2026 年初，YOLO 迎來了新變化，都非常亮眼 ??

一是 YOLO-Master：引入了 ES-MoE，讓計算更智能。會依據(jù)輸入自適應(yīng)做計算，不再是靜態(tài)的。

二是 YOLO26：更簡化的設(shè)計，直接端到端出檢測結(jié)果。不用再搞后處理了呢，利于邊緣做部署。

既然 YOLO 上新了，那一起玩一玩吧 ??

• 體驗 YOLO-Master: https://huggingface.co/spaces/gatilin/YOLO-Master-WebUI-Demo
• 體驗 YOLO26: https://platform.ultralytics.com/ultralytics/yolo26

以下則是動手實踐了呢。

• 代碼: https://github.com/ikuokuo/start-deep-learning

YOLO-Master

環(huán)境

準備 Conda 環(huán)境，

conda create -n yolom python=3.12
conda activate yolom

# Install PyTorch (CPU version)
pip install torch torchvision
# Install PyTorch with CUDA (version <= nvidia-smi shown)
#  https://pytorch.org/get-started/locally
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu130

準備 YOLO-Master，

# Clone the repository
git clone --depth 1 https://github.com/Tencent/YOLO-Master
cd YOLO-Master

# Install dependencies
pip install -r requirements.txt
pip install -e .

# Optional: Install FlashAttention for faster training (CUDA required)
#   https://github.com/Dao-AILab/flash-attention
pip install flash_attn

準備模型，

# 獲取模型，YOLO-Master-EsMoE-N
wget https://huggingface.co/gatilin/YOLO-Master-ckpts-v0/resolve/main/YOLO-Master-EsMoE-N/YOLO-Master-EsMoE-N.pt?download=true
ln -s YOLO-Master-EsMoE-N.pt yolo_master_n.pt

推理

from ultralytics import YOLO

# 加載模型，Nano 版
model = YOLO("yolo_master_n.pt")
# model = YOLO("runs/detect/train/weights/best.pt")

# 檢測圖像
results = model.predict("data/dog.jpg")
results[0].show()

# 保存結(jié)果
results[0].save("result/dog.jpg")

訓(xùn)練

from ultralytics import YOLO

# 加載模型，從 YAML 構(gòu)建（從零開始）
# model = YOLO('cfg/models/master/v0/det/yolo-master-n.yaml')
# 加載模型，Nano 版（全量微調(diào)）
model = YOLO("yolo_master_n.pt")

# 訓(xùn)練模型，用 COCO8 數(shù)據(jù)集測試
#   結(jié)果在 runs/detect/train10/weights/best.pt
results = model.train(
    data='coco8.yaml',  # data='coco.yaml',
    epochs=100,         # epochs=600,
    batch=8,            # batch=256,
    imgsz=640,
    device="0",         # device="0,1,2,3",  # 如果使用多 GPU
    scale=0.5,
    mosaic=1.0,
    mixup=0.0,
    copy_paste=0.1
)

YOLO-Master 還可以 LoRA 微調(diào) ??

YOLO26

環(huán)境

準備 Conda 環(huán)境，

conda create -n yolo26 python=3.12
conda activate yolo26

# Install PyTorch (CPU version)
pip install torch torchvision
# Install PyTorch with CUDA (version <= nvidia-smi shown)
#  https://pytorch.org/get-started/locally
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu130

準備 YOLO26，

# Install ultralytics package
pip install -U ultralytics

# Install dependencies
pip install faster-coco-eval

推理

from ultralytics import YOLO

# 加載模型，Nano 版
model = YOLO("yolo26n.pt")
# model = YOLO("runs/detect/train/weights/best.pt")

# 檢測圖像
results = model.predict("data/dog.jpg")
results[0].show()

# 保存結(jié)果
results[0].save("result/dog.jpg")

訓(xùn)練

from ultralytics import YOLO

# 加載模型，Nano 版
model = YOLO("yolo26n.pt")

# 訓(xùn)練模型，用 COCO8 數(shù)據(jù)集測試
#   結(jié)果在 runs/detect/train/weights/best.pt
results = model.train(data="coco8.yaml", epochs=100, imgsz=640)

結(jié)語

YOLO 訓(xùn)練越來越簡便了呢，部署仍舊看平臺咯。