fix precision issue #126

Xtra-Computing · Feb 18, 2019 · 7b6b743 · 7b6b743
1 parent a824a1a
commit 7b6b743
Show file tree

Hide file tree

Showing 8 changed files with 33 additions and 3 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -19,6 +19,10 @@ endif()
 set(USE_CUDA ON CACHE BOOL "Compile with CUDA")
 set(USE_EIGEN OFF CACHE BOOL "Compile with Eigen")
 set(BUILD_TESTS OFF CACHE BOOL "Build Tests")
+set(USE_DOUBLE OFF CACHE BOOL "Use double as kernel_type")
+if(USE_DOUBLE)
+    message("Use double as kernel_type")
+endif()
 if (NOT CMAKE_BUILD_TYPE)
     set(CMAKE_BUILD_TYPE Release)
 endif ()

diff --git a/docs/faq.md b/docs/faq.md
@@ -3,6 +3,9 @@ Frequently Asked Questions (FAQs)
 This page is dedicated to summarizing some frequently asked questions about ThunderSVM.
 
 ## FAQs of users
+* **Why does ThunderSVM have large errors? My data set is not normalized.** 
+   To reduce the errors, you need to add ``-DUSE_DOUBLE=ON`` in cmake (e.g., ``cmake -DUSE_DOUBLE=ON ..``). The reason is that ThunderSVM uses ``float`` to store kernel values for better efficiency. For problems require high precision, you need to tell ThunderSVM to use ``double``.
+
 * **How can I use the source code?** 
    Please refere to [How To](how-to.md) page.
 

diff --git a/include/thundersvm/config.h.in b/include/thundersvm/config.h.in
@@ -1,3 +1,4 @@
 #cmakedefine DATASET_DIR "@DATASET_DIR@"
 #cmakedefine USE_CUDA
 #cmakedefine USE_EIGEN
+#cmakedefine USE_DOUBLE
diff --git a/include/thundersvm/thundersvm.h b/include/thundersvm/thundersvm.h
@@ -14,5 +14,10 @@
 using std::string;
 using std::vector;
 typedef double float_type;
+
+#ifdef USE_DOUBLE
+typedef double kernel_type;
+#else
 typedef float kernel_type;
+#endif
 #endif //THUNDERSVM_THUNDERSVM_H
diff --git a/src/thundersvm/cmdparser.cpp b/src/thundersvm/cmdparser.cpp
@@ -184,6 +184,9 @@ void CMDParser::parse_command_line(int argc, char **argv) {
                 case 'u':
                     gpu_id = atoi(argv[i]);
                     break;
+                case 'o':
+                    n_cores = atoi(argv[i]);
+                    break;
                 case 'm':
                     param_cmd.max_mem_size = static_cast<size_t>(max(atoi(argv[i]), 0)) << 20;//MB to Byte
                     break;
@@ -192,6 +195,11 @@ void CMDParser::parse_command_line(int argc, char **argv) {
                     HelpInfo_svmpredict();
             }
         }
+        if (n_cores > 0) {
+            omp_set_num_threads(n_cores);
+        } else if (n_cores != -1) {
+            LOG(ERROR) << "the number of cpu cores must be positive or -1";
+        }
         if (i >= argc - 2)
             HelpInfo_svmpredict();
         svmpredict_input_file = argv[i];

diff --git a/src/thundersvm/kernel/kernelmatrix_kernel.cpp b/src/thundersvm/kernel/kernelmatrix_kernel.cpp
@@ -149,7 +149,7 @@ namespace svm_kernel {
     void dns_csr_mul(int m, int n, int k, const SyncArray<kernel_type> &dense_mat, const SyncArray<kernel_type> &csr_val,
                      const SyncArray<int> &csr_row_ptr, const SyncArray<int> &csr_col_ind, int nnz,
                      SyncArray<kernel_type> &result) {
-        Eigen::Map<const Eigen::MatrixXf> denseMat(dense_mat.host_data(), k, n);
+        Eigen::Map<const Eigen::Matrix<kernel_type, Eigen::Dynamic, Eigen::Dynamic, Eigen::ColMajor>> denseMat(dense_mat.host_data(), k, n);
         Eigen::Map<const Eigen::SparseMatrix<kernel_type, Eigen::RowMajor>> sparseMat(m, k, nnz, csr_row_ptr.host_data(),
                                                                                 csr_col_ind.host_data(),
                                                                                 csr_val.host_data());
@@ -177,7 +177,7 @@ namespace svm_kernel {
 
     void dns_dns_mul(int m, int n, int k, const SyncArray<kernel_type> &dense_mat,
                      const SyncArray<kernel_type> &origin_dense, SyncArray<kernel_type> &result){
-        Eigen::Map<const Eigen::MatrixXf> denseMat(dense_mat.host_data(), k, n);
+        Eigen::Map<const Eigen::Matrix<kernel_type, Eigen::Dynamic, Eigen::Dynamic, Eigen::ColMajor>> denseMat(dense_mat.host_data(), k, n);
         Eigen::Map<const Eigen::Matrix<kernel_type, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>>
                 originDenseMat(origin_dense.host_data(), m, k);
         Eigen::Matrix<kernel_type, Eigen::Dynamic, Eigen::Dynamic, Eigen::ColMajor> retMat = originDenseMat * denseMat;

diff --git a/src/thundersvm/kernel/kernelmatrix_kernel.cu b/src/thundersvm/kernel/kernelmatrix_kernel.cu
@@ -144,10 +144,19 @@ namespace svm_kernel {
         }
         kernel_type one(1);
         kernel_type zero(0);
+#ifdef USE_DOUBLE
+        cusparseDcsrmm2(handle, CUSPARSE_OPERATION_NON_TRANSPOSE, CUSPARSE_OPERATION_TRANSPOSE,
+                        m, n, k, nnz, &one, descr, csr_val.device_data(), csr_row_ptr.device_data(),
+                        csr_col_ind.device_data(),
+                        dense_mat.device_data(), n, &zero, result.device_data(), m);
+#else//kernel type is float
         cusparseScsrmm2(handle, CUSPARSE_OPERATION_NON_TRANSPOSE, CUSPARSE_OPERATION_TRANSPOSE,
                         m, n, k, nnz, &one, descr, csr_val.device_data(), csr_row_ptr.device_data(),
                         csr_col_ind.device_data(),
                         dense_mat.device_data(), n, &zero, result.device_data(), m);
+#endif
+
+
         //cusparseScsrmm return row-major matrix, so no transpose is needed
     }
 }
diff --git a/src/thundersvm/model/svc.cpp b/src/thundersvm/model/svc.cpp
@@ -239,7 +239,7 @@ vector<float_type> SVC::predict_label(const SyncArray<float_type> &dec_values, i
                 }
             }
             int maxVoteClass = 0;
-            for (int i = 0; i < n_classes; ++i) {
+            for (int i = 1; i < n_classes; ++i) {
                 if (votes[i] > votes[maxVoteClass])
                     maxVoteClass = i;
             }