Skip to content
/ deep_lda Public

Implementation of Deep Linear Discriminant Analysis (DeepLDA)

Notifications You must be signed in to change notification settings

CPJKU/deep_lda

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Deep Linear Discriminant Analysis (DeepLDA)

This repository contains code for reproducing the experiments reported in the ICLR 2016 paper Deep Linear Discriminant Analysis by Matthias Dorfer, Rainer Kelz and Gerhard Widmer from the Department of Computational Perception at JKU Linz.

Requirements

The implementation is based on Theano and the neural networks library Lasagne. For installing Theano and Lasagne please follow the installation instruction on the respective github pages.

You will also need: matplotlib, numpy and scipy

Experiments

We report results for three different benchmark datasets in our paper. The datasets will be downloaded automatically form the corresponding data set pages.

Training

For training the models just run the following commands:

MNIST: the model should train up to a validation accuracy of around 99.7%.:

python exp_dlda.py --model mnist_dlda --data mnist --train

CIFAR10: the model should train up to 91.5% validation accuracy. This is a little bit less than the results reported in the paper as no GCN-ZCA normalization is applied:

python exp_dlda.py --model cifar10_dlda --data cifar10 --train

Train on the first fold of STL10:

python exp_dlda.py --model stl10_dlda --data stl10 --train --fold 0
  • the model should train up to a validation accuracy of around 67% on fold 0.
  • you can of course train on the remaining folds by changing the --fold argument
  • we train this model on NVIDIA Tesla K40 (12GB memory) so it might be to large for less powerful GPUs (in this case you could try to reduce the batch size).

Evaluation

For evaluating the trained models run the following commands. The script will report:

  • The accuracies on train, validation and test set
  • Report the magnitudes of the individual eigenvalues after solving the general (Deep)LDA eigenvalue problem compare Figure 5 in paper)
  • Produce some plots visualizing the structure of the latent representation produced by the model (compare Figure 5 in paper)
python exp_dlda.py --model mnist_dlda --data mnist --eval
python exp_dlda.py --model cifar10_dlda --data cifar10 --eval
python exp_dlda.py --model stl10_dlda --data stl10 --eval --fold 0

About

Implementation of Deep Linear Discriminant Analysis (DeepLDA)

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages