Updates to consistency of MNMG PCA/TSVD solvers (docs + code consolidation)

cpp/include/cuml/decomposition/params.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2018-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -77,7 +77,11 @@ class paramsPCATemplate : public paramsTSVDTemplate<enum_solver> {
 };
 
 typedef paramsTSVDTemplate<> paramsTSVD;
-
 typedef paramsPCATemplate<> paramsPCA;
 
+enum class mg_solver { COV_EIG_DQ, COV_EIG_JACOBI, QR };
+
+typedef paramsPCATemplate<mg_solver> paramsPCAMG;
+typedef paramsTSVDTemplate<mg_solver> paramsTSVDMG;
+
 };  // end namespace ML

cpp/include/cuml/decomposition/pca_mg.hpp

@@ -22,13 +22,6 @@
 #include "pca.hpp"
 
 namespace ML {
-
-enum class mg_solver { COV_EIG_DQ, COV_EIG_JACOBI, QR };
-
-typedef paramsTSVDTemplate<mg_solver> paramsTSVDMG;
-
-typedef paramsPCATemplate<mg_solver> paramsPCAMG;
-
 namespace PCA {
 namespace opg {
 

cpp/include/cuml/decomposition/tsvd_mg.hpp

@@ -42,7 +42,7 @@ void fit(raft::handle_t& handle,
          MLCommon::Matrix::floatData_t** input,
          float* components,
          float* singular_vals,
-         paramsTSVD prms,
+         paramsTSVDMG& prms,
          bool verbose = false);
 
 void fit(raft::handle_t& handle,
@@ -51,7 +51,7 @@ void fit(raft::handle_t& handle,
          MLCommon::Matrix::doubleData_t** input,
          double* components,
          double* singular_vals,
-         paramsTSVD prms,
+         paramsTSVDMG& prms,
          bool verbose = false);
 
 /**
@@ -77,7 +77,7 @@ void fit_transform(raft::handle_t& handle,
                    float* explained_var,
                    float* explained_var_ratio,
                    float* singular_vals,
-                   paramsTSVD prms,
+                   paramsTSVDMG& prms,
                    bool verbose);
 
 void fit_transform(raft::handle_t& handle,
@@ -89,7 +89,7 @@ void fit_transform(raft::handle_t& handle,
                    double* explained_var,
                    double* explained_var_ratio,
                    double* singular_vals,
-                   paramsTSVD prms,
+                   paramsTSVDMG& prms,
                    bool verbose);
 
 /**
@@ -109,7 +109,7 @@ void transform(raft::handle_t& handle,
                MLCommon::Matrix::Data<float>** input,
                float* components,
                MLCommon::Matrix::Data<float>** trans_input,
-               paramsTSVD prms,
+               paramsTSVDMG& prms,
                bool verbose);
 
 void transform(raft::handle_t& handle,
@@ -118,7 +118,7 @@ void transform(raft::handle_t& handle,
                MLCommon::Matrix::Data<double>** input,
                double* components,
                MLCommon::Matrix::Data<double>** trans_input,
-               paramsTSVD prms,
+               paramsTSVDMG& prms,
                bool verbose);
 
 /**
@@ -138,7 +138,7 @@ void inverse_transform(raft::handle_t& handle,
                        MLCommon::Matrix::Data<float>** trans_input,
                        float* components,
                        MLCommon::Matrix::Data<float>** input,
-                       paramsTSVD prms,
+                       paramsTSVDMG& prms,
                        bool verbose);
 
 void inverse_transform(raft::handle_t& handle,
@@ -147,7 +147,7 @@ void inverse_transform(raft::handle_t& handle,
                        MLCommon::Matrix::Data<double>** trans_input,
                        double* components,
                        MLCommon::Matrix::Data<double>** input,
-                       paramsTSVD prms,
+                       paramsTSVDMG& prms,
                        bool verbose);
 
