zzy34407230
/
mindspore2022
Not watched
1
0
Fork 0
Code
Releases 22 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
Tree: a64920c468
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'a64920c468'
${ noResults }
mindspore2022/tests/st/scipy_st/test_ops.py

390 lines
17 kB
Raw Blame History 

  1. # Copyright 2021 Huawei Technologies Co., Ltd

  2. #

  3. # Licensed under the Apache License, Version 2.0 (the "License");

  4. # you may not use this file except in compliance with the License.

  5. # You may obtain a copy of the License at

  6. #

  7. # http://www.apache.org/licenses/LICENSE-2.0

  8. #

  9. # Unless required by applicable law or agreed to in writing, software

  10. # distributed under the License is distributed on an "AS IS" BASIS,

  11. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

  12. # See the License for the specific language governing permissions and

  13. # limitations under the License.

  14. # ============================================================================

  15. """st for scipy.ops."""

  16. from typing import Generic

  17. import pytest

  18. import numpy as np

  19. import scipy as scp

  20. from scipy.linalg import solve_triangular

  21. 

  22. from mindspore import Tensor, context

  23. from mindspore.scipy.ops import EighNet, EigNet, Cholesky, SolveTriangular

  24. from tests.st.scipy_st.utils import create_sym_pos_matrix

  25. 

  26. np.random.seed(0)

  27. 

  28. 

  29. @pytest.mark.level0

  30. @pytest.mark.platform_x86_cpu

  31. @pytest.mark.platform_x86_gpu_training

  32. @pytest.mark.env_onecard

  33. @pytest.mark.parametrize('n', [3, 5, 7])

  34. @pytest.mark.parametrize('lower', [True, False])

  35. @pytest.mark.parametrize('dtype', [np.float64])

  36. def test_cholesky(n: int, lower: bool, dtype: Generic):

  37.     """

  38.     Feature: ALL TO ALL

  39.     Description:  test cases for cholesky [N,N]

  40.     Expectation: the result match scipy cholesky

  41.     """

  42.     context.set_context(mode=context.GRAPH_MODE)

  43.     a = create_sym_pos_matrix((n, n), dtype)

  44.     tensor_a = Tensor(a)

  45.     expect = scp.linalg.cholesky(a, lower=lower)

  46.     cholesky_net = Cholesky(lower=lower, clean=True)

  47.     output = cholesky_net(tensor_a)

  48.     assert np.allclose(expect, output.asnumpy())

  49. 

  50. 

  51. @pytest.mark.level0

  52. @pytest.mark.platform_x86_cpu

  53. @pytest.mark.env_onecard

  54. @pytest.mark.parametrize('n', [4, 6, 9, 10])

  55. def test_eig_net(n: int):

  56.     """

  57.     Feature: ALL To ALL

  58.     Description: test cases for eigen decomposition test cases for Ax= lambda * x /( A- lambda * E)X=0

  59.     Expectation: the result match to numpy

  60.     """

  61.     context.set_context(mode=context.GRAPH_MODE)

  62.     # test for real scalar float 32

  63.     rtol = 1e-3

  64.     atol = 1e-4

  65.     msp_eig = EigNet(True)

  66.     A = np.array(np.random.rand(n, n), dtype=np.float32)

  67.     tensor_a = Tensor(np.array(A).astype(np.float32))

  68.     msp_w, msp_v = msp_eig(tensor_a)

  69.     assert np.allclose(A @ msp_v.asnumpy() - msp_v.asnumpy() @ np.diag(msp_w.asnumpy()), np.zeros((n, n)), rtol, atol)

  70. 

  71.     # test case for real scalar double 64

  72.     A = np.array(np.random.rand(n, n), dtype=np.float64)

  73.     rtol = 1e-5

  74.     atol = 1e-8

  75.     msp_eig = EigNet(True)

  76.     msp_w, msp_v = msp_eig(Tensor(np.array(A).astype(np.float64)))

  77. 

  78.     # Compare with scipy

  79.     # sp_w, sp_v = sp.linalg.eig(A.astype(np.float64))

  80.     assert np.allclose(A @ msp_v.asnumpy() - msp_v.asnumpy() @ np.diag(msp_w.asnumpy()), np.zeros((n, n)), rtol, atol)

  81. 

