Merge pull request #4 from rapidsai/bug-tsvd-explainedvar

Bug tsvd explained_var

README.md

@@ -7,18 +7,18 @@ The cuML repository contains:
 1. ***python***: Python based GPU Dataframe (GDF) machine learning package that takes [cuDF](https://github.com/rapidsai/cudf-alpha) dataframes as input. cuML connects the data to C++/CUDA based cuML and ml-prims libraries without ever leaving GPU memory.
 
 2. ***cuML***: C++/CUDA machine learning algorithms. This library currently includes the following five algorithms;
-   a. Single GPU Truncated Singular Value Decomposition (tSVD).
-   b. Single GPU Principal Component Analysis (PCA).
-   c. Single GPU Density-based Spatial Clustering of Applications with Noise (DBSCAN).
-   d. Single GPU Kalman Filtering
-   e. Multi-GPU K-Means Clustering
+   a. Single GPU Truncated Singular Value Decomposition (tSVD),
+   b. Single GPU Principal Component Analysis (PCA),
+   c. Single GPU Density-based Spatial Clustering of Applications with Noise (DBSCAN),
+   d. Single GPU Kalman Filtering,
+   e. Multi-GPU K-Means Clustering.
 
 3. ***ml-prims***: Low level machine learning primitives used in cuML. ml-prims is comprised of the following components;
-   a. Linear Algebra
-   b. Statistics
-   c. Basic Matrix Operations
-   d. Distance Functions
-   e. Random Number Generation
+   a. Linear Algebra,
+   b. Statistics,
+   c. Basic Matrix Operations,
+   d. Distance Functions,
+   e. Random Number Generation.
 
 #### Available Algorithms for version 0.1alpha:
 
@@ -30,7 +30,7 @@ The cuML repository contains:
 
 Upcoming algorithms for version 0.1:
 
-- kmeans
+- K-Means Clustering
 
 - Kalman Filter
 
@@ -49,7 +49,7 @@ List of dependencies:
 
 1. zlib
 2. cmake (>= 3.8, version 3.11.4 is recommended and there are issues with version 3.12)
-3. CUDA Toolkit (>= 9.0)
+3. CUDA (>= 9.0)
 4. Cython (>= 0.28)
 5. gcc (>=5.4.0)
 6. [cuDF](https://github.com/rapidsai/cudf-alpha)
@@ -65,27 +65,27 @@ git clone --recurse-submodules https://github.com/rapidsai/cuml.git
 To build the python package, in the repository root folder:
 
 ```
-cd python
 python setup.py install
 ```
 
 ### Building CuML:
 
 ### Running tests
 
-To test the C++ algorithms using googletests, without compiling the
+To test the C++ algorithms using googletests, in the repository root folder:
 
 ```bash
+$ cd cuML
 $ mkdir build
 $ cd build
 $ cmake ..
 $ make -j
 $ ./ml_test
 ```
 
-### Python Tests
+### Python Notebooks
 
-Additional python tests can be found in the pythontests folder, along some useful scripts. Py.test based unit testing are being added for the final version 0.1.
+Demo notebooks can be found in python/notebooks folder.
 
 ## External
 

cuML/src/pca/pca.h

@@ -33,8 +33,39 @@ namespace ML {
 
 using namespace MLCommon;
 
-// TODO: Implement for the case prms.trans_input = false
-// TODO: Check sign flip algorithm
+template<typename math_t>
+void truncCompExpVars(math_t *in, math_t *components, math_t *explained_var,
+		math_t *explained_var_ratio, paramsTSVD prms,
+		cusolverDnHandle_t cusolver_handle, cublasHandle_t cublas_handle) {
+
+	math_t *components_all;
+	math_t *explained_var_all;
+	math_t *explained_var_ratio_all;
+
+	int len = prms.n_cols * prms.n_cols;
+	allocate(components_all, len);
+	allocate(explained_var_all, prms.n_cols);
+	allocate(explained_var_ratio_all, prms.n_cols);
+
+	calEig(in, components_all, explained_var_all, prms, cusolver_handle,
+			cublas_handle);
+
+	Matrix::truncZeroOrigin(components_all, prms.n_cols, components,
+			prms.n_components, prms.n_cols);
+
+	Matrix::ratio(explained_var_all, explained_var_ratio_all, prms.n_cols);
+
+	Matrix::truncZeroOrigin(explained_var_all, prms.n_cols, explained_var,
+			prms.n_components, 1);
+
+	Matrix::truncZeroOrigin(explained_var_ratio_all, prms.n_cols,
+			explained_var_ratio, prms.n_components, 1);
+
+	CUDA_CHECK(cudaFree(components_all));
+	CUDA_CHECK(cudaFree(explained_var_all));
+	CUDA_CHECK(cudaFree(explained_var_ratio_all));
+}
+
 /**
  * @brief perform fit operation for the pca. Generates eigenvectors, explained vars, singular vals, etc.
  * @input param input: the data is fitted to PCA. Size n_rows x n_cols. The size of the data is indicated in prms.
@@ -66,27 +97,19 @@ void pcaFit(math_t *input, math_t *components, math_t *explained_var,
 
