* impl - linalg matrix_rank for cpu and gpu implemented

* fix - python interface

python/mxnet/ndarray/numpy/linalg.py

@@ -23,7 +23,53 @@
 from . import _api_internal
 
 __all__ = ['norm', 'svd', 'cholesky', 'qr', 'inv', 'det', 'slogdet', 'solve', 'tensorinv', 'tensorsolve',
-           'pinv', 'eigvals', 'eig', 'eigvalsh', 'eigh', 'lstsq']
+           'pinv', 'eigvals', 'eig', 'eigvalsh', 'eigh', 'lstsq', 'matrix_rank']
+
+
+def matrix_rank(M, tol=None, hermitian=False):
+    """
+    Return matrix rank of array using SVD method
+
+    Rank of the array is the number of singular values of the array that are
+    greater than `tol`.
+
+    Parameters
+    M : {(M,), (..., M, N)} ndarray
+        Input vector or stack of matrices.
+    tol : (...) ndarray, float, optional
+        Threshold below which SVD values are considered zero. If `tol` is
+        None, and ``S`` is an array with singular values for `M`, and
+        ``eps`` is the epsilon value for datatype of ``S``, then `tol` is
+        set to ``S.max() * max(M.shape) * eps``.
+    hermitian : bool, optional
+        If True, `M` is assumed to be Hermitian (symmetric if real-valued),
+        enabling a more efficient method for finding singular values.
+        Defaults to False.
+
+    Returns
+    -------
+    rank : (...) ndarray
+        Rank of M.
+
+    Examples
+    --------
+    >>> from mxnet import np
+    >>> np.matrix_rank(np.eye(4)) # Full rank matrix
+    4
+    >>> I=np.eye(4); I[-1,-1] = 0. # rank deficient matrix
+    >>> np.matrix_rank(I)
+    3
+    >>> np.matrix_rank(np.ones((4,))) # 1 dimension - rank 1 unless all 0
+    1
+    >>> np.matrix_rank(np.zeros((4,)))
+    0
+    """
+    finfo_eps_32 = _np.finfo(_np.float32).eps
+    finfo_eps_64 = _np.finfo(_np.float64).eps
+    if tol is None:
+        return _npi.matrix_rank_none_tol(M, finfo_eps_32, finfo_eps_64, hermitian)
+    else:
+        return _npi.matrix_rank(M, tol, hermitian)
 
 
 def lstsq(a, b, rcond='warn'):

python/mxnet/numpy/fallback_linalg.py

@@ -24,11 +24,9 @@
 __all__ = [
     'cond',
     'matrix_power',
-    'matrix_rank',
     'multi_dot'
 ]
 
 cond = onp.linalg.cond
 matrix_power = onp.linalg.matrix_power
-matrix_rank = onp.linalg.matrix_rank
 multi_dot = onp.linalg.multi_dot

python/mxnet/numpy/linalg.py

@@ -22,10 +22,51 @@
 from . import fallback_linalg
 
 __all__ = ['norm', 'svd', 'cholesky', 'qr', 'inv', 'det', 'slogdet', 'solve', 'tensorinv', 'tensorsolve',
-           'pinv', 'eigvals', 'eig', 'eigvalsh', 'eigh', 'lstsq']
+           'pinv', 'eigvals', 'eig', 'eigvalsh', 'eigh', 'lstsq', 'matrix_rank']
 __all__ += fallback_linalg.__all__
 
 
+def matrix_rank(M, tol=None, hermitian=False):
+    """
+    Return matrix rank of array using SVD method
+
+    Rank of the array is the number of singular values of the array that are
+    greater than `tol`.
+
+    Parameters
+    M : {(M,), (..., M, N)} ndarray
+        Input vector or stack of matrices.
+    tol : (...) ndarray, float, optional
+        Threshold below which SVD values are considered zero. If `tol` is
+        None, and ``S`` is an array with singular values for `M`, and
+        ``eps`` is the epsilon value for datatype of ``S``, then `tol` is
+        set to ``S.max() * max(M.shape) * eps``.
+    hermitian : bool, optional
+        If True, `M` is assumed to be Hermitian (symmetric if real-valued),
+        enabling a more efficient method for finding singular values.
+        Defaults to False.
+
+    Returns
+    -------
+    rank : (...) ndarray
+        Rank of M.
+
+    Examples
+    --------
+    >>> from mxnet import np
+    >>> np.matrix_rank(np.eye(4)) # Full rank matrix
+    4
+    >>> I=np.eye(4); I[-1,-1] = 0. # rank deficient matrix
+    >>> np.matrix_rank(I)
+    3
+    >>> np.matrix_rank(np.ones((4,))) # 1 dimension - rank 1 unless all 0
+    1
+    >>> np.matrix_rank(np.zeros((4,)))
+    0
+    """
+    return _mx_nd_np.linalg.matrix_rank(M, tol, hermitian)
+
+
 def lstsq(a, b, rcond='warn'):
     r"""
     Return the least-squares solution to a linear matrix equation.

python/mxnet/numpy_dispatch_protocol.py

@@ -166,6 +166,7 @@ def _run_with_array_ufunc_proto(*args, **kwargs):
     'linalg.eigvalsh',
     'linalg.eigh',
     'linalg.qr',
+    'linalg.matrix_rank',
     'shape',
     'trace',
     'tril',

python/mxnet/symbol/numpy/linalg.py

@@ -23,7 +23,40 @@
 from . import _internal as _npi
 
 __all__ = ['norm', 'svd', 'cholesky', 'qr', 'inv', 'det', 'slogdet', 'solve', 'tensorinv', 'tensorsolve',
-           'pinv', 'eigvals', 'eig', 'eigvalsh', 'eigh', 'lstsq']
+           'pinv', 'eigvals', 'eig', 'eigvalsh', 'eigh', 'lstsq', 'matrix_rank']
+
+
+def matrix_rank(M, tol=None, hermitian=False):
+    """
+    Return matrix rank of array using SVD method
+
+    Rank of the array is the number of singular values of the array that are
+    greater than `tol`.
+
+    Parameters
+    M : {(M,), (..., M, N)} _Symbol
+        Input vector or stack of matrices.
+    tol : (...) _Symbol, float, optional
+        Threshold below which SVD values are considered zero. If `tol` is
+        None, and ``S`` is an array with singular values for `M`, and
+        ``eps`` is the epsilon value for datatype of ``S``, then `tol` is
+        set to ``S.max() * max(M.shape) * eps``.
+    hermitian : bool, optional
+        If True, `M` is assumed to be Hermitian (symmetric if real-valued),
+        enabling a more efficient method for finding singular values.
+        Defaults to False.
+
+    Returns
+    -------
+    rank : (...) _Symbol
+        Rank of M.
+    """
+    finfo_eps_32 = _np.finfo(_np.float32).eps
+    finfo_eps_64 = _np.finfo(_np.float64).eps
+    if tol is None:
+        return _npi.matrix_rank_none_tol(M, finfo_eps_32, finfo_eps_64, hermitian)
+    else:
+        return _npi.matrix_rank(M, tol, hermitian)
 
 
 def lstsq(a, b, rcond='warn'):