YOLOv4-TT100K 中国交通标志识别 (Ascend NPU Adapted)

# 确认 NPU 环境
npu-smi info

# Python 环境
pip install torch==2.9.0 torch_npu==2.9.0.post1
pip install opencv-python numpy modelscope

pip install modelscope
modelscope download --model Genius-Society/yolov4_tt100k --local_dir ./yolo_model

import torch
import torch_npu
from yolov4_pytorch import YOLOv4

# 加载模型
model = YOLOv4('yolov4-tt100k.cfg')
model.load_darknet_weights('yolov4-tt100k_best.weights')
model.eval()
model = model.to('npu:0')

# 推理
import cv2
import numpy as np

img = cv2.imread('traffic_sign.jpg')
img = cv2.resize(img, (608, 608))
tensor = torch.from_numpy(img.transpose(2,0,1)/255.0).float().unsqueeze(0).to('npu:0')

with torch.no_grad():
    outputs = model(tensor)

# 图片推理
python3 inference.py --source traffic.jpg --output result.jpg --device npu

# 性能基准测试
python3 inference.py --device npu --benchmark --benchmark-iters 100

# 精度评估
python3 evaluate.py --device npu --iters 100

YOLOv4-TT100K
├── CSPDarknet53 Backbone (Mish activation)
│   ├── Stem: Conv 3→32, stride=1
│   ├── Stage1: Conv 32→64, stride=2 + CSPBlock(1×)
│   ├── Stage2: Conv 64→128, stride=2 + CSPBlock(2×)
│   ├── Stage3: Conv 128→256, stride=2 + CSPBlock(8×)
│   ├── Stage4: Conv 256→512, stride=2 + CSPBlock(8×)
│   └── Stage5: Conv 512→1024, stride=2 + CSPBlock(4×)
├── SPP (Spatial Pyramid Pooling)
├── PANet (Path Aggregation Network)
│   ├── FPN Top-down path (2× Upsample)
│   └── PAN Bottom-up path (2× Downsample)
└── YOLOv3 Head (3 scales)
    ├── Head1: 76×76 (stride 8, small objects)
    ├── Head2: 38×38 (stride 16, medium objects)
    └── Head3: 19×19 (stride 32, large objects)