Statistics funcs cov (#2)

* naive numpy.cov implementation

* del statistics_tests

* add statics_tests

* throw errors in np.cov

Author: Alexander-Makaryev

dpnp/backend_statistics.pyx

@@ -38,10 +38,44 @@ from dpnp.dpnp_utils cimport checker_throw_type_error, normalize_axis
 
 
 __all__ += [
+    "dpnp_cov",
     "dpnp_mean"
 ]
 
 
+cpdef dparray dpnp_cov(dparray array1):
+    # behaviour of original numpy
+    if array1.ndim > 2:
+        raise ValueError("array has more than 2 dimensions")
+
+    if array1.ndim < 2:
+        raise NotImplementedError
+
+    # numpy provide result as float64 for any input type
+    cdef dparray mean = dparray(array1.shape[0], dtype=numpy.float64)
+    cdef dparray X = dparray(array1.shape, dtype=numpy.float64)
+
+    # mean(array1, axis=1) #################################
+    # dpmp.mean throws: 'dpnp.dparray.dparray' object is not callable
+    for i in range(array1.shape[0]):
+        sum = 0.0
+        for j in range(array1.shape[1]):
+            sum += array1[i, j]
+        mean[i] = sum / array1.shape[1]
+    ########################################################
+    #X = array1 - mean[:, None]
+    #X = array1 - mean[:, numpy.newaxis]
+    #X = array1 - mean.reshape((array1.shape[0], 1))
+    for i in range(array1.shape[0]):
+        for j in range(array1.shape[1]):
+            X[i, j] = array1[i, j] - mean[i]
+    ########################################################
+    Y = X.transpose()
+    res = dpnp_matmul(X, Y) / (array1.shape[1] - 1)
+
+    return res
+
+
 cpdef dparray dpnp_mean(dparray a, axis):
     cdef dparray_shape_type shape_a = a.shape
     cdef long size_a = a.size

dpnp/dpnp_iface_statistics.py

@@ -47,10 +47,37 @@
 from dpnp.dpnp_utils import checker_throw_value_error, use_origin_backend
 
 __all__ = [
+    'cov',
     'mean'
 ]
 
 
+def cov(in_array1, y=None, rowvar=True, bias=False, ddof=None, fweights=None, aweights=None):
+    """
+    Estimate a covariance matrix, given data and weights.
+    """
+
+    is_dparray1 = isinstance(in_array1, dparray)
+
+    if (not use_origin_backend(in_array1) and is_dparray1):
+        if y is not None:
+            checker_throw_value_error("cov", "y", type(y), None)
+        if rowvar is not True:
+            checker_throw_value_error("cov", "rowvar", rowvar, True)
+        if bias is not False:
+            checker_throw_value_error("cov", "bias", bias, False)
+        if ddof is not None:
+            checker_throw_value_error("cov", "ddof", type(ddof), None)
+        if fweights is not None:
+            checker_throw_value_error("cov", "fweights", type(fweights), None)
+        if aweights is not None:
+            checker_throw_value_error("cov", "aweights", type(aweights), None)
+
+        return dpnp_cov(in_array1)
+
+    return numpy.cov(in_array1)
+
+
 def mean(a, axis=None):
     """
     Compute the arithmetic mean along the specified axis.

tests/third_party/cupy/statistics_tests/__init__.py renamed to tests/third_party/cupy/statics_tests/__init__.py

