search

syfafterzy / SVDNet-for-Pedestrian-Retrieval

Code for https://arxiv.org/abs/1703.05693
138 stars 47 forks source link

About solver.prototxt #3

Closed rainsoulsrx closed 6 years ago

rainsoulsrx commented 6 years ago

I notice that base_lr is set to be 0.001 in all solver_restraint.prototxt, normally, base_lr should change smaller as the training process goes on, but the given code keep on using 0.001. will it converg very well in the end?

ghost commented 6 years ago

Thanks for Ur question.

Actually, the base_lr do change smaller as the training process goes on.

Please see details at: https://github.com/syfafterzy/SVDNet-for-Pedestrian-Retrieval/blob/master/SVDNet/resnet/base_lr.txt.

rainsoulsrx commented 6 years ago

Thx, but the caffenet solver_restraint.prototxt, base_lr: 0.001, lr_policy: "step", stepsize: 2000, max_iter: 2000, that is to say, the lr keeps to be 0.01 during the restraint stage, right?

syfafterzy commented 6 years ago

Yes, you are right. For CaffeNet, we only fine-tune the network to good convergence during the last RRI, to reduce training time.

rainsoulsrx commented 6 years ago

Thank you~~