ustc-vim / dualnet

DualNet: Learn Complementary Features for Image Recognition

DualNet

This repository is for the paper "DualNet: Learn Complementary Features for Image Recognition"

Download Links: [Paper] [Supplementary Material]

./caffe-dualnet: modified from Caffe (https://github.com/BVLC/caffe)
./dualnet-dataset: the prototxt defining the models for each dataset

Usage

We illustrate the training process taking DNI on CIFAR100 as an example:

Training standard deep model

./build/tools/caffe.bin train -solver data/pklcifar100/model/v4_ninnet/solver.prototxt 2>&1 | tee -a data/pklcifar100/model/v4_ninnet/pklcifar100_nin_log.txt

Iterative Training (max_iter is set to 1 in this case)

ln -s data/pklcifar100/model/v4_ninnet/snapshot/v4_pklcifar100_nin_iter_120000.caffemodel data/pklcifar100/model/v4_ninnet/pklcifar100_nin_train_iter_120000.caffemodel_coarse
./build/tools/caffe train -solver data/pklcifar100/model/v4_ninnet/res_e1/res_e1_solver.prototxt -weights data/pklcifar100/model/v4_ninnet/pklcifar100_nin_train_iter_120000.caffemodel_coarse 2>&1 | tee -a data/pklcifar100/model/v4_ninnet/res_e1/pklcifar100_nin_res_e1_log.txt

The pretrained models are available at here.

Citation

Please cite the following paper if you find this useful in your research:

@InProceedings{Hou2017DualNet,
Title = {DualNet: Learn Complementary Features for Image Recognition},
Author = {Saihui Hou, Xu Liu and Zilei Wang},
Booktitle = {IEEE International Conference on Computer Vision (ICCV)},
Year = {2017}
}

dualnet's Issues

why there is only two softmax layer in prototxt

hi~

I have read your paper, but I only find two softmaxloss layer is ' res_e1_train_val.prototxt' . But there are three softmaxloss Layers in paper. May be it is due to 'iterative training' ? Thanks for your help.

Does the net is defined purely by protxt?

Hi,hshustc!

    Thanks for your great work.  Could you tell me if you use python API to define DualNet?  It seeems that no python file in your project is used to define DualNet.  I only see the protxtfile.  Hope get ur response, thanks.

Question about the architecture of resnet-20 in the paper

Does the number of output channels vary from 64->128->256->512 just like the origin ResNet paper?

训练60000次后，损失函数降不下来，结果的精度很低

你好，最近一直在读你的这篇文章，用caffe跑程序，迭代完60000次后，损失函数从最开始的7.67降到3.76左右（学习率也调整过，batchsize调整为3），训练完后，精度只有12%左右，我想请教下，怎么修改源程序参数呢？才能使损失函数收敛，精度有论文中那么高。。。期待你的答复，谢谢

Iterative training

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

where is the code of JointFinetunning?

hi hshustc,
Though i have read your code for many times, I didn't find which part is used in JointFinetunning, Could u please tell me where I can find it ?
thx for your help!

initialize of the itertive training

hello,
I want to ask you whether you use finetuned model to initialize both S1 Extractor(or include S1 Classifier) and S2 Extractor ?

thanks