Linear Learning Rate Scaling

• If [He Initialization ] is used, 0.1 is a good learning rate for batch size 256 and for a larger b, $0.1\times \frac{\mathrm{b}}{256}$ is okay