Contents
Description of demo_multiclass_CNN.m
Comparison of multiclass classification using multiclass logistic regression and simplest possible CNN with one convolution and one mean pooling layer feeding into softmax
Note: implementations have not yet been parallelized and do not make use of GPUs in order to keep the algorithm easy to understand and extend. Consequently, this demo will take significant time to run.
% load data from MNIST
loadMNISTDataset
Decompressing MNIST files...Done. Loading MNIST files into Workspace...Ready to train model
usage of multi-class logistic classification (MNIST data)
options_lg = []; options_lg.addBias = 1; model_lg = ml_multiclass_logistic(Xtrain, ytrain, options_lg); yhat_lg = model_lg.predict(model_lg, Xtest); testError_lg = mean(yhat_lg ~= ytest); fprintf('Averaged misclassification test error with %s is: %.3f\n', ... model_lg.name, testError_lg);
Averaged misclassification test error with Multiclass Logistic Classification is: 0.079
usage of multi-class CNN classification (MNIST data)
options_cnn.imageDim = 28; options_cnn.nClasses = 10; options_cnn.filterDim = 9; % Filter size for conv layer options_cnn.nFilters = 20; % Number of filters for conv layer options_cnn.poolDim = 2; model_cnn = ml_multiclass_CNN(Xtrain, ytrain, options_cnn); yhat_cnn = model_cnn.predict(model_cnn, Xtest); testError_cnn = mean(yhat_cnn ~= ytest); fprintf('Averaged misclassification test error with %s is: %.3f\n', ... model_cnn.name, testError_cnn);
Training example: 1 Training example: 10001 Training example: 20001 Training example: 30001 Training example: 40001 Training example: 50001 Training example: 60001 Training example: 70001 Training example: 80001 Training example: 90001 Averaged misclassification test error with CNN for Multiclass Image Classification is: 0.020