YOLO12-OBB pre-trained models are not currently available in the Ultralytics model zoo. For oriented bounding box detection, we recommend using YOLO11-OBB which offers stable performance and pre-trained weights. You can find implementation details in the Oriented Object Detection documentation and YOLO11 Model Overview.

We appreciate your interest in pushing the boundaries of computer vision with Ultralytics! 🚀

[LiangChaoWei512]

May I ask if I want to use YOLOV12's pre trained weights to train my dataset? According to the introduction, FlashAttention is not enabled by default, but my hardware supports it. Where should I enable FlashAttention

[EtahReign]

Hello,

Are there any pretrained yolov12n-face.pt models available for download for face tracking purposes?

[clousy188]

在5090 32G上使用yolo12l训练模型的时候，数据量大概7万左右，batch_size=24 内存占用在24G左右，为啥速度只有2-3it/s 是哪里设置有问题？下面是训练参数：
rain_args = dict(
data="data.yaml",
epochs=500,
imgsz=640,
batch=24,
device=device,
workers=16,

    cache=False,      
    amp=True, 
    
    # --- 减少 CPU 负担 ---
    mosaic=0.0,        # 暂时关掉 Mosaic 看看速度是否飙升
    mixup=0.0,         # 同上
)

[glenn-jocher]

yolo12l 本身是重注意力结构，吞吐会比同尺度的 CNN 系列低不少，所以你看到的 2–3 it/s 在 batch=24 下未必异常（折算大约是 48–72 img/s）；如果你目标是更高训练/CPU 端速度，建议直接换到更稳更快的 Ultralytics YOLO26（YOLO12 更偏研究/基准用途，见 YOLO12 文档）。想确认是不是环境/数据读取瓶颈的话，请把 yolo checks 输出和完整文本训练日志（含第一轮的进度条行、CUDA:/显卡名、DataLoader 相关信息）贴在这里；另外你现在 cache=False 可能会被磁盘 IO 卡住、workers=16 在部分系统上也会造成额外调度开销，建议先用下面这组参数做一次对照测试（同一数据/同一 imgsz），看 it/s 是否明显变化：

yolo checks
yolo train model=yolo12l.pt data=data.yaml epochs=2 imgsz=640 batch=24 device=0 workers=8 cache=True mosaic=0.0 mixup=0.0