search

ustc-vim / dualnet

DualNet: Learn Complementary Features for Image Recognition
19 stars 13 forks source link

Iterative training #1

Open xiaoboCASIA opened 7 years ago

xiaoboCASIA commented 7 years ago

Hello! How to do the iterative training? In the given demo, we only need to set the max_iter=1? Thanks for your help!

hshustc commented 7 years ago

In most cases, the performance gain is largely obtained in the first iteration. In our experiments, we have tried max_iter=2 on UEC-FOOD-100 and no larger value is tested.

Zheng222 commented 7 years ago

Hello, hshustc. With respect to the usage of this code, I found that the iterative training is just frozening the subnetwork 1 and training the subnetwork 2 from dualnet/dualnet-dataset/pklcifar100/model/v4_ninnet/res_e1/pklcifar100_nin_res_e1_log.txt file, dualnet/dualnet-dataset/pklcifar100/model/v4_ninnet/res_e1/res_e1_solver.prototxt file and dualnet/dualnet-dataset/pklcifar100/model/v4_ninnet/res_e1/res_el_train_val.prototxt file. Meanwhile, from these files, I found the result of log file corresponds to iterative training (Fused Classifier) of the paper, while iterative training (classifier average) is not provided in this code. From my point of view, this training strategy is little different from that of paper. Could you please explain in detail? Thank you a lot! @ustc-vim

hshustc commented 6 years ago

Iterative training (class average) is only for the test phase. You can use the model after iterative training to get the output of three classifiers and average them.