def mnp_clip(x):
a = mnp.clip(x, mnp.asarray(10.0), mnp.asarray([
2,
]))
b = mnp.clip(x, 0, 1)
c = mnp.clip(x, mnp.asarray(0), mnp.asarray(10), dtype=mnp.float32)
return a, b, c
def test_asarray():
test_case = Cases()
for array in test_case.array_sets:
# Check for dtype matching
actual = onp.asarray(array)
expected = mnp.asarray(array).asnumpy()
# Since we set float32/int32 as the default dtype in mindspore, we need
# to make a conversion between numpy.asarray and mindspore.numpy.asarray
if actual.dtype is onp.dtype('float64'):
assert expected.dtype == onp.dtype('float32')
elif actual.dtype is onp.dtype('int64'):
assert expected.dtype == onp.dtype('int32')
else:
assert actual.dtype == expected.dtype
match_array(actual, expected, error=7)
for i in range(len(test_case.onp_dtypes)):
actual = onp.asarray(array, test_case.onp_dtypes[i])
expected = mnp.asarray(array, test_case.mnp_dtypes[i]).asnumpy()
match_array(actual, expected, error=7)
# Additional tests for nested tensor mixture
mnp_input = [(mnp.ones(3, ), mnp.ones(3)), [[1, 1, 1], (1, 1, 1)]]
onp_input = [(onp.ones(3, ), onp.ones(3)), [[1, 1, 1], (1, 1, 1)]]
actual = onp.asarray(onp_input)
expected = mnp.asarray(mnp_input).asnumpy()
match_array(actual, expected, error=7)
def test_append():
onp_array = onp.random.random((4, 3, 2)).astype('float32')
onp_value = onp.random.random((4, 3, 2)).astype('float32')
mnp_array = mnp.asarray(onp_array)
mnp_value = mnp.asarray(onp_value)
onp_res = onp_append(onp_array, onp_value)
mnp_res = mnp_append(mnp_array, mnp_value)
check_all_results(onp_res, mnp_res)
def test_type_promotion():
arr = rand_int(2, 3)
onp_sum = onp_add(arr, arr)
a = mnp.asarray(arr, dtype='float16')
b = mnp.asarray(arr, dtype='float32')
c = mnp.asarray(arr, dtype='int32')
match_array(mnp_add(a, b).asnumpy(), onp_sum)
match_array(mnp_add(b, c).asnumpy(), onp_sum)
def test_squeeze():
onp_array = onp.random.random((1, 3, 1, 4, 2)).astype('float32')
mnp_array = mnp.asarray(onp_array)
o_squeezed = onp_squeeze(onp_array)
m_squeezed = mnp_squeeze(mnp_array)
check_all_results(o_squeezed, m_squeezed)
onp_array = onp.random.random((1, 1, 1, 1, 1)).astype('float32')
mnp_array = mnp.asarray(onp_array)
o_squeezed = onp_squeeze(onp_array)
m_squeezed = mnp_squeeze(mnp_array)
check_all_results(o_squeezed, m_squeezed)
def test_array():
