yyht/apdlm

Training Deep Neural Networks via Direct Loss Minimization

Training Deep Neural Networks via Direct Loss Minimization

This is the code for our ICML'16 paper:

• Yang Song, Alexander G. Schwing, Richard S. Zemel, Raquel Urtasun, Training Deep Neural Networks via Direct Loss Minimization, International Conference on Machine Learning, New York City, USA, 2016.

Please cite the above paper if you use our code. For simplicity, we only include the code for action classification. We basically use the same code for object detection with minor modifications.

Dependencies

Our codes depend on qmake, CUDA, MPICH, CImg, LibLSDN (see the following section) and some MATLAB codes from PASCAL VOC2012 development toolkits. It is compulsory to have these downloaded (and installed).

The main source code was written in C++11.

LibLSDN

Our code is based on a deep learning library LibLSDN, which is written and maintained by Alexander G. Schwing. We include part of that library in our repository, including CPrecisionTimer.h, LSDN_Common.h, LSDN_CudaCommon.h and all files in LibLSDN/.

In order to install this library, please first modify the paths accordingly in Makefile and then run

make LibLSDN.a

Data

We use the PASCAL VOC2012 dataset, which can be downloaded at the website. The dataset contains a VOCcode/ folder, which contains a lot of utility MATLAB codes helpful for preprocessing data.

After extracting VOC2012 dataset, you should put preprocessing/save_action_split.m under the VOCdevkit2012/ folder and run it to get preprocessed data files. They have the format of *_info.dat and *_ims.dat.

Training and Validation / Testing

Previous to running the program, please modify the paths and parameters in src/APDLM.pro and src/Conf.cpp according to your own settings.

First run the following commands

qmake
make

to compile the program.

Then run

mpiexec -n 2 ./APDLM [alpha] [beta] [epsilon] [class_name] [gpu1] [gpu2]

Here alpha is the learning rate, beta is the L2 regularization coefficient, epsilon is the sharpness parameter as defined in the paper, class_name is the name of the class, gpu1 and gpu2 are indices of GPUs. You can also change the number of GPUs.

Notes

The AlexNet weight AlexNetWeightsSingleDimensionOutput.dat used for finetuning is too large for GitHub. Please contact the authors for it.