Learning Rate Configuration about efficientps

[mohan1914]

To freeze the iABN layers, you can set the require_grad=False in the config file. And then add the extra 10 epochs in the mutli-step parameters and the total epoch.
However, we don't use this lr schedule for training our network with this repository. Instead, we follow the following additional training protocol:
We start another round of training with the same initial lr, but halve the multi-step iteration steps as well as the total epoch.
(the model of this round is initialized from the previous one)