The code is pre-compiled in 64-bit Linux (ubuntu 12.04 + NVIDIA driver 331.20 + cuda 5.5)please follow the installation instructions on caffe's page.

Run ./train.sh to reproduce the experiment.
The result will be shown in standard stream and recorded in .log files specified in train.sh
Before running the experiment please edit the corresponding .solver file to modify the GPU ID you want to use.
By default, I output some DEBUG information (weight/gradient/error magnitude and variance), you can also modify the output frequency in the solver file (info blocks).

I might have made mistakes in wrting the solver and script file when cleaning the code. contact [email protected] if you have any trouble in running the experiment.