@@ -0,0 +1,166 @@
+import unittest
+
+import numpy
+import pytest
+
+import dpnp as cupy
+from tests.third_party.cupy import testing
+
+
+@testing.gpu
+class TestCorrcoef(unittest.TestCase):
+
+    @testing.for_all_dtypes()
+    @testing.numpy_cupy_allclose()
+    def test_corrcoef(self, xp, dtype):
+        a = testing.shaped_arange((2, 3), xp, dtype)
+        return xp.corrcoef(a)
+
+    @testing.for_all_dtypes()
+    @testing.numpy_cupy_allclose()
+    def test_corrcoef_diag_exception(self, xp, dtype):
+        a = testing.shaped_arange((1, 3), xp, dtype)
+        return xp.corrcoef(a)
+
+    @testing.for_all_dtypes()
+    @testing.numpy_cupy_allclose()
+    def test_corrcoef_y(self, xp, dtype):
+        a = testing.shaped_arange((2, 3), xp, dtype)
+        y = testing.shaped_arange((2, 3), xp, dtype)
+        return xp.corrcoef(a, y=y)
+
+    @testing.for_all_dtypes()
+    @testing.numpy_cupy_allclose()
+    def test_corrcoef_rowvar(self, xp, dtype):
+        a = testing.shaped_arange((2, 3), xp, dtype)
+        y = testing.shaped_arange((2, 3), xp, dtype)
+        return xp.corrcoef(a, y=y, rowvar=False)
+
+
+@testing.gpu
+class TestCov(unittest.TestCase):
+
+    def generate_input(self, a_shape, y_shape, xp, dtype):
+        a = testing.shaped_arange(a_shape, xp, dtype)
+        y = None
+        if y_shape is not None:
+            y = testing.shaped_arange(y_shape, xp, dtype)
+        return a, y
+
+    @testing.for_all_dtypes()
+    @testing.numpy_cupy_allclose()
+    def check(self, a_shape, y_shape=None, rowvar=True, bias=False,
+              ddof=None, xp=None, dtype=None):
+        a, y = self.generate_input(a_shape, y_shape, xp, dtype)
+        return xp.cov(a, y, rowvar, bias, ddof)
+
+    @testing.for_all_dtypes()
+    @testing.numpy_cupy_allclose()
+    def check_warns(self, a_shape, y_shape=None, rowvar=True, bias=False,
+                    ddof=None, xp=None, dtype=None):
+        with testing.assert_warns(RuntimeWarning):
+            a, y = self.generate_input(a_shape, y_shape, xp, dtype)
+            return xp.cov(a, y, rowvar, bias, ddof)
+
+    @testing.for_all_dtypes()
+    def check_raises(self, a_shape, y_shape=None, rowvar=True, bias=False,
+                     ddof=None, dtype=None):
+        for xp in (numpy, cupy):
+            a, y = self.generate_input(a_shape, y_shape, xp, dtype)
+            with pytest.raises(ValueError):
+                xp.cov(a, y, rowvar, bias, ddof)
+
+    def test_cov(self):
+        self.check((2, 3))
+        self.check((2,), (2,))
+        self.check((1, 3), (1, 3), rowvar=False)
+        self.check((2, 3), (2, 3), rowvar=False)
+        self.check((2, 3), bias=True)
+        self.check((2, 3), ddof=2)
+
+    def test_cov_warns(self):
+        self.check_warns((2, 3), ddof=3)
+        self.check_warns((2, 3), ddof=4)
+
+    def test_cov_raises(self):
+        self.check_raises((2, 3), ddof=1.2)
+        self.check_raises((3, 4, 2))
+        self.check_raises((2, 3), (3, 4, 2))
+
+    def test_cov_empty(self):
+        self.check((0, 1))
+
+
+@testing.gpu
+@testing.parameterize(*testing.product({
+    'mode': ['valid', 'same', 'full'],
+    'shape1': [(5,), (6,), (20,), (21,)],
+    'shape2': [(5,), (6,), (20,), (21,)],
+}))
+class TestCorrelateShapeCombination(unittest.TestCase):
+
+    @testing.for_all_dtypes(no_float16=True)
+    @testing.numpy_cupy_allclose(rtol=1e-4)
+    def test_correlate(self, xp, dtype):
+        a = testing.shaped_arange(self.shape1, xp, dtype)
+        b = testing.shaped_arange(self.shape2, xp, dtype)
+        return xp.correlate(a, b, mode=self.mode)
+
+
+@testing.gpu
+@testing.parameterize(*testing.product({
+    'mode': ['valid', 'full', 'same']
+}))
+class TestCorrelate(unittest.TestCase):
+
+    @testing.for_all_dtypes()
+    @testing.numpy_cupy_allclose(rtol=1e-5)
+    def test_correlate_non_contiguous(self, xp, dtype):
+        a = testing.shaped_arange((300,), xp, dtype)
+        b = testing.shaped_arange((100,), xp, dtype)
+        return xp.correlate(a[::200], b[10::70], mode=self.mode)
+
+    @testing.for_all_dtypes(no_float16=True)
+    @testing.numpy_cupy_allclose(rtol=1e-4)
+    def test_correlate_large_non_contiguous(self, xp, dtype):
+        a = testing.shaped_arange((10000,), xp, dtype)
+        b = testing.shaped_arange((1000,), xp, dtype)
+        return xp.correlate(a[200::], b[10::700], mode=self.mode)
+
+    @testing.for_all_dtypes_combination(names=['dtype1', 'dtype2'])
+    @testing.numpy_cupy_allclose(rtol=1e-2)
+    def test_correlate_diff_types(self, xp, dtype1, dtype2):
+        a = testing.shaped_random((200,), xp, dtype1)
+        b = testing.shaped_random((100,), xp, dtype2)
+        return xp.correlate(a, b, mode=self.mode)
+
+
+@testing.gpu
+@testing.parameterize(*testing.product({
+    'mode': ['valid', 'same', 'full']
+}))
+class TestCorrelateInvalid(unittest.TestCase):
+
+    @testing.with_requires('numpy>=1.18')
+    @testing.for_all_dtypes()
+    def test_correlate_empty(self, dtype):
+        for xp in (numpy, cupy):
+            a = xp.zeros((0,), dtype)
+            with pytest.raises(ValueError):
+                xp.correlate(a, a, mode=self.mode)
+
+    @testing.for_all_dtypes()
+    def test_correlate_ndim(self, dtype):
+        for xp in (numpy, cupy):
+            a = testing.shaped_arange((5, 10, 2), xp, dtype)
+            b = testing.shaped_arange((3, 4, 4), xp, dtype)
+            with pytest.raises(ValueError):
+                xp.correlate(a, b, mode=self.mode)
+
+    @testing.for_all_dtypes()
+    def test_correlate_zero_dim(self, dtype):
+        for xp in (numpy, cupy):
+            a = testing.shaped_arange((), xp, dtype)
+            b = testing.shaped_arange((1,), xp, dtype)
+            with pytest.raises(ValueError):
+                xp.correlate(a, b, mode=self.mode)