  82.     # test case for complex64

  83.     rtol = 1e-3

  84.     atol = 1e-4

  85.     A = np.array(np.random.rand(n, n), dtype=np.complex64)

  86.     for i in range(0, n):

  87.         for j in range(0, n):

  88.             if i == j:

  89.                 A[i][j] = complex(np.random.rand(1, 1), 0)

  90.             else:

  91.                 A[i][j] = complex(np.random.rand(1, 1), np.random.rand(1, 1))

  92.     msp_eig = EigNet(True)

  93.     msp_w, msp_v = msp_eig(Tensor(np.array(A).astype(np.complex64)))

  94.     # Compare with scipy, scipy passed

  95.     # sp_w, sp_v = sp.linalg.eig(A.astype(np.complex128))

  96.     # assert np.allclose(A @ sp_v - sp_v @ np.diag(sp_w), np.zeros((n, n)), rtol, atol)

  97. 

  98.     # print(A @ msp_v.asnumpy() - msp_v.asnumpy() @ np.diag(msp_w.asnumpy()))

  99.     assert np.allclose(A @ msp_v.asnumpy() - msp_v.asnumpy() @ np.diag(msp_w.asnumpy()), np.zeros((n, n)), rtol, atol)

  100. 

  101.     # test for complex128

  102.     rtol = 1e-5

  103.     atol = 1e-8

  104.     A = np.array(np.random.rand(n, n), dtype=np.complex128)

  105.     for i in range(0, n):

  106.         for j in range(0, n):

  107.             if i == j:

  108.                 A[i][j] = complex(np.random.rand(1, 1), 0)

  109.             else:

  110.                 A[i][j] = complex(np.random.rand(1, 1), np.random.rand(1, 1))

  111.     msp_eig = EigNet(True)

  112.     msp_w, msp_v = msp_eig(Tensor(np.array(A).astype(np.complex128)))

  113.     # Com`pare with scipy, scipy passed

  114.     # sp_w, sp_v = sp.linalg.eig(A.astype(np.complex128))

  115.     # assert np.allclose(A @ sp_v - sp_v @ np.diag(sp_w), np.zeros((n, n)), rtol, atol)

  116.     assert np.allclose(A @ msp_v.asnumpy() - msp_v.asnumpy() @ np.diag(msp_w.asnumpy()), np.zeros((n, n)), rtol, atol)

  117.     msp_eig = EigNet(False)

  118.     msp_w0 = msp_eig(Tensor(np.array(A).astype(np.complex128)))

  119.     assert np.allclose(msp_w0.asnumpy() - msp_w.asnumpy(), np.zeros((n, n)), rtol, atol)

  120. 

  121. 

  122. @pytest.mark.level0

  123. @pytest.mark.platform_x86_cpu

  124. @pytest.mark.env_onecard

  125. @pytest.mark.parametrize('n', [4, 6, 9, 10])

  126. def test_eigh_net_cpu(n: int):

  127.     """

  128.     Feature: ALL To ALL

  129.     Description: test cases for eigen decomposition test cases for Ax= lambda * x /( A- lambda * E)X=0

  130.     Expectation: the result match to numpy

  131.     """

  132.     context.set_context(mode=context.GRAPH_MODE)

  133.     # test for real scalar float 32

  134.     rtol = 1e-3

  135.     atol = 1e-4

  136. 

  137.     A = create_sym_pos_matrix((n, n), np.float32)

  138.     msp_eigh = EighNet(True, True)

  139.     msp_wl, msp_vl = msp_eigh(Tensor(np.array(A).astype(np.float32)))

  140.     msp_eigh = EighNet(True, False)

  141.     msp_wu, msp_vu = msp_eigh(Tensor(np.array(A).astype(np.float32)))

  142.     sym_Al = (np.tril((np.tril(A) - np.tril(A).T)) + np.tril(A).T)

  143.     sym_Au = (np.triu((np.triu(A) - np.triu(A).T)) + np.triu(A).T)

  144.     assert np.allclose(sym_Al @ msp_vl.asnumpy() - msp_vl.asnumpy() @ np.diag(msp_wl.asnumpy()), np.zeros((n, n)), rtol,

  145.                        atol)

  146.     assert np.allclose(sym_Au @ msp_vu.asnumpy() - msp_vu.asnumpy() @ np.diag(msp_wu.asnumpy()), np.zeros((n, n)), rtol,

  147.                        atol)

  148. 