# array's function is very similar to asarray, so we mainly test the
# `copy` argument.
test_case = Cases()
for array in test_case.array_sets:
arr1 = mnp.asarray(array)
arr2 = mnp.array(arr1, copy=False)
arr3 = mnp.array(arr1)
arr4 = mnp.asarray(array, dtype='int32')
arr5 = mnp.asarray(arr4, dtype=mnp.int32)
assert arr1 is arr2
assert arr1 is not arr3
assert arr4 is arr5
def test_cumsum():
x = mnp.ones((16, 16), dtype="bool")
match_array(mnp.cumsum(x).asnumpy(), onp.cumsum(x.asnumpy()))
match_array(
mnp.cumsum(x, axis=0).asnumpy(), onp.cumsum(x.asnumpy(), axis=0))
match_meta(mnp.cumsum(x).asnumpy(), onp.cumsum(x.asnumpy()))
x = rand_int(3, 4, 5)
match_array(
mnp.cumsum(mnp.asarray(x), dtype="bool").asnumpy(),
onp.cumsum(x, dtype="bool"))
match_array(
mnp.cumsum(mnp.asarray(x), axis=-1).asnumpy(), onp.cumsum(x, axis=-1))
def test_ndarray_astype():
onp_array = onp.random.random((3, 4, 5)).astype('float32')
mnp_array = mnp.asarray(onp_array)
o_astype = onp_ndarray_astype(onp_array)
m_astype = mnp_ndarray_astype(mnp_array)
for arr1, arr2 in zip(o_astype, m_astype):
assert arr1.dtype == arr2.asnumpy().dtype
def run_single_test(mnp_fn, onp_fn, arr, error=0):
mnp_arr = mnp.asarray(arr)
for actual, expected in zip(mnp_fn(mnp_arr), onp_fn(arr)):
if isinstance(expected, tuple):
for actual_arr, expected_arr in zip(actual, expected):
match_array(actual_arr.asnumpy(), expected_arr, error)
match_array(actual.asnumpy(), expected, error)
def test_tensor_flatten():
lst = [[1.0, 2.0], [3.0, 4.0]]
tensor_list = mnp.asarray(lst)
assert tensor_list.flatten().asnumpy().tolist() == [1.0, 2.0, 3.0, 4.0]
assert tensor_list.flatten(order='F').asnumpy().tolist() == [
1.0, 3.0, 2.0, 4.0
]
def test_tensor_ravel():
lst = [[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0]]
tensor_list = mnp.asarray(lst)
assert tensor_list.ravel().shape == (8, )
assert tensor_list.ravel().asnumpy().tolist() == [
1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0
]
def test_negative():
arr = onp.arange(-6, 6).reshape((2, 2, 3)).astype('float32')
out_lst = [
onp.ones((2, 2, 3)).astype('float32'),
onp.ones((5, 2, 2, 3)).astype('float32')
]
where_lst = [
onp.full((2, 2, 3), [True, False, True]),
onp.full((2, 3), False)
]
for out in out_lst:
for where in where_lst:
onp_neg = onp_negative(arr, out=out, where=where)
mnp_neg = mnp_negative(mnp.asarray(arr), mnp.asarray(out),
mnp.asarray(where))
match_array(mnp_neg.asnumpy(), onp_neg, 1e-5)
def test_tril_indices_from():
m = int(rand_int().tolist())
n = int(rand_int().tolist())
t = mnp.asarray(rand_int(m, n).tolist())
k = rand_int().tolist()
mnp_res = mnp.tril_indices_from(t, k)
onp_res = onp.tril_indices_from(t.asnumpy(), k)
match_all_arrays(mnp_res, onp_res)
def test_tensor_swapaxes():
lst = [[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]
tensor_list = mnp.asarray(lst)
with pytest.raises(TypeError):
tensor_list = tensor_list.swapaxes(0, (1,))
with pytest.raises(ValueError):
tensor_list = tensor_list.swapaxes(0, 3)
assert tensor_list.swapaxes(0, 1).shape == (3, 2)
def test_tensor_squeeze():
lst = [[[1.0], [2.0], [3.0]]]
tensor_list = mnp.asarray(lst)
with pytest.raises(TypeError):
tensor_list = tensor_list.squeeze(1.2)
with pytest.raises(ValueError):
tensor_list = tensor_list.squeeze(4)
assert tensor_list.squeeze().shape == (3,)
assert tensor_list.squeeze(axis=2).shape == (1, 3)
def test_tensor_reshape():
lst = [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0]
tensor_list = mnp.asarray(lst)
