The matlab code for the IJCAI-16 paper "Self-Paced Boost Learning for Classification"
-
Store the data as "Data/fea.mat" (in the "Data/" directory), which has two matrices: fea: n-by-d matrix, where each row is the feature of a sample; gnd: n-by-1 vector, where each element is the class label index of a sample.
-
Run "Gen_Split.m" to generate the split of traning/validation/test set of the data. Set the "train_ratio" and "vali_ratio" variables for the proportions of the training and the validation samples, respectively.
-
(Optional) Run "Gen_Noise" to generate label noise in the training set. Set the "n_ratio" variable for the proportion of the noisily labeled samples.
-
Run "SPBLmain.m" to train the SPBL model with the training set and test the learned model on the test set. The output experimental results: ResObj.test_err: test error rate at each iteration, with the first (second) column as the top-1 (top-5) error rates. ResObj.vali_err: validation error rate at each iteration, with the first (second) column as the top-1 (top-5) error rates. OptRes: struct varaible. The optimal results on the test set based on the validation.