加载 torchvision 模型

You rarely need to build a famous net from scratch. PyTorchs torchvision library ships them ready to use.

From ResNet to VGG to modern nets, torchvision.models is a library of proven architectures at your fingertips.

Grabbing an architecture is a single call. Here you get a ResNet-18 with random, untrained weights.

The real power is loading pretrained weights learned on ImageNet, so the model already knows useful visual features.

Those weights are a huge head start. You can fine-tune them on your own task with far less data.

Pretrained models expect specific input sizes and normalization. The weights object carries the exact transforms to apply.

Before predicting, call eval so layers like batch norm and dropout behave correctly for inference.

ImageNet has 1000 classes. To reuse the net, replace its final fully connected layer to match your classes.

For quick transfer learning, freeze the backbone so only your new head trains, which is fast and data-light.

The weights metadata even lists the human-readable categories, so you can turn an index back into a label.

With one import you reuse years of research and compute. This is the everyday workflow of applied deep learning.

With torchvision you load proven architectures and pretrained weights in a line, then adapt them to your task. You are ready to build!