 	Stats::mean(mu, input, prms.n_cols, prms.n_rows, true, false);
 
-	if (prms.algorithm == solver::RANDOMIZED) {
-		Stats::meanCenter(input, mu, prms.n_cols, prms.n_rows, false);
-		calCompExpVarsSvd(input, components, singular_vals, explained_var,
-				explained_var_ratio, prms, cusolver_handle, cublas_handle);
-	} else {
-		math_t *cov;
-		int len = prms.n_cols * prms.n_cols;
-		allocate(cov, len);
-
-		Stats::cov(cov, input, mu, prms.n_cols, prms.n_rows, true, false, true,
-				cublas_handle);
-		calCompExpVarsEig(cov, components, explained_var, explained_var_ratio,
-				prms, cusolver_handle, cublas_handle);
-
-		math_t scalar = (prms.n_rows - 1);
-		Matrix::seqRoot(explained_var, singular_vals, scalar,
-				prms.n_components);
-
-		if (cov)
-			CUDA_CHECK(cudaFree(cov));
-	}
+	math_t *cov;
+	int len = prms.n_cols * prms.n_cols;
+	allocate(cov, len);
+
+	Stats::cov(cov, input, mu, prms.n_cols, prms.n_rows, true, false, true,
+			cublas_handle);
+	truncCompExpVars(cov, components, explained_var, explained_var_ratio, prms,
+			cusolver_handle, cublas_handle);
+
+	math_t scalar = (prms.n_rows - 1);
+	Matrix::seqRoot(explained_var, singular_vals, scalar, prms.n_components);
+
+	CUDA_CHECK(cudaFree(cov));
 
 	Stats::meanAdd(input, mu, prms.n_cols, prms.n_rows, false);
 }
@@ -117,7 +140,8 @@ void pcaFitTransform(math_t *input, math_t *trans_input, math_t *components,
 	pcaTransform(input, components, trans_input, singular_vals, mu, prms,
 			cublas_handle);
 
-	signFlip(trans_input, prms.n_rows, prms.n_components, components, prms.n_cols);
+	signFlip(trans_input, prms.n_rows, prms.n_components, components,
+			prms.n_cols);
 }
 
 // TODO: implement pcaGetCovariance function

cuML/src/tsvd/tsvd.cu

@@ -24,31 +24,27 @@ namespace ML {
 
 using namespace MLCommon;
 
-void tsvdFit(float *input, float *components, float *explained_var,
-		float *explained_var_ratio, float *singular_vals, paramsTSVD prms) {
+void tsvdFit(float *input, float *components, float *singular_vals, paramsTSVD prms) {
 	cublasHandle_t cublas_handle;
 	CUBLAS_CHECK(cublasCreate(&cublas_handle));
 
 	cusolverDnHandle_t cusolver_handle = NULL;
 	CUSOLVER_CHECK(cusolverDnCreate(&cusolver_handle));
 
-	tsvdFit(input, components, explained_var, explained_var_ratio,
-			singular_vals, prms, cublas_handle, cusolver_handle);
+	tsvdFit(input, components, singular_vals, prms, cublas_handle, cusolver_handle);
 
 	CUBLAS_CHECK(cublasDestroy(cublas_handle));
 	CUSOLVER_CHECK(cusolverDnDestroy(cusolver_handle));
 }
 
-void tsvdFit(double *input, double *components, double *explained_var,
-		double *explained_var_ratio, double *singular_vals, paramsTSVD prms) {
+void tsvdFit(double *input, double *components, double *singular_vals, paramsTSVD prms) {
 	cublasHandle_t cublas_handle;
 	CUBLAS_CHECK(cublasCreate(&cublas_handle));
 
 	cusolverDnHandle_t cusolver_handle = NULL;
 	CUSOLVER_CHECK(cusolverDnCreate(&cusolver_handle));
 
-	tsvdFit(input, components, explained_var, explained_var_ratio,
-			singular_vals, prms, cublas_handle, cusolver_handle);
+	tsvdFit(input, components, singular_vals, prms, cublas_handle, cusolver_handle);
 
 	CUBLAS_CHECK(cublasDestroy(cublas_handle));
 	CUSOLVER_CHECK(cusolverDnDestroy(cusolver_handle));