Make matrix well-conditioned when unittest inverse

paddle/math/Matrix.cpp

@@ -202,6 +202,17 @@ void GpuMatrix::resetOne() {
   CHECK(data_ != NULL);
   one();
 }
+
+void GpuMatrix::setDiag(real value) {
+  CHECK(data_ != NULL);
+  CHECK_EQ(height_, width_);
+
+  zeroMem();
+  for (size_t i = 0; i < height_; i++) {
+    hl_memcpy_host2device(&data_[i * stride_ + i], &value, sizeof(real));
+  }
+}
+
 void GpuMatrix::resize(size_t newHeight, size_t newWidth) {
   size_t newSize = newHeight * newWidth;
   if (NULL == memoryHandle_.get() ||
@@ -1244,6 +1255,16 @@ void CpuMatrix::resetOne() {
   BaseMatrix::one();
 }
 
+void CpuMatrix::setDiag(real value) {
+  CHECK(data_ != NULL);
+  CHECK_EQ(height_, width_);
+
+  zeroMem();
+  for (size_t i = 0; i < height_; i++) {
+    data_[i * stride_ + i] = value;
+  }
+}
+
 void CpuMatrix::copyFrom(const Matrix& src) {
   CHECK(isContiguous());
   if (typeid(src) == typeid(GpuMatrix)) {

paddle/math/Matrix.h

@@ -195,6 +195,8 @@ public:
 
   virtual void resetOne() { LOG(FATAL) << "Not implemented"; }
 
+  virtual void setDiag(real value) { LOG(FATAL) << "Not implemented"; }
+
   virtual void copyFrom(const Matrix& src) { LOG(FATAL) << "Not implemented"; }
 
   virtual void trimFrom(const CpuSparseMatrix& src) {
@@ -330,6 +332,7 @@ public:
 
   virtual MatrixPtr getInverse() {
     LOG(FATAL) << "Not implemented";
+    return nullptr;
   }
 
   /**
@@ -1016,6 +1019,7 @@ public:
 
   void zeroMem();
   void resetOne();
+  void setDiag(real value);
 
   void resize(size_t newHeight, size_t newWidth);
   void resize(size_t newHeight, size_t newWidth,
@@ -1280,6 +1284,8 @@ public:
 
   void zeroMem();
   void resetOne();
+  void setDiag(real value);
+
   void resize(size_t newHeight, size_t newWidth);
   void resize(size_t newHeight, size_t newWidth,
               size_t newNnz, /* used to allocate space */

paddle/math/tests/test_matrixCompare.cpp

@@ -647,20 +647,23 @@ void testMatrixInverse(int height) {
   MatrixPtr cpuI = std::make_shared<CpuMatrix>(height, height);
   MatrixPtr gpuI = std::make_shared<GpuMatrix>(height, height);
 
+  /* Make matrix well conditioned: cpu * cpuT + Identity */
   cpu->randomizeUniform();
+  MatrixPtr cpuT = cpu->getTranspose();
+  MatrixPtr outputCheck = std::make_shared<CpuMatrix>(height, height);
+  outputCheck->mul(cpu, cpuT);
+  cpu->setDiag(1.0);
+  cpu->add(*outputCheck);
+
   gpu->copyFrom(*cpu);
   cpu->inverse(cpuI, false);
   gpu->inverse(gpuI, false);
 
-  MatrixPtr outputCheck = std::make_shared<CpuMatrix>(height, height);
   outputCheck->copyFrom(*gpuI);
   MatrixCheckErr(*cpuI, *outputCheck);
 
   outputCheck->mul(cpu, cpuI);
-  cpu->zeroMem();
-  for (int i = 0; i < height; i++) {
-    cpu->getRowBuf(i)[i] = 1.0;
-  }
+  cpu->setDiag(1.0);
   MatrixCheckErr(*cpu, *outputCheck);
 }