  149.     # test case for real scalar double 64

  150.     A = np.random.rand(n, n)

  151.     rtol = 1e-5

  152.     atol = 1e-8

  153.     msp_eigh = EighNet(True, True)

  154.     msp_wl, msp_vl = msp_eigh(Tensor(np.array(A).astype(np.float64)))

  155.     msp_eigh = EighNet(True, False)

  156.     msp_wu, msp_vu = msp_eigh(Tensor(np.array(A).astype(np.float64)))

  157.     sym_Al = (np.tril((np.tril(A) - np.tril(A).T)) + np.tril(A).T)

  158.     sym_Au = (np.triu((np.triu(A) - np.triu(A).T)) + np.triu(A).T)

  159.     assert np.allclose(sym_Al @ msp_vl.asnumpy() - msp_vl.asnumpy() @ np.diag(msp_wl.asnumpy()), np.zeros((n, n)), rtol,

  160.                        atol)

  161.     assert np.allclose(sym_Au @ msp_vu.asnumpy() - msp_vu.asnumpy() @ np.diag(msp_wu.asnumpy()), np.zeros((n, n)), rtol,

  162.                        atol)

  163.     # test for real scalar float64 no vector

  164.     msp_eigh = EighNet(False, True)

  165.     msp_wl0 = msp_eigh(Tensor(np.array(A).astype(np.float64)))

  166.     msp_eigh = EighNet(False, False)

  167.     msp_wu0 = msp_eigh(Tensor(np.array(A).astype(np.float64)))

  168.     assert np.allclose(msp_wl.asnumpy() - msp_wl0.asnumpy(), np.zeros((n, n)), rtol, atol)

  169.     assert np.allclose(msp_wu.asnumpy() - msp_wu0.asnumpy(), np.zeros((n, n)), rtol, atol)

  170. 

  171.     # test case for complex64

  172.     rtol = 1e-3

  173.     atol = 1e-4

  174.     A = np.array(np.random.rand(n, n), dtype=np.complex64)

  175.     for i in range(0, n):

  176.         for j in range(0, n):

  177.             if i == j:

  178.                 A[i][j] = complex(np.random.rand(1, 1), 0)

  179.             else:

  180.                 A[i][j] = complex(np.random.rand(1, 1), np.random.rand(1, 1))

  181.     sym_Al = (np.tril((np.tril(A) - np.tril(A).T)) + np.tril(A).conj().T)

  182.     sym_Au = (np.triu((np.triu(A) - np.triu(A).T)) + np.triu(A).conj().T)

  183.     msp_eigh = EighNet(True, True)

  184.     msp_wl, msp_vl = msp_eigh(Tensor(np.array(A).astype(np.complex64)))

  185.     msp_eigh = EighNet(True, False)

  186.     msp_wu, msp_vu = msp_eigh(Tensor(np.array(A).astype(np.complex64)))

  187.     assert np.allclose(sym_Al @ msp_vl.asnumpy() - msp_vl.asnumpy() @ np.diag(msp_wl.asnumpy()), np.zeros((n, n)), rtol,

  188.                        atol)

  189.     assert np.allclose(sym_Au @ msp_vu.asnumpy() - msp_vu.asnumpy() @ np.diag(msp_wu.asnumpy()), np.zeros((n, n)), rtol,

  190.                        atol)

  191. 

  192.     # test for complex128

  193.     rtol = 1e-5

  194.     atol = 1e-8

  195.     A = np.array(np.random.rand(n, n), dtype=np.complex128)

  196.     for i in range(0, n):

  197.         for j in range(0, n):

  198.             if i == j:

  199.                 A[i][j] = complex(np.random.rand(1, 1), 0)

  200.             else:

  201.                 A[i][j] = complex(np.random.rand(1, 1), np.random.rand(1, 1))

  202.     sym_Al = (np.tril((np.tril(A) - np.tril(A).T)) + np.tril(A).conj().T)

  203.     sym_Au = (np.triu((np.triu(A) - np.triu(A).T)) + np.triu(A).conj().T)