 };  // end namespace opg

cpp/src/tsvd/tsvd_mg.cu

@@ -46,7 +46,7 @@ void fit_impl(raft::handle_t& handle,
               Matrix::PartDescriptor& input_desc,
               T* components,
               T* singular_vals,
-              paramsTSVD prms,
+              paramsTSVDMG& prms,
               cudaStream_t* streams,
               std::uint32_t n_streams,
               bool verbose)
@@ -93,7 +93,7 @@ void fit_impl(raft::handle_t& handle,
               Matrix::Data<T>** input,
               T* components,
               T* singular_vals,
-              paramsTSVD prms,
+              paramsTSVDMG& prms,
               bool verbose)
 {
   int rank = handle.get_comms().get_rank();
@@ -128,7 +128,7 @@ void transform_impl(raft::handle_t& handle,
                     Matrix::PartDescriptor input_desc,
                     T* components,
                     std::vector<Matrix::Data<T>*>& trans_input,
-                    paramsTSVD prms,
+                    paramsTSVDMG& prms,
                     cudaStream_t* streams,
                     std::uint32_t n_streams,
                     bool verbose)
@@ -180,7 +180,7 @@ void transform_impl(raft::handle_t& handle,
                     Matrix::Data<T>** input,
                     T* components,
                     Matrix::Data<T>** trans_input,
-                    paramsTSVD prms,
+                    paramsTSVDMG& prms,
                     bool verbose)
 {
   int rank = handle.get_comms().get_rank();
@@ -215,7 +215,7 @@ void inverse_transform_impl(raft::handle_t& handle,
                             Matrix::PartDescriptor trans_input_desc,
                             T* components,
                             std::vector<Matrix::Data<T>*>& input,
-                            paramsTSVD prms,
+                            paramsTSVDMG& prms,
                             cudaStream_t* streams,
                             std::uint32_t n_streams,
                             bool verbose)
@@ -265,7 +265,7 @@ void inverse_transform_impl(raft::handle_t& handle,
                             Matrix::Data<T>** trans_input,
                             T* components,
                             Matrix::Data<T>** input,
-                            paramsTSVD prms,
+                            paramsTSVDMG& prms,
                             bool verbose)
 {
   int rank = handle.get_comms().get_rank();
@@ -318,7 +318,7 @@ void fit_transform_impl(raft::handle_t& handle,
                         T* explained_var,
                         T* explained_var_ratio,
                         T* singular_vals,
-                        paramsTSVD prms,
+                        paramsTSVDMG& prms,
                         bool verbose)
 {
   int rank = handle.get_comms().get_rank();
@@ -386,7 +386,7 @@ void fit(raft::handle_t& handle,
          Matrix::floatData_t** input,
          float* components,
          float* singular_vals,
-         paramsTSVD prms,
+         paramsTSVDMG& prms,
          bool verbose)
 {
   fit_impl(handle, rank_sizes, n_parts, input, components, singular_vals, prms, verbose);
@@ -398,7 +398,7 @@ void fit(raft::handle_t& handle,
          Matrix::doubleData_t** input,
          double* components,
          double* singular_vals,
-         paramsTSVD prms,
+         paramsTSVDMG& prms,
          bool verbose)
 {
   fit_impl(handle, rank_sizes, n_parts, input, components, singular_vals, prms, verbose);
@@ -413,7 +413,7 @@ void fit_transform(raft::handle_t& handle,
                    float* explained_var,
                    float* explained_var_ratio,
                    float* singular_vals,
-                   paramsTSVD prms,
+                   paramsTSVDMG& prms,
                    bool verbose)
 {
   fit_transform_impl(handle,
@@ -438,7 +438,7 @@ void fit_transform(raft::handle_t& handle,
                    double* explained_var,
                    double* explained_var_ratio,
                    double* singular_vals,
-                   paramsTSVD prms,
+                   paramsTSVDMG& prms,
                    bool verbose)
 {
   fit_transform_impl(handle,
@@ -460,7 +460,7 @@ void transform(raft::handle_t& handle,
                Matrix::Data<float>** input,
                float* components,
                Matrix::Data<float>** trans_input,
-               paramsTSVD prms,
+               paramsTSVDMG& prms,
                bool verbose)
 {
   transform_impl(handle, rank_sizes, n_parts, input, components, trans_input, prms, verbose);
@@ -472,7 +472,7 @@ void transform(raft::handle_t& handle,
                Matrix::Data<double>** input,
                double* components,
                Matrix::Data<double>** trans_input,
-               paramsTSVD prms,
+               paramsTSVDMG& prms,
                bool verbose)
 {
   transform_impl(handle, rank_sizes, n_parts, input, components, trans_input, prms, verbose);
@@ -484,7 +484,7 @@ void inverse_transform(raft::handle_t& handle,
                        Matrix::Data<float>** trans_input,
                        float* components,
                        Matrix::Data<float>** input,
-                       paramsTSVD prms,
+                       paramsTSVDMG prms,
                        bool verbose)
 {
   inverse_transform_impl(
@@ -497,7 +497,7 @@ void inverse_transform(raft::handle_t& handle,
                        Matrix::Data<double>** trans_input,
                        double* components,
                        Matrix::Data<double>** input,
-                       paramsTSVD prms,
+                       paramsTSVDMG prms,
                        bool verbose)
 {
   inverse_transform_impl(

python/cuml/dask/decomposition/pca.py

@@ -1,4 +1,4 @@
-# Copyright (c) 2019-2021, NVIDIA CORPORATION.
+# Copyright (c) 2019-2022, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -32,9 +32,11 @@ class PCA(BaseDecomposition,
     the data. N_components is usually small, say at 3, where it can be used for
     data visualization, data compression and exploratory analysis.
 
-    cuML's multi-node multi-gpu (MNMG) PCA expects a dask cuDF input, and
-    provides a "Full" algorithm. It uses a full eigendecomposition
-    then selects the top K eigenvectors.
+    cuML's multi-node multi-gpu (MNMG) PCA expects a dask-cuDF object as input
+    and provides 2 algorithms, Full and Jacobi. Full (default) uses a full
+    eigendecomposition then selects the top K eigenvectors. The Jacobi
+    algorithm can be much faster as it iteratively tries to correct the top K
+    eigenvectors, but might be less accurate.
 