with pytest.raises(TypeError):
tensor_list = tensor_list.reshape({0, 1, 2})
with pytest.raises(ValueError):
tensor_list = tensor_list.reshape(1, 2, 3)
assert tensor_list.reshape([-1, 4]).shape == (2, 4)
assert tensor_list.reshape(1, -1, 4).shape == (1, 2, 4)
def test_dsplit():
onp_arrs = [onp.random.randint(1, 5, size=(5, 4, 9)).astype('float32')]
mnp_arrs = [mnp.asarray(arr) for arr in onp_arrs]
for onp_arr, mnp_arr in zip(onp_arrs, mnp_arrs):
o_dsplit = onp_dsplit(onp_arr)
m_dsplit = mnp_dsplit(mnp_arr)
for expect_lst, actual_lst in zip(o_dsplit, m_dsplit):
for expect, actual in zip(expect_lst, actual_lst):
match_array(expect, actual.asnumpy())
def construct_arrays(n=1, ndim=1, axis=None, low=1, high=5):
onp_array_lst = []
mnp_array_lst = []
shape = onp.random.randint(low=low, high=high, size=ndim)
new_shape = [sh for sh in shape]
while n > 0:
n -= 1
onp_array1 = onp.random.randint(
low=low, high=high, size=shape).astype(onp.float32)
onp_array_lst.append(onp_array1)
mnp_array_lst.append(mnp.asarray(onp_array1))
if axis is not None and axis < ndim:
new_shape[axis] += onp.random.randint(2)
onp_array2 = onp.random.randint(
low=low, high=high, size=new_shape).astype(onp.float32)
onp_array_lst.append(onp_array2)
mnp_array_lst.append(mnp.asarray(onp_array2))
return onp_array_lst, mnp_array_lst
def test_pad_gpu():
x_np = onp.random.random([2, 1, 4, 3]).astype("float32")
x_ms = mnp.asarray(x_np.tolist())
# pad symmetric odd
mnp_res = mnp.pad(x_ms, ((10, 3), (5, 2), (3, 0), (2, 6)),
mode='symmetric',
reflect_type='odd')
onp_res = onp.pad(x_np, ((10, 3), (5, 2), (3, 0), (2, 6)),
mode='symmetric',
reflect_type='odd')
match_all_arrays(mnp_res, onp_res, error=1e-5)
# pad symmetric even
mnp_res = mnp.pad(x_ms, ((10, 13), (5, 12), (3, 0), (2, 6)),
mode='symmetric',
reflect_type='even')
onp_res = onp.pad(x_np, ((10, 13), (5, 12), (3, 0), (2, 6)),
mode='symmetric',
reflect_type='even')
match_all_arrays(mnp_res, onp_res, error=1e-5)
# pad reflect odd
mnp_res = mnp.pad(x_ms, ((10, 3), (5, 2), (3, 0), (2, 6)),
mode='reflect',
reflect_type='odd')
onp_res = onp.pad(x_np, ((10, 3), (5, 2), (3, 0), (2, 6)),
mode='reflect',
reflect_type='odd')
match_all_arrays(mnp_res, onp_res, error=1e-5)
# pad reflect even
mnp_res = mnp.pad(x_ms, ((10, 13)), mode='reflect', reflect_type='even')
onp_res = onp.pad(x_np, ((10, 13)), mode='reflect', reflect_type='even')
match_all_arrays(mnp_res, onp_res, error=1e-5)
# pad func
x_np = onp.random.random([2, 4]).astype("float32")
x_ms = mnp.asarray(x_np.tolist())
mnp_res = mnp.pad(x_ms, ((5, 5)), mode=pad_with_msfunc, padder=99)
onp_res = onp.pad(x_np, ((5, 5)), mode=pad_with_npfunc, padder=99)
match_all_arrays(mnp_res, onp_res, error=1e-5)
def test_asarray():
test_case = Cases()
for array in test_case.array_sets:
# Check for dtype matching
actual = onp.asarray(array)
expected = mnp.asarray(array).asnumpy()
# Since we set float32/int32 as the default dtype in mindspore, we need
# to make a conversion between numpy.asarray and mindspore.numpy.asarray
if actual.dtype is onp.dtype('float64'):
assert expected.dtype == onp.dtype('float32')
elif actual.dtype is onp.dtype('int64'):
assert expected.dtype == onp.dtype('int32')
else:
assert actual.dtype == expected.dtype
match_array(actual, expected, error=7)
for i in range(len(test_case.onp_dtypes)):
actual = onp.asarray(array, test_case.onp_dtypes[i])
expected = mnp.asarray(array, test_case.mnp_dtypes[i]).asnumpy()
match_array(actual, expected, error=7)
def test_array():
