Why do you change initial GoogleNet loss1 output from 1000 to 128? Also,why do you change the lr_mult simultaneously ?

[jsjs0827]

[XinyiXuXD]

The 128 issue: the author sets the embedding size of each submodel to 128, and finally obtains a 384 ensemble feature vector to represent each sample when testing. This embedding size is refer to "Deep metric learning via lifted structured feature embedding"

[abcdvzz]

So why do you change the lr_mult simultaneously?

[XinyiXuXD]

because the final fc has no pretrain model to fine tune on, but the otherlayers fine tune on googlenet model that pretrained on imagenet dataset