     Examples
     --------
@@ -107,9 +109,10 @@ class PCA(BaseDecomposition,
     n_components : int (default = 1)
         The number of top K singular vectors / values you want.
         Must be <= number(columns).
-    svd_solver : 'full', 'jacobi', or 'tsqr'
-        'full': run exact full SVD and select the components by postprocessing
-        'jacobi': iteratively compute SVD of the covariance matrix
+    svd_solver : 'full', 'jacobi', 'auto'
+        'full': Run exact full SVD and select the components by postprocessing
+        'jacobi': Iteratively compute SVD of the covariance matrix
+        'auto': For compatiblity with Scikit-learn. Alias for 'jacobi'.
     verbose : int or boolean, default=False
         Sets logging level. It must be one of `cuml.common.logger.level_*`.
         See :ref:`verbosity-levels` for more info.

python/cuml/dask/decomposition/tsvd.py

@@ -1,4 +1,4 @@
-# Copyright (c) 2019-2021, NVIDIA CORPORATION.
+# Copyright (c) 2019-2022, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -97,7 +97,7 @@ class TruncatedSVD(BaseDecomposition,
     n_components : int (default = 1)
         The number of top K singular vectors / values you want.
         Must be <= number(columns).
-    svd_solver : 'full'
+    svd_solver : 'full', 'jacobi'
         Only Full algorithm is supported since it's significantly faster on GPU
         then the other solvers including randomized SVD.
     verbose : int or boolean, default=False

python/cuml/decomposition/base_mg.pyx

@@ -37,6 +37,14 @@ from cuml.decomposition.utils cimport *
 from cuml.common import input_to_cuml_array
 from cuml.common.opg_data_utils_mg cimport *
 
+from enum import IntEnum
+
+
+class MGSolver(IntEnum):
+    COV_EIG_DQ = <underlying_type_t_solver> mg_solver.COV_EIG_DQ
+    COV_EIG_JACOBI = <underlying_type_t_solver> mg_solver.COV_EIG_JACOBI
+    QR = <underlying_type_t_solver> mg_solver.QR
+
 
 class BaseDecompositionMG(object):
 

python/cuml/decomposition/pca_mg.pyx

@@ -35,30 +35,12 @@ import cuml.common.opg_data_utils_mg as opg
 import cuml.internals
 from cuml.common.base import Base
 from raft.common.handle cimport handle_t
-from cuml.decomposition.utils cimport paramsSolver
+from cuml.decomposition.utils cimport *
 from cuml.common.opg_data_utils_mg cimport *
 
-from cuml.decomposition import PCA
-from cuml.decomposition.base_mg import BaseDecompositionMG
-
-
-ctypedef int underlying_type_t_solver
-
-
-cdef extern from "cuml/decomposition/pca_mg.hpp" namespace "ML":
 
-    ctypedef enum mg_solver "ML::mg_solver":
-        COV_EIG_DQ "ML::mg_solver::COV_EIG_DQ"
-        COV_EIG_JACOBI "ML::mg_solver::COV_EIG_JACOBI"
-        QR "ML::mg_solver::QR"
-
-    cdef cppclass paramsTSVDMG(paramsSolver):
-        size_t n_components
-        mg_solver algorithm  # = solver::COV_EIG_DQ
-
-    cdef cppclass paramsPCAMG(paramsTSVDMG):
-        bool copy
-        bool whiten
+from cuml.decomposition import PCA
+from cuml.decomposition.base_mg import BaseDecompositionMG, MGSolver
 
 
 cdef extern from "cuml/decomposition/pca_mg.hpp" namespace "ML::PCA::opg":
@@ -88,12 +70,6 @@ cdef extern from "cuml/decomposition/pca_mg.hpp" namespace "ML::PCA::opg":
                   bool verbose) except +
 
 
-class MGSolver(IntEnum):
-    COV_EIG_DQ = <underlying_type_t_solver> mg_solver.COV_EIG_DQ
-    COV_EIG_JACOBI = <underlying_type_t_solver> mg_solver.COV_EIG_JACOBI
-    QR = <underlying_type_t_solver> mg_solver.QR
-
-
 class PCAMG(BaseDecompositionMG, PCA):
 
     def __init__(self, **kwargs):
@@ -103,6 +79,7 @@ class PCAMG(BaseDecompositionMG, PCA):
         algo_map = {
             'full': MGSolver.COV_EIG_DQ,
             'auto': MGSolver.COV_EIG_JACOBI,
+            'jacobi': MGSolver.COV_EIG_JACOBI,
             # 'arpack': NOT_SUPPORTED,
             # 'randomized': NOT_SUPPORTED,
         }
@@ -118,7 +95,6 @@ class PCAMG(BaseDecompositionMG, PCA):
         params.n_rows = n_rows
         params.n_cols = n_cols
         params.whiten = self.whiten
-        params.n_iterations = self.iterated_power
         params.tol = self.tol
         params.algorithm = <mg_solver> (<underlying_type_t_solver> (
             self.c_algorithm))