cait_xxs36_384 - 昇腾 NPU 适配

# 安装依赖（使用清华 PyPI 镜像）
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple torch torchvision timm modelscope pillow numpy

# 安装 torch_npu（昇腾 NPU 支持）
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple torch_npu

cd cait_xxs36_384
python3 inference.py --model cait_xxs36_384 --device cpu

cd cait_xxs36_384
python3 inference.py --model cait_xxs36_384 --device npu

cd cait_xxs36_384
python3 compare_cpu_npu.py --model cait_xxs36_384

cait_xxs36_384/
├── inference.py          # NPU/CPU 推理脚本
├── compare_cpu_npu.py    # CPU vs NPU 精度对比脚本
├── requirements.txt      # 依赖包列表
├── precision_results.json # 精度测试结果 (JSON)
├── precision_test.log    # 精度测试日志
├── terminal_screenshot.png # 模拟终端输出截图
└── README.md             # 本文件

import torch
import torch_npu
from PIL import Image
from timm import create_model
from timm.data import create_transform, IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD

model_name = "cait_xxs36_384"
model = create_model(model_name, pretrained=False, num_classes=1000)

# 加载本地权重
import timm
timm.models.load_checkpoint(model, "./model.safetensors")
model.eval()
model.to("npu:0")

# 预处理图像
transform = create_transform(
    input_size=(3, 384, 384),
    is_training=False,
    mean=IMAGENET_DEFAULT_MEAN,
    std=IMAGENET_DEFAULT_STD,
    interpolation='bicubic',
)
image = Image.open("test.jpg").convert("RGB")
input_tensor = transform(image).unsqueeze(0).to("npu:0")

# 推理
with torch.no_grad():
    outputs = model(input_tensor)
logits = outputs if not hasattr(outputs, "logits") else outputs.logits
probs = torch.softmax(logits, dim=-1)
pred = torch.argmax(probs, dim=-1).item()
print(f"Predicted class: {pred}")

python3 compare_cpu_npu.py --model cait_xxs36_384