cv_resnet18_license-plate-detection_damo Ascend NPU 部署指南

cv_resnet18_license-plate-detection_damo-ascend/
├── inference.py              # 推理测试脚本
├── log.txt                   # 测试日志
├── README.md                 # 本文档
├── test_license_plate.jpg    # 测试图片
├── result_visualization.jpg   # 检测结果可视化
├── precision_result.json     # 精度测试结果
├── inference_result.json     # 推理输出结果
└── fusion_result.json        # 图融合结果

docker exec -it test-modelagent bash

source /usr/local/Ascend/ascend-toolkit/set_env.sh

pip install torch torch_npu opencv-python numpy

cd /data/ysws/agentsp/5-19-1/cv_resnet18_license-plate-detection_damo-ascend/

# 运行车牌检测
python3 inference.py

cd /data/ysws/agentsp/5-19-1/cv_resnet18_license-plate-detection_damo-ascend/

# 运行完整精度测试
python3 inference.py precision_test

============================================================
License Plate Detection - Ascend NPU Test
Output: /data/ysws/agentsp/5-19-1/cv_resnet18_license-plate-detection_damo-ascend
============================================================

Mode: PRECISION TEST
NPU available: True
Model: .../pytorch_model.pt (exists: True)

============================================================
Loading Test Image
============================================================
Original image shape: (435, 580, 3)
Input tensor shape: torch.Size([1, 3, 1024, 1024])
Scale: 1.7655172413793103, Padded size: (768, 1024)

============================================================
Building Model
============================================================
Loading state dict...
Model built successfully

============================================================
Running CPU Inference
============================================================
CPU inference time: 44.0268s
Heatmap shape: torch.Size([1, 1, 512, 512])
Reg shape: torch.Size([1, 2, 512, 512])
WH shape: torch.Size([1, 8, 512, 512])
Cls shape: torch.Size([1, 4, 512, 512])
Ftype shape: torch.Size([1, 11, 512, 512])
Detected license plates (CPU): 0

============================================================
Running NPU Inference
============================================================
NPU inference time: 3.8880s
Detected license plates (NPU): 0

============================================================
Precision Test Results
============================================================
Heatmap max diff: 1.649661e-01
Reg max diff: 3.209072e-02
WH max diff: 1.525805e-01
Cls max diff: 1.359105e-01
Ftype max diff: 1.226125e-01
Heatmap relative error: 3.554848e-02 (3.5548%)
Reg relative error: 5.299350e-02 (5.2994%)
WH relative error: 3.462220e-02 (3.4622%)
Cls relative error: 2.428787e-02 (2.4288%)
Ftype relative error: 2.512269e-02 (2.5123%)

Heatmap: FAIL (threshold: 1.0%)
Reg: FAIL (threshold: 1.0%)
WH: FAIL (threshold: 1.0%)
Cls: FAIL (threshold: 1.0%)
Ftype: FAIL (threshold: 1.0%)
Detection count match: PASS

Overall Status: FAIL

============================================================
Saving Results
============================================================
Visualization saved: .../result_visualization.jpg

============================================================
Test Complete!
============================================================

import torch
import cv2
import numpy as np

MODEL_PATH = "/data/ysws/agentsp/5-19-1/cv_resnet18_license-plate-detection_damo/iic/cv_resnet18_license-plate-detection_damo/pytorch_model.pt"

model = LicensePlateDetectionModel(MODEL_PATH)
device = torch.device("npu:0")
model = model.to(device)
model.eval()

img = cv2.imread("test_license_plate.jpg")
img_tensor, scale, new_h, new_w = preprocess(img)
img_tensor = torch.from_numpy(img_tensor).unsqueeze(0).to(device)

with torch.no_grad():
    hm, reg, wh, cls, ftype = model(img_tensor)

detections = decode_predictions(hm.cpu(), reg.cpu(), wh.cpu(), cls.cpu(), ftype.cpu(), scale, img.shape)
print(f"检测到 {len(detections)} 个车牌")

import os

image_dir = "/path/to/images"
output_dir = "/path/to/output"
os.makedirs(output_dir, exist_ok=True)

image_files = [f for f in os.listdir(image_dir) if f.endswith('.jpg')]

for img_file in image_files:
    img_path = os.path.join(image_dir, img_file)
    img = cv2.imread(img_path)

    img_tensor, scale, new_h, new_w = preprocess(img)
    img_tensor = torch.from_numpy(img_tensor).unsqueeze(0).to(device)

    with torch.no_grad():
        hm, reg, wh, cls, ftype = model(img_tensor)

    detections = decode_predictions(hm.cpu(), reg.cpu(), wh.cpu(), cls.cpu(), ftype.cpu(), scale, img.shape)

    vis_path = os.path.join(output_dir, f"result_{img_file}")
    visualize(img, detections, vis_path)

# 图像预处理
target_size = 1024  # 长边缩放到 1024
scale = target_size / max(h, w)
new_h, new_w = int(h * scale), int(w * scale)

# 归一化 (ImageNet mean/std)
MEAN = np.array([123.675, 116.28, 103.53], dtype=np.float32)
STD = np.array([58.395, 57.12, 57.375], dtype=np.float32)
img_norm = (img_resized.astype(np.float32) - MEAN) / STD

# Pad to square (1024 x 1024)
pad_h = target_size - new_h
pad_w = target_size - new_w