1 warm up: The first ten epochs learn to increase linearly first, and then to maintain, at 20 and 40 epochs, the learning rate decays to 0.1
2 Random erasure of data
3 label smooth, make a change to the cross entropy
4 Set the last downsampling and convolution step size of resnet50 to 1
5 When using triplet and ID loss, the features sent to triplet loss are passed through a BN layer, and then sent to ID Loss. In this BN layer, bias is reset to zero and not used
6 加center loss