a course work which is uesd in cloud platform

train.py
Train WideResNet from scratch on CIFAR-10/100 and save a checkpoint
→ used to obtain our own pretrained model (e.g. wrn_cifar_best.pth).

从零开始在 CIFAR-10/100 上训练 WideResNet，并保存 checkpoint，
→ 用于生成自训练预训练权重（如 wrn_cifar_best.pth）。

• wideresnet.py
  Implementation of WideResNet-XX-k with residual BasicBlock and NetworkBlock.
  This is the shared supervised backbone used by both training and transfer learning.

  WideResNet-XX-k 的网络结构实现（残差 BasicBlock + NetworkBlock）。
  在本项目中作为监督学习主干网络，同时被 train.py 和 transferlearning_train.py 调用。

• transferlearning_train.py
  Fine-tune WideResNet on CIFAR with transfer learning.
  It loads a pretrained checkpoint (e.g. the one produced by train.py),
  uses stronger data augmentation (RandomErasing), and trains with a smaller LR.

  利用 迁移学习 对 WideResNet 进行微调。
  从预训练 checkpoint（例如 train.py 训练得到的模型）加载权重，
  使用更强的数据增强（RandomErasing），并采用更小的学习率进行微调。