# array's function is very similar to asarray, so we mainly test the
# `copy` argument.
test_case = Cases()
for array in test_case.array_sets:
arr1 = mnp.asarray(array)
arr2 = mnp.array(arr1, copy=False)
arr3 = mnp.array(arr1)
arr4 = mnp.asarray(array, dtype='int32')
arr5 = mnp.asarray(arr4, dtype=mnp.int32)
assert arr1 is arr2
assert arr1 is not arr3
assert arr4 is arr5
# Additional tests for nested tensor/numpy_array mixture
mnp_input = [(mnp.ones(3, ), mnp.ones(3)), [[1, 1, 1], (1, 1, 1)]]
onp_input = [(onp.ones(3, ), onp.ones(3)), [[1, 1, 1], (1, 1, 1)]]
actual = onp.array(onp_input)
expected = mnp.array(mnp_input).asnumpy()
match_array(actual, expected, error=7)
def test_concatenate_type_promotion():
onp_array = onp.random.random((5, 1)).astype('float32')
mnp_array = mnp.asarray(onp_array)
onp_array1 = onp_array.astype(onp.float16)
onp_array2 = onp_array.astype(onp.bool_)
onp_array3 = onp_array.astype(onp.float32)
onp_array4 = onp_array.astype(onp.int32)
mnp_array1 = mnp_array.astype(onp.float16)
mnp_array2 = mnp_array.astype(onp.bool_)
mnp_array3 = mnp_array.astype(onp.float32)
mnp_array4 = mnp_array.astype(onp.int32)
o_concatenate = onp_concatenate_type_promotion(
onp_array1, onp_array2, onp_array3, onp_array4).astype('float32')
m_concatenate = mnp_concatenate_type_promotion(
mnp_array1, mnp_array2, mnp_array3, mnp_array4)
check_all_results(o_concatenate, m_concatenate, error=1e-7)
def test_pad():
x_np = onp.random.random([2, 3, 4]).astype("float32")
x_ms = mnp.asarray(x_np.tolist())
# pad constant
mnp_res = mnp.pad(x_ms, ((1, 1), (2, 2), (3, 4)))
onp_res = onp.pad(x_np, ((1, 1), (2, 2), (3, 4)))
match_all_arrays(mnp_res, onp_res, error=1e-5)
mnp_res = mnp.pad(x_ms, ((1, 1), (2, 3), (4, 5)),
constant_values=((3, 4), (5, 6), (7, 8)))
onp_res = onp.pad(x_np, ((1, 1), (2, 3), (4, 5)),
constant_values=((3, 4), (5, 6), (7, 8)))
match_all_arrays(mnp_res, onp_res, error=1e-5)
# pad statistic
mnp_res = mnp.pad(x_ms, ((1, 1), (2, 2), (3, 4)),
mode="mean",
stat_length=((1, 2), (2, 10), (3, 4)))
onp_res = onp.pad(x_np, ((1, 1), (2, 2), (3, 4)),
mode="mean",
stat_length=((1, 2), (2, 10), (3, 4)))
match_all_arrays(mnp_res, onp_res, error=1e-5)
# pad edge
mnp_res = mnp.pad(x_ms, ((1, 1), (2, 2), (3, 4)), mode="edge")
onp_res = onp.pad(x_np, ((1, 1), (2, 2), (3, 4)), mode="edge")
match_all_arrays(mnp_res, onp_res, error=1e-5)
# pad wrap
mnp_res = mnp.pad(x_ms, ((1, 1), (2, 2), (3, 4)), mode="wrap")
onp_res = onp.pad(x_np, ((1, 1), (2, 2), (3, 4)), mode="wrap")
match_all_arrays(mnp_res, onp_res, error=1e-5)
# pad linear_ramp
mnp_res = mnp.pad(x_ms, ((1, 3), (5, 2), (3, 0)),
mode="linear_ramp",
end_values=((0, 10), (9, 1), (-10, 99)))
onp_res = onp.pad(x_np, ((1, 3), (5, 2), (3, 0)),
mode="linear_ramp",
end_values=((0, 10), (9, 1), (-10, 99)))
match_all_arrays(mnp_res, onp_res, error=1e-5)
def test_linspace():
actual = onp.linspace(2.0, 3.0, dtype=onp.float32)
expected = mnp.linspace(2.0, 3.0).asnumpy()
match_array(actual, expected, error=6)
actual = onp.linspace(2.0, 3.0, num=5, dtype=onp.float32)
expected = mnp.linspace(2.0, 3.0, num=5).asnumpy()
match_array(actual, expected, error=6)
actual = onp.linspace(
2.0, 3.0, num=5, endpoint=False, dtype=onp.float32)
expected = mnp.linspace(2.0, 3.0, num=5, endpoint=False).asnumpy()
match_array(actual, expected, error=6)
