shentqlf / softmaxcpp Goto Github PK

The modules here relate to the softmax classifier model in my mini-project report "Linear classification models applied to the classification of hand-written digits", submitted towards the degree of MPhil in Scientific Computing and Machine Learning, University of Cambridge.

The MNIST data files should be downloaded from http://cis.jhu.edu/~sachin/digit/digit.html and organised by digit into a folder MNISTdata, with MNISTdata/data0 corresponding to digit 0 etc. all the way up to digit 9.

The project uses the c++ armadillo linear algebra library. The included Makefile has the required linker flags to this library for running the code on the LSC computers. 

Before compiling:

Set the tunable parameters in main.C to generate results for a specific test case. These parameters have been labelled with self-explanatory names.

To compile the program:

make Softmax 

To run it:

./Softmax

Numerical gradient checking can be performed by toggling on/off the DEBUG_GRADIENTS macro in Softmax.C. 

The file plotSoftmax.py can be used to generate all the plots. Note that output data files for the plotting of learning curves and validation curves will only be produced if the TUNING macro in main.C is set to true.