Codes for SJTU CS385 homework based on MNIST dataset.
-sub: load a subset, still 10-category--traintest 10000 5000: 10000, 5000 images respectively in trainset and testset-1vsr: load one-vs-rest, binary classes--loadone 1: 1-vs-rest-1vs1: load one-vs-one, binary classes--loadtwo 3 5: 3-vs-5- else: load full dataset
python logistic.py --loss LASSO -lam 0.01: logistic regression using intercept term, full dataset (10-class cross entropy), LASSO loss with regularizatation penalty lambda=0.01.python logistic.py --kernel RBF -1vs1 --loadtwo 0 1: logistic regression using RBF kernel, binary 0&1 dataset, RBF kernel. On the first run krenel matrix will be precomputed and saved for more runs.
Weights are plotted.
python LDA.py --trainset 10000 5000: LDA on 10-categories, load 10000 train images and 5000 test images.python LDA.py -1vsr --loadone 1: LDA on binary digits 1-vs-rest.
Threshold for binary classification, bivariate Gaussian distributions and projected samples for multi-classification are plotted. Repeated trials shows LDA's sensitivity on datasets.
python SVM.py --kernel poly: SVM with polynomial kernel, on full dataset.- Or PCA before SVM. See codes for details.
Confusion matrix, support vectors and decision boundary are plotted.
