chen1474147 / Deep3DPose

MIT License
65 stars 32 forks source link

Circular Training #7

Open priyapaul opened 7 years ago

priyapaul commented 7 years ago

I have spent some time in understanding your paper. I have some questions on domain adaptation

point 1:

After 1st stage of training for several iterations for a good classifier and regressor, you change the dataset to change the labels to 0.5 , freeze the domain mixer/classifier layers , by keeping the lr_mult: =0. train again(stage 2) for several iterations ? Please spare some time to correct me if I am wrong

point 2:

Note that we train the network in a circular way, stage1 is alternate with stage2. This is because domain classifier is based on current feature extractor. If we modify the extractor in stage2, then we need to refine new classifier in stage1. So it is a circular process

After 2 stages of training you have already got the pose regressor without any gap between syn and real data. Then why do we need to do it circularly again? Could you please explain that? " This is because domain classifier is based on current feature extractor" , but feature extractor changes in every iterations anyway? I didn't understand this point ! Thanks in advance .