cv_mobileface_hand-static Ascend NPU 部署指南

cv_mobileface_hand-static-ascend/
├── inference.py          # 推理测试脚本
├── log.txt               # 测试日志
├── precision_result.json # 精度测试结果 JSON
├── README.md             # 本文档
└── test_hand.jpg         # 测试图片 (模型目录下)

docker exec -it test-modelagent bash

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

pip install torch_npu opencv-python numpy

cd /data/ysws/agentsp/5-19-1/cv_mobileface_hand-static-ascend/

# 运行默认推理
python3 inference.py

cd /data/ysws/agentsp/5-19-1/cv_mobileface_hand-static-ascend/

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

============================================================
Hand Static Detection - Ascend NPU Test
Output: /data/ysws/agentsp/5-19-1/cv_mobileface_hand-static-ascend
============================================================

Mode: PRECISION TEST

============================================================
Loading Model
============================================================
Model bin: /data/ysws/agentsp/5-19-1/cv_mobileface_hand-static/iic/cv_mobileface_hand-static/pytorch_model.bin (exists: True)
PyTorch: /data/ysws/agentsp/5-19-1/cv_mobileface_hand-static-ascend/hand_npu.pt (exists: True)
Test image: /data/ysws/agentsp/5-19-1/cv_mobileface_hand-static/iic/cv_mobileface_hand-static/test_hand.jpg

============================================================
Loading Test Image
============================================================
Image shape: (309, 310, 3) (H, W, C)
Input tensor shape: (1, 3, 112, 112)

============================================================
Running CPU Inference
============================================================
CPU output shape: torch.Size([1, 15])
CPU prediction: d_bixin (score: 0.0715)
CPU time: 0.7390s

============================================================
Running NPU Inference
============================================================
NPU output shape: torch.Size([1, 15])
NPU prediction: unrecog (score: 0.0727)
NPU time: 4.0450s
Speedup: 0.18x

============================================================
Precision Test Results
============================================================
CPU prediction: 1 (d_bixin), score: 0.071512
NPU prediction: 14 (unrecog), score: 0.072718
Prediction match: False
Score difference: 0.001206
Prediction match: FAIL
Score threshold (0.01): PASS

Status: FAIL

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

import cv2
import torch
import numpy as np

MODEL_DIR = "/data/ysws/agentsp/5-19-1/cv_mobileface_hand-static/iic/cv_mobileface_hand-static"
LABELS = ['bixin', 'd_bixin', 'd_first_left', 'd_fist_right', 'd_hand',
          'fashe', 'fist', 'five', 'ok', 'one', 'tuoju', 'two', 'yaogun', 'zan', 'unrecog']

# 图像预处理
def preprocess(img):
    img = cv2.resize(img, (112, 112))
    img = img.astype(np.float32) / 255.0
    mean = np.array([0.5, 0.5, 0.5], dtype=np.float32).reshape(1, 1, 3)
    std = np.array([0.5, 0.5, 0.5], dtype=np.float32).reshape(1, 1, 3)
    img = (img - mean) / std
    img = np.transpose(img, (2, 0, 1))
    return img[np.newaxis, :, :, :].astype(np.float32)

# 加载模型
state = torch.load(f"{MODEL_DIR}/pytorch_model.bin", map_location='cpu')
model = HandModel(state)
model.eval()

# 读取并预处理图像
img = cv2.imread(f"{MODEL_DIR}/test_hand.jpg")
img_tensor = torch.from_numpy(preprocess(img))

# NPU 推理
device = torch.device("npu:0")
model = model.to(device)
img_tensor = img_tensor.to(device)

with torch.no_grad():
    output = model(img_tensor)
    probs = torch.softmax(output, dim=1)[0]
    pred = torch.argmax(probs).item()
    score = probs[pred].item()

print(f"Prediction: {LABELS[pred]} (score: {score:.4f})")

import torch
import numpy as np
import cv2

# 批量图像路径
image_paths = ["hand1.jpg", "hand2.jpg", "hand3.jpg"]

# 批量预处理
batch = []
for path in image_paths:
    img = cv2.imread(path)
    img_tensor = preprocess(img)
    batch.append(img_tensor)

batch_tensor = torch.from_numpy(np.concatenate(batch, axis=0))

# 批量推理
batch_tensor = batch_tensor.to(device)
with torch.no_grad():
    outputs = model(batch_tensor)
    probs = torch.softmax(outputs, dim=1)
    preds = torch.argmax(probs, dim=1)

for i, pred in enumerate(preds):
    print(f"Image {i}: {LABELS[pred.item()]} (score: {probs[i][pred].item():.4f})")

{
  "input_size": 112,
  "input_channels": 3,
  "num_classes": 15,
  "batch_size": 1,
  "preprocess": "mean=0.5, std=0.5"
}