  204.     msp_eigh = EighNet(True, True)

  205.     msp_wl, msp_vl = msp_eigh(Tensor(np.array(A).astype(np.complex128)))

  206.     msp_eigh = EighNet(True, False)

  207.     msp_wu, msp_vu = msp_eigh(Tensor(np.array(A).astype(np.complex128)))

  208.     assert np.allclose(sym_Al @ msp_vl.asnumpy() - msp_vl.asnumpy() @ np.diag(msp_wl.asnumpy()), np.zeros((n, n)), rtol,

  209.                        atol)

  210.     assert np.allclose(sym_Au @ msp_vu.asnumpy() - msp_vu.asnumpy() @ np.diag(msp_wu.asnumpy()), np.zeros((n, n)), rtol,

  211.                        atol)

  212. 

  213.     # test for real scalar complex128 no vector

  214.     msp_eigh = EighNet(False, True)

  215.     msp_wl0 = msp_eigh(Tensor(np.array(A).astype(np.complex128)))

  216.     msp_eigh = EighNet(False, False)

  217.     msp_wu0 = msp_eigh(Tensor(np.array(A).astype(np.complex128)))

  218.     assert np.allclose(msp_wl.asnumpy() - msp_wl0.asnumpy(), np.zeros((n, n)), rtol, atol)

  219.     assert np.allclose(msp_wu.asnumpy() - msp_wu0.asnumpy(), np.zeros((n, n)), rtol, atol)

  220. 

  221. 

  222. @pytest.mark.level0

  223. @pytest.mark.platform_x86_gpu_training

  224. @pytest.mark.env_onecard

  225. @pytest.mark.parametrize('n', [4, 6, 8, 20])

  226. def test_eigh_net_gpu(n: int):

  227.     """

  228.     Feature: ALL To ALL

  229.     Description: test cases for eigen decomposition test cases for Ax= lambda * x /( A- lambda * E)X=0

  230.     Expectation: the result match to numpy

  231.     """

  232.     context.set_context(mode=context.GRAPH_MODE)

  233.     # test for real scalar float 32

  234.     rtol = 1e-3

  235.     atol = 1e-4

  236.     A = create_sym_pos_matrix((n, n), np.float32)

  237.     msp_eigh = EighNet(True, True)

  238.     msp_wl, msp_vl = msp_eigh(Tensor(np.array(A).astype(np.float32)))

  239.     msp_eigh = EighNet(True, False)

  240.     msp_wu, msp_vu = msp_eigh(Tensor(np.array(A).astype(np.float32)))

  241.     assert np.allclose(A @ msp_vl.asnumpy() - msp_vl.asnumpy() @ np.diag(msp_wl.asnumpy()), np.zeros((n, n)), rtol,

  242.                        atol)

  243.     assert np.allclose(A @ msp_vu.asnumpy() - msp_vu.asnumpy() @ np.diag(msp_wu.asnumpy()), np.zeros((n, n)), rtol,

  244.                        atol)

  245. 

  246.     # test case for real scalar double 64

  247.     A = create_sym_pos_matrix((n, n), np.float64)

  248.     rtol = 1e-5

  249.     atol = 1e-8

  250.     msp_eigh = EighNet(True, True)

  251.     msp_wl, msp_vl = msp_eigh(Tensor(np.array(A).astype(np.float64)))

  252.     msp_eigh = EighNet(True, False)

  253.     msp_wu, msp_vu = msp_eigh(Tensor(np.array(A).astype(np.float64)))

  254.     assert np.allclose(A @ msp_vl.asnumpy() - msp_vl.asnumpy() @ np.diag(msp_wl.asnumpy()), np.zeros((n, n)), rtol,

  255.                        atol)

  256.     assert np.allclose(A @ msp_vu.asnumpy() - msp_vu.asnumpy() @ np.diag(msp_wu.asnumpy()), np.zeros((n, n)), rtol,

  257.                        atol)

  258.     # test for real scalar float64 no vector

  259.     msp_eigh = EighNet(False, True)

  260.     msp_wl0 = msp_eigh(Tensor(np.array(A).astype(np.float64)))

  261.     msp_eigh = EighNet(False, False)

  262.     msp_wu0 = msp_eigh(Tensor(np.array(A).astype(np.float64)))

  263.     assert np.allclose(msp_wl.asnumpy() - msp_wl0.asnumpy(), np.zeros((n, n)), rtol, atol)

  264.     assert np.allclose(msp_wu.asnumpy() - msp_wu0.asnumpy(), np.zeros((n, n)), rtol, atol)

  265. 

