Training parameters (lr and weight decay) change.

[wuyuebupt]

Change the lr schedule from steplr to multistep lr following the paper. Change weight decay according to the incremental stages. Early stages have larger weight decay than late stages. I got one close result [0.856, 0.72125, 0.6556666666666666, 0.6015, 0.5577], see the log file in the folder logs.

The distillation part is also different from our implementation. But I think your implementation is kind of fine.