本文共 414 字，大约阅读时间需要 1 分钟。

Deep residual learning for image recognition

参考论文：《Deep residual learning for image recognition》

pytorch 源码：https://pytorch.org/docs/stable/_modules/torchvision/models/resnet.html

细节：

1. We adopt batch normalization(BN) right after each convolution and before activation
2. SGD (batch size: 256)
3. weight decay: 0.0001; momentum: 0.9
4. 每一层（3，4，5）刚开始，都会有一个stride = 2的卷积用于降采样
5. basic block channel翻倍，而bottleneck block channel翻四倍