  266.     # test case for complex64

  267.     rtol = 1e-3

  268.     atol = 1e-4

  269.     A = np.array(np.random.rand(n, n), dtype=np.complex64)

  270.     for i in range(0, n):

  271.         for j in range(0, n):

  272.             if i == j:

  273.                 A[i][j] = complex(np.random.rand(1, 1), 0)

  274.             else:

  275.                 A[i][j] = complex(np.random.rand(1, 1), np.random.rand(1, 1))

  276.     sym_Al = (np.tril((np.tril(A) - np.tril(A).T)) + np.tril(A).conj().T)

  277.     sym_Au = (np.triu((np.triu(A) - np.triu(A).T)) + np.triu(A).conj().T)

  278.     msp_eigh = EighNet(True, True)

  279.     msp_wl, msp_vl = msp_eigh(Tensor(np.array(A).astype(np.complex64)))

  280.     msp_eigh = EighNet(True, False)

  281.     msp_wu, msp_vu = msp_eigh(Tensor(np.array(A).astype(np.complex64)))

  282.     assert np.allclose(sym_Al @ msp_vl.asnumpy() - msp_vl.asnumpy() @ np.diag(msp_wl.asnumpy()),

  283.                        np.zeros((n, n)), rtol, atol)

  284.     assert np.allclose(sym_Au @ msp_vu.asnumpy() - msp_vu.asnumpy() @ np.diag(msp_wu.asnumpy()),

  285.                        np.zeros((n, n)), rtol, atol)

  286. 

  287.     # test for complex128

  288.     rtol = 1e-5

  289.     atol = 1e-8

  290.     A = np.array(np.random.rand(n, n), dtype=np.complex128)

  291.     for i in range(0, n):

  292.         for j in range(0, n):

  293.             if i == j:

  294.                 A[i][j] = complex(np.random.rand(1, 1), 0)

  295.             else:

  296.                 A[i][j] = complex(np.random.rand(1, 1), np.random.rand(1, 1))

  297.     sym_Al = (np.tril((np.tril(A) - np.tril(A).T)) + np.tril(A).conj().T)

  298.     sym_Au = (np.triu((np.triu(A) - np.triu(A).T)) + np.triu(A).conj().T)

  299.     msp_eigh = EighNet(True, True)

  300.     msp_wl, msp_vl = msp_eigh(Tensor(np.array(A).astype(np.complex128)))

  301.     msp_eigh = EighNet(True, False)

  302.     msp_wu, msp_vu = msp_eigh(Tensor(np.array(A).astype(np.complex128)))

  303.     assert np.allclose(sym_Al @ msp_vl.asnumpy() - msp_vl.asnumpy() @ np.diag(msp_wl.asnumpy()),

  304.                        np.zeros((n, n)), rtol, atol)

  305.     assert np.allclose(sym_Au @ msp_vu.asnumpy() - msp_vu.asnumpy() @ np.diag(msp_wu.asnumpy()),

  306.                        np.zeros((n, n)), rtol, atol)

  307.     # test for real scalar complex128 no vector

  308.     msp_eigh = EighNet(False, True)

  309.     msp_wl0 = msp_eigh(Tensor(np.array(A).astype(np.complex128)))

  310.     msp_eigh = EighNet(False, False)

  311.     msp_wu0 = msp_eigh(Tensor(np.array(A).astype(np.complex128)))

  312.     assert np.allclose(msp_wl.asnumpy() - msp_wl0.asnumpy(), np.zeros((n, n)), rtol, atol)

  313.     assert np.allclose(msp_wu.asnumpy() - msp_wu0.asnumpy(), np.zeros((n, n)), rtol, atol)

  314. 

  315. 

