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

python/cuml/dask/decomposition/pca.py

@@ -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

@@ -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/pca_mg.pyx

@@ -35,31 +35,13 @@ import cuml.common.opg_data_utils_mg as opg
 import cuml.internals
 from cuml.common.base import Base
 from cuml.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
-
 
 cdef extern from "cuml/decomposition/pca_mg.hpp" namespace "ML::PCA::opg":
 
@@ -87,13 +69,6 @@ cdef extern from "cuml/decomposition/pca_mg.hpp" namespace "ML::PCA::opg":
                   paramsPCAMG &prms,
                   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 +78,7 @@ class PCAMG(BaseDecompositionMG, PCA):
         algo_map = {
             'full': MGSolver.COV_EIG_DQ,
             'auto': MGSolver.COV_EIG_JACOBI,
+            'jacobi': MGSoler.COV_EIG_JACOBI,
             # 'arpack': NOT_SUPPORTED,
             # 'randomized': NOT_SUPPORTED,
         }

python/cuml/decomposition/tsvd_mg.pyx

@@ -48,7 +48,7 @@ cdef extern from "cuml/decomposition/tsvd_mg.hpp" namespace "ML::TSVD::opg":
                             float *explained_var,
                             float *explained_var_ratio,
                             float *singular_vals,
-                            paramsTSVD &prms,
+                            paramsTSVDMG &prms,
                             bool verbose) except +
 
     cdef void fit_transform(handle_t& handle,
@@ -60,7 +60,7 @@ cdef extern from "cuml/decomposition/tsvd_mg.hpp" namespace "ML::TSVD::opg":
                             double *explained_var,
                             double *explained_var_ratio,
                             double *singular_vals,
-                            paramsTSVD &prms,
+                            paramsTSVDMG &prms,
                             bool verbose) except +
 
 
@@ -69,6 +69,18 @@ class TSVDMG(BaseDecompositionMG, TruncatedSVD):
     def __init__(self, **kwargs):
         super(TSVDMG, self).__init__(**kwargs)
 
+    def _build_params(self, n_rows, n_cols):
+        cpdef paramsPCAMG *params = new paramsTSVDMG()
+        params.n_components = self._n_components
+        params.n_rows = n_rows
+        params.n_cols = n_cols
+        params.n_iterations = self.iterated_power
+        params.tol = self.tol
+        params.algorithm = <mg_solver> (<underlying_type_t_solver> (
+            self.c_algorithm))
+
+        return <size_t>params
+
     @cuml.internals.api_base_return_any_skipall
     def _call_fit(self, X, trans, rank, input_desc,
                   trans_desc, arg_params):
@@ -80,7 +92,7 @@ class TSVDMG(BaseDecompositionMG, TruncatedSVD):
         cdef uintptr_t singular_vals_ptr = self.singular_values_.ptr
         cdef handle_t* handle_ = <handle_t*><size_t>self.handle.getHandle()
 
-        cdef paramsTSVD *params = <paramsTSVD*><size_t>arg_params
+        cdef paramsTSVDMG *params = <paramsTSVDMG*><size_t>arg_params
 
         if self.dtype == np.float32:
 

python/cuml/decomposition/utils.pxd

@@ -18,6 +18,26 @@ from libcpp cimport bool
 
 ctypedef int underlying_type_t_solver
 
+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
+
+cdef extern from "cuml/decomposition/pca_mg.hpp" namespace "ML" nogil:
+
+    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
+
 cdef extern from "cuml/decomposition/params.hpp" namespace "ML" nogil:
 
     ctypedef enum solver "ML::solver":