search

sanghoon / pva-faster-rcnn

Demo code for PVANet
https://arxiv.org/abs/1611.08588
Other
650 stars 241 forks source link

What is the train details on ImageNet2012? #93

Open qingzhouzhen opened 6 years ago

qingzhouzhen commented 6 years ago

I want to pretrain pvanet on ImageNet, but can not find the details arguments such opotimizer, how to initilize weight and so on.(except learning rate and 192*192 I found in article) @sanghoon

twmht commented 6 years ago

@sanghoon

Can you share your plateau window size when training imageNet?

qingzhouzhen commented 6 years ago

what your problem? @twmht , my problem is I pretrain pvanet on ImageNet2012 got 66%, not 70%, I want to transfer pvanet to MXNet.

twmht commented 6 years ago

@qingzhouzhen

I go 61%, but the learning rate policy is step, not plateau.

Since @sanghoon does not provide the solver to train imageNet, I don't know how to set the plateau parameter to reach 2M iterations.

twmht commented 6 years ago

@qingzhouzhen

I am curious about what's your learning rate policy?

qingzhouzhen commented 6 years ago

--lr-factor=0.36 --lr-step-epoch=30,60,80,100,120,140 --num-epochs=160 by the way, what is your framework? Caffe , TensorFlow, or Mxnet

twmht commented 6 years ago

@qingzhouzhen

I am using caffe, so you do not use the learning rate policy plateau?

qingzhouzhen commented 6 years ago

I do not no what is plateau, MXNET does not have this , at least I do not know.