actual = onp.linspace(2.0, 3.0, num=5, retstep=True, dtype=onp.float32)
expected = mnp.linspace(2.0, 3.0, num=5, retstep=True)
match_array(actual[0], expected[0].asnumpy())
assert actual[1] == expected[1].asnumpy()
actual = onp.linspace(2.0, [3, 4, 5], num=5,
endpoint=False, dtype=onp.float32)
expected = mnp.linspace(
2.0, [3, 4, 5], num=5, endpoint=False).asnumpy()
match_array(actual, expected, error=6)
start = onp.random.random([2, 1, 4]).astype("float32")
stop = onp.random.random([1, 5, 1]).astype("float32")
actual = onp.linspace(start, stop, num=20, retstep=True,
endpoint=False, dtype=onp.float32)
expected = mnp.linspace(mnp.asarray(start), mnp.asarray(stop), num=20,
retstep=True, endpoint=False)
match_array(actual[0], expected[0].asnumpy(), error=6)
match_array(actual[1], expected[1].asnumpy(), error=6)
actual = onp.linspace(start, stop, num=20, retstep=True,
endpoint=False, dtype=onp.int16)
expected = mnp.linspace(mnp.asarray(start), mnp.asarray(stop), num=20,
retstep=True, endpoint=False, dtype=mnp.int16)
match_array(actual[0], expected[0].asnumpy(), error=6)
match_array(actual[1], expected[1].asnumpy(), error=6)
for axis in range(2):
actual = onp.linspace(start, stop, num=20, retstep=False,
endpoint=False, dtype=onp.float32, axis=axis)
expected = mnp.linspace(mnp.asarray(start), mnp.asarray(stop), num=20,
retstep=False, endpoint=False, dtype=mnp.float32, axis=axis)
match_array(actual, expected.asnumpy(), error=6)
def test_absolute():
arr = rand_int(2, 3)
a = mnp.asarray(arr, dtype='float16')
b = mnp.asarray(arr, dtype='float32')
c = mnp.asarray(arr, dtype='uint8')
d = mnp.asarray(arr, dtype='bool')
match_array(mnp_absolute(a).asnumpy(), onp_absolute(a.asnumpy()))
match_array(mnp_absolute(b).asnumpy(), onp_absolute(b.asnumpy()))
match_array(mnp_absolute(c).asnumpy(), onp_absolute(c.asnumpy()))
match_array(mnp_absolute(d).asnumpy(), onp_absolute(d.asnumpy()))
where = rand_int(2, 3).astype('bool')
out = rand_int(2, 3)
match_array(mnp.absolute(a, out=mnp.asarray(out), where=mnp.asarray(where)).asnumpy(),
onp.absolute(a.asnumpy(), out=out, where=where))
def test_ndarray_transpose():
onp_array = onp.random.random((3, 4, 5)).astype('float32')
mnp_array = mnp.asarray(onp_array)
o_transposed = onp_ndarray_transpose(onp_array)
m_transposed = mnp_ndarray_transpose(mnp_array)
check_all_results(o_transposed, m_transposed)
def test_ndarray_flatten():
onp_array = onp.random.random((3, 4, 5)).astype('float32')
mnp_array = mnp.asarray(onp_array)
o_flatten = onp_ndarray_flatten(onp_array)
m_flatten = mnp_ndarray_flatten(mnp_array)
check_all_results(o_flatten, m_flatten)
def test_concatenate():
onp_array = onp.random.random((5, 4, 3, 2)).astype('float32')
mnp_array = mnp.asarray(onp_array)
o_concatenate = onp_concatenate(onp_array)
m_concatenate = mnp_concatenate(mnp_array)
check_all_results(o_concatenate, m_concatenate)
def test_ravel():
onp_array = onp.random.random((2, 3, 4)).astype('float32')
mnp_array = mnp.asarray(onp_array)
o_ravel = onp_ravel(onp_array)
m_ravel = mnp_ravel(mnp_array).asnumpy()
match_array(o_ravel, m_ravel)
def test_reshape():
onp_array = onp.random.random((2, 3, 4)).astype('float32')
mnp_array = mnp.asarray(onp_array)
o_reshaped = onp_reshape(onp_array)
m_reshaped = mnp_reshape(mnp_array)
check_all_results(o_reshaped, m_reshaped)