Deep residual learning for image recognition
参考论文：《Deep residual learning for image recognition》 pytorch 源码：pytorch.org/docs/stable…
细节：
- We adopt batch normalization(BN) right after each convolution and before activation
- SGD (batch size: 256)
- weight decay: 0.0001; momentum: 0.9
- 每一层（3，4，5）刚开始，都会有一个stride = 2的卷积用于降采样
- basic block channel翻倍，而bottleneck block channel翻四倍
