fixed #20

initialize f_val in smaller batch
Xtra-Computing · Nov 5, 2017 · 468c1e8 · 468c1e8
1 parent 461dbf5
commit 468c1e8
Show file tree

Hide file tree

Showing 3 changed files with 30 additions and 12 deletions.
diff --git a/src/test/test_main.cpp b/src/test/test_main.cpp
@@ -9,6 +9,6 @@ int main(int argc, char **argv) {
     ::testing::InitGoogleTest(&argc, argv);
     el::Loggers::reconfigureAllLoggers(el::ConfigurationType::Format, "%datetime %level %fbase:%line : %msg");
     el::Loggers::addFlag(el::LoggingFlag::ColoredTerminalOutput);
-    cudaSetDevice(1);
+    cudaSetDevice(0);
     return RUN_ALL_TESTS();
 }
diff --git a/src/thundersvm/solver/csmosolver.cu b/src/thundersvm/solver/csmosolver.cu
@@ -132,24 +132,27 @@ CSMOSolver::calculate_rho(const SyncData <real> &f_val, const SyncData<int> &y,
 
 void CSMOSolver::init_f(const SyncData<real> &alpha, const SyncData<int> &y, const KernelMatrix &k_mat,
                         SyncData<real> &f_val) const {
-    //TODO initialize with smaller batch to reduce memory usage
+    //todo auto set batch size
+    int batch_size = 100;
     vector<int> idx_vec;
     vector<real> alpha_diff_vec;
     for (int i = 0; i < alpha.size(); ++i) {
         if (alpha[i] != 0) {
             idx_vec.push_back(i);
             alpha_diff_vec.push_back(-alpha[i] * y[i]);
         }
-    }
-    if (idx_vec.size() > 0) {
-        SyncData<int> idx(idx_vec.size());
-        SyncData<real> alpha_diff(idx_vec.size());
-        idx.copy_from(idx_vec.data(), idx_vec.size());
-        alpha_diff.copy_from(alpha_diff_vec.data(), idx_vec.size());
-        SyncData<real> kernel_rows(idx.size() * k_mat.n_instances());
-        k_mat.get_rows(idx, kernel_rows);
-        SAFE_KERNEL_LAUNCH(update_f, f_val.device_data(), idx.size(), alpha_diff.device_data(),
-                           kernel_rows.device_data(), k_mat.n_instances());
+        if (idx_vec.size() > batch_size || (i == alpha.size() - 1 && idx_vec.size() > 0)) {
+            SyncData<int> idx(idx_vec.size());
+            SyncData<real> alpha_diff(idx_vec.size());
+            idx.copy_from(idx_vec.data(), idx_vec.size());
+            alpha_diff.copy_from(alpha_diff_vec.data(), idx_vec.size());
+            SyncData<real> kernel_rows(idx.size() * k_mat.n_instances());
+            k_mat.get_rows(idx, kernel_rows);
+            SAFE_KERNEL_LAUNCH(update_f, f_val.device_data(), idx.size(), alpha_diff.device_data(),
+                               kernel_rows.device_data(), k_mat.n_instances());
+            idx_vec.clear();
+            alpha_diff_vec.clear();
+        }
     }
 }
 

diff --git a/src/thundersvm/thundersvm-train.cpp b/src/thundersvm/thundersvm-train.cpp
@@ -39,6 +39,21 @@ int main(int argc, char **argv) {
             break;
     }
 
+    //todo add this to check_parameter method
+    if (parser.param_cmd.svm_type == SvmParam::NU_SVC) {
+        train_dataset.group_classes();
+        for (int i = 0; i < train_dataset.n_classes(); ++i) {
+            int n1 = train_dataset.count()[i];
+            for (int j = i + 1; j < train_dataset.n_classes(); ++j) {
+                int n2 = train_dataset.count()[j];
+                if (parser.param_cmd.nu * (n1 + n2) / 2 > min(n1, n2)) {
+                    printf("specified nu is infeasible\n");
+                    return 1;
+                }
+            }
+        }
+    }
+
     CUDA_CHECK(cudaSetDevice(parser.gpu_id));
 
     if (parser.do_cross_validation) {