Update toy example script

toy_example/run_toy_example.sh

@@ -1,20 +1,35 @@
 #!/bin/bash
-
+################################################################################
+# svm-gsu: toy example                                                         #
+# This is a toy example ...                                                                             #
+#                                                                              #
+# [1] C. Tzelepis, V. Mezaris, I. Patras, "Linear Maximum Margin Classifier    #
+# for Learning from Uncertain Data", IEEE Transactions on Pattern Analysis and #
+# Machine Intelligence, accepted for publication.                              #
+# DOI:10.1109/TPAMI.2017.2772235.                                              #
+################################################################################
 # Define experiment's parameters
-verb=1
-cov_mat="diag"
-kernel=0
-lambda=100
-T=3000
-k=5000
-#k=1000
-p=1.00
+verb=1            # Set verbose mode (On: 1 / Off: 0)
+cov_mat="diag"    # Choose covariance matrices type:
+                  # -- "full": Full covariance matrices
+                  # -- "diag": Diagonal covariance matrices
+                  # -- "iso": Isotropic covariance matrices
+kernel=0          # Choose kernel type:
+                  # -- 0: linear kernel
+                  # -- 2: RBF kernel
+lambda=0.001      # Set SVM-GSU's reguralization (lambda) parameter
+T=3000            # Set number of iterations for the SGD algorithm
+k=100             # Set the sampling size for the SGD algorithm (1<=k<=T)
+p=1.00            # Set the fraction of the total variance preserved for each
+                  # input Gaussian -- 0.0<p<=1.0 (learning in linear subspaces
+                  # -- See Sect. 3.2 of [1])
 
 # Define input/output directories
 data_dir="./data"
 models_dir="./models"
 output_dir="./output"
-bin_dir=../build/gsvm-train
+bin_dir_train=../build/gsvm-train
+bin_dir_predict=../build/gsvm-predict
 
 if [[ ${kernel} -eq 0 ]];
 then
@@ -27,34 +42,49 @@ else
     exit;
 fi
 
+if [ "${cov_mat}" == "full" ]
+then
+  cov_mat_opt=0
+elif [ "${cov_mat}" == "diag" ]
+then
+  cov_mat_opt=1
+elif [ "${cov_mat}" == "iso" ]
+then
+  cov_mat_opt=2
+fi
+
 # Training files
 train_mean_vec="${data_dir}/mean.dat"
 train_labels="${data_dir}/labels.dat"
 train_cov_mat="${data_dir}/cov_${cov_mat}.dat"
 train_model_file="${models_dir}/svmgsu_${kernel_type}_${cov_mat}_T${T}_k${k}_l${lambda}.model"
 
-# Testing files
+# Testing files (use )
 test_mean_vec="${data_dir}/mean.dat"
 test_labels="${data_dir}/labels.dat"
 test_cov_mat="${data_dir}/cov_${cov_mat}.dat"
 test_output_file="${output_dir}/svmgsu_${kernel_type}_${cov_mat}_T${T}_k${k}_l${lambda}.output"
+test_metrics_file="${output_dir}/svmgsu_${kernel_type}_${cov_mat}_T${T}_k${k}_l${lambda}.metrics"
 
 # Construct training options
-if [ "${cov_mat}" == "full" ]
-then
-  cov_mat_opt=0
-elif [ "${cov_mat}" == "diag" ]
-then
-  cov_mat_opt=1
-elif [ "${cov_mat}" == "iso" ]
-then
-  cov_mat_opt=2
-fi
+
 opt="-v ${verb} -t ${kernel} -d ${cov_mat_opt} -l ${lambda} -T ${T} -k ${k} -p ${p}"
 
 # Construct training command
-train_cmd="${bin_dir}/gsvm-train ${opt} ${train_mean_vec} ${train_labels} ${train_cov_mat} ${train_model_file}"
+train_cmd="${bin_dir_train}/gsvm-train ${opt} ${train_mean_vec} ${train_labels} ${train_cov_mat} ${train_model_file}"
 
 # Train SVM-GSU
-echo "Training a ${kernel_type} SVM-GSU..."
+echo -n "Re-build gsvm-train..."
+#make clean -C ../build/gsvm-train
+make -s -C ../build/gsvm-train && echo "Done!"
+echo "Run gsvm-train..."
 eval ${train_cmd}
+
+# Evaluate the above trained SVM-GSU model
+echo -n "Re-build gsvm-predict..."
+#make clean -C ../build/gsvm-predict
+make -s -C ../build/gsvm-predict && echo "Done!"
+predict_opt="-v ${verb} -t ${test_labels} -m ${test_metrics_file}"
+predict_cmd="${bin_dir_predict}/gsvm-predict ${predict_opt} ${train_mean_vec} ${train_model_file} ${test_output_file}"
+echo "Run gsvm-predict..."
+eval ${predict_cmd}