torch_verify.py

import torch
from torchvision import models, transforms
from PIL import Image


# Downloading: "https://download.pytorch.org/models/resnet50-0676ba61.pth" to /root/.cache/torch/hub/checkpoints/resnet50-0676ba61.pth
# 加载预训练的ResNet50模型，并指定不使用GPU
model = models.resnet50(pretrained=True)
model.eval()  # 将模型设置为评估模式
device = torch.device("cuda:0")  # 指定设备为CPU
model.to(device)  # 将模型转移到CPU上

# 图片预处理
transform = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

image_path = 'dog.jpg'  # 替换为你的图片路径
image = Image.open(image_path).convert('RGB')
image = transform(image)

# 添加batch维度
image = image.unsqueeze(0)

# 在CPU上进行推理
with torch.no_grad():
    outputs = model(image.to(device))
    
# 获取预测类别
_, predicted_class = torch.max(outputs, 1)
predicted_class_idx = predicted_class.item()

# 打印预测类别
print(f"Predicted class: {predicted_class_idx}")

# 验证分类结果:
# https://github.com/HoldenCaulfieldRye/caffe/blob/master/data/ilsvrc12/synset_words.txt
synset_words_path = "synset_words.txt"
with open(synset_words_path, 'r') as f:
    synset_words = f.readlines()

# 每行前去除空格和制表符，并根据索引获取类别名称
class_names = [line.strip() for line in synset_words]
predicted_class_name = class_names[predicted_class_idx+1]  # 注意索引可能从1开始，所以+1
print(f"Predicted class name: {predicted_class_name}")  # 彭布罗克，彭布罗克威尔士柯基犬

with torch.no_grad():
    input_tensor = torch.ones((1,3,224,224), dtype=torch.float).to("cuda:0")
    save_path = "./resnet50.onnx"
    dynamic_shape = {
        'input': {0: 'batch_size', 2: 'height', 3: 'width'},
        'output': {0: 'batch_size', 1: 'n_class'}
    }

    torch.onnx.export(
        model, 
        (input_tensor, ),
        f=save_path,
        verbose=False,
        opset_version=17,
        input_names=['input'],
        output_names=['output'],
        dynamic_axes=dynamic_shape,
    )

    print("export success")