ResNet-18 NPU 适配模型

pip install -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements.txt

python inference.py --device cpu \
  --model-dir /path/to/model \
  --output-dir /path/to/output

python inference.py --device npu \
  --model-dir /path/to/model \
  --output-dir /path/to/output

python compare_cpu_npu.py /path/to/output /path/to/output /path/to/output

resnet-18-npu/
├── inference.py          # 推理脚本（支持 CPU/NPU）
├── compare_cpu_npu.py    # CPU/NPU 精度对比脚本
├── requirements.txt      # 依赖列表
├── readme.md             # 本文件
├── model/                # 原始模型文件
└── output/               # 推理和对比结果
    ├── cpu_result.json   # CPU 推理结果
    ├── cpu_logits.npy    # CPU Logits
    ├── npu_result.json   # NPU 推理结果
    ├── npu_logits.npy    # NPU Logits
    └── comparison_result.json  # 精度对比结果

import torch
import torch_npu
from transformers import AutoImageProcessor, ResNetForImageClassification
from PIL import Image
import numpy as np

# 加载模型
model_dir = "./model/microsoft/resnet-18"
image_processor = AutoImageProcessor.from_pretrained(model_dir)
model = ResNetForImageClassification.from_pretrained(model_dir)

# 迁移到 NPU
device = torch.device("npu:0")
model = model.to(device)
model.eval()

# 准备输入
img = Image.open("your_image.jpg")
inputs = image_processor(img, return_tensors="pt")
inputs = {k: v.to(device) for k, v in inputs.items()}

# 推理
with torch.no_grad():
    outputs = model(**inputs)

# 获取预测结果
logits = outputs.logits.cpu().numpy()
probs = torch.nn.functional.softmax(torch.tensor(logits), dim=-1).numpy()
top5_idx = np.argsort(logits[0])[-5:][::-1]

for rank, idx in enumerate(top5_idx):
    label = model.config.id2label[int(idx)]
    print(f"{rank+1}. {label} (prob: {probs[0][idx]:.4f})")

# NPU 推理
python3 inference.py --device npu

# CPU 推理
python3 inference.py --device cpu