  316. @pytest.mark.level0

  317. @pytest.mark.platform_x86_cpu

  318. @pytest.mark.platform_x86_gpu_training

  319. @pytest.mark.env_onecard

  320. @pytest.mark.parametrize('n', [10, 20])

  321. @pytest.mark.parametrize('trans', ["N", "T"])

  322. @pytest.mark.parametrize('dtype', [np.float32, np.float64])

  323. @pytest.mark.parametrize('lower', [False, True])

  324. @pytest.mark.parametrize('unit_diagonal', [False])

  325. def test_solve_triangular_2d(n: int, dtype, lower: bool, unit_diagonal: bool, trans: str):

  326.     """

  327.     Feature: ALL TO ALL

  328.     Description:  test cases for [N x N] X [N X 1]

  329.     Expectation: the result match scipy

  330.     """

  331.     context.set_context(mode=context.GRAPH_MODE)

  332.     a = (np.random.random((n, n)) + np.eye(n)).astype(dtype)

  333.     b = np.random.random((n, 1)).astype(dtype)

  334.     expect = solve_triangular(a, b, lower=lower, unit_diagonal=unit_diagonal, trans=trans)

  335.     solve = SolveTriangular(lower, unit_diagonal, trans)

  336.     output = solve(Tensor(a), Tensor(b)).asnumpy()

  337.     np.testing.assert_almost_equal(expect, output, decimal=5)

  338. 

  339. 

  340. @pytest.mark.level0

  341. @pytest.mark.platform_x86_cpu

  342. @pytest.mark.platform_x86_gpu_training

  343. @pytest.mark.env_onecard

  344. @pytest.mark.parametrize('n', [10, 20])

  345. @pytest.mark.parametrize('trans', ["N", "T"])

  346. @pytest.mark.parametrize('dtype', [np.float32, np.float64])

  347. @pytest.mark.parametrize('lower', [False, True])

  348. @pytest.mark.parametrize('unit_diagonal', [False, True])

  349. def test_solve_triangular_1d(n: int, dtype, lower: bool, unit_diagonal: bool, trans: str):

  350.     """

  351.     Feature: ALL TO ALL

  352.     Description: test cases for [N x N] X [N]

  353.     Expectation: the result match scipy

  354.     """

  355.     context.set_context(mode=context.GRAPH_MODE)

  356.     a = (np.random.random((n, n)) + np.eye(n)).astype(dtype)

  357.     b = np.random.random(n).astype(dtype)

  358.     expect = solve_triangular(a, b, lower=lower, unit_diagonal=unit_diagonal, trans=trans)

  359.     solve = SolveTriangular(lower, unit_diagonal, trans)

  360.     output = solve(Tensor(a), Tensor(b)).asnumpy()

  361.     np.testing.assert_almost_equal(expect, output, decimal=5)

  362. 

  363. 

  364. @pytest.mark.level0

  365. @pytest.mark.platform_x86_cpu

  366. @pytest.mark.platform_x86_gpu_training

  367. @pytest.mark.env_onecard

  368. @pytest.mark.parametrize('shape', [(4, 5), (10, 20)])

  369. @pytest.mark.parametrize('trans', ["N", "T"])

  370. @pytest.mark.parametrize('dtype', [np.float32, np.float64])

  371. @pytest.mark.parametrize('lower', [False, True])

  372. @pytest.mark.parametrize('unit_diagonal', [False, True])

  373. def test_solve_triangular_matrix(shape: int, dtype, lower: bool, unit_diagonal: bool, trans: str):

  374.     """

  375.     Feature: ALL TO ALL

  376.     Description: test cases for [N x N] X [N]

  377.     Expectation: the result match scipy

  378.     """

  379.     if trans == 'T':

  380.         n, m = shape

  381.     else:

  382.         m, n = shape

  383.     context.set_context(mode=context.GRAPH_MODE)

  384.     a = (np.random.random((m, m)) + np.eye(m)).astype(dtype)

  385.     b = np.random.random((m, n)).astype(dtype)

  386.     expect = solve_triangular(a, b, lower=lower, unit_diagonal=unit_diagonal, trans=trans)

  387.     solve = SolveTriangular(lower, unit_diagonal, trans)

  388.     output = solve(Tensor(a), Tensor(b)).asnumpy()

  389.     np.testing.assert_almost_equal(expect, output, decimal=5)