Skip to content
This repository has been archived by the owner on Nov 17, 2023. It is now read-only.

Commit

Permalink
setup and concatenate, copy, expand_dims, expm1 (#16493)
Browse files Browse the repository at this point in the history 
norm, svd, split, squeeze, std, swapaxes, tensorfot, tile, trace, transpose, tril

fix swapaxes, sanity check issues

current svd not compatible

inputs of indentical axes supported for swapaxes, skip int tests of linalg.norm

re-factor the code structure

delete an extra blank line
  • Loading branch information
@xiezhq-hermann @reminisce
xiezhq-hermann authored and reminisce committed Oct 20, 2019
1 parent 5b67a69 commit c1d02ce
Show file tree
Hide file tree
Showing 2 changed files with 223 additions and 56 deletions.
3 changes: 3 additions & 0 deletions python/mxnet/numpy_dispatch_protocol.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -118,6 +118,9 @@ def _run_with_array_ufunc_proto(*args, **kwargs):
'unique',
'var',
'zeros_like',
'linalg.norm',
'trace',
'tril',
'meshgrid',
'outer',
'einsum'
Expand Down
276 changes: 220 additions & 56 deletions tests/python/unittest/test_numpy_interoperability.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,7 @@
from mxnet import np
from mxnet.test_utils import assert_almost_equal
from mxnet.test_utils import use_np
from mxnet.test_utils import is_op_runnable
from common import assertRaises, with_seed
from mxnet.numpy_dispatch_protocol import with_array_function_protocol, with_array_ufunc_protocol
from mxnet.numpy_dispatch_protocol import _NUMPY_ARRAY_FUNCTION_LIST, _NUMPY_ARRAY_UFUNC_LIST
Expand All @@ -48,6 +49,201 @@ def get_workloads(name):
return OpArgMngr._args.get(name, None)


def _add_workload_concatenate(array_pool):
OpArgMngr.add_workload('concatenate', [array_pool['4x1'], array_pool['4x1']])
OpArgMngr.add_workload('concatenate', [array_pool['4x1'], array_pool['4x1']], axis=1)
OpArgMngr.add_workload('concatenate', [np.random.uniform(size=(3, 3))])
OpArgMngr.add_workload('concatenate', (np.arange(4).reshape((2, 2)), np.arange(4).reshape((2, 2))))
OpArgMngr.add_workload('concatenate', (np.arange(4),))
OpArgMngr.add_workload('concatenate', (np.array(np.arange(4)),))
OpArgMngr.add_workload('concatenate', (np.arange(4), np.arange(3)))
OpArgMngr.add_workload('concatenate', (np.array(np.arange(4)), np.arange(3)))
OpArgMngr.add_workload('concatenate', (np.arange(4), np.arange(3)), axis=0)
OpArgMngr.add_workload('concatenate', (np.arange(4), np.arange(3)), axis=-1)
a23 = np.random.uniform(size=(2, 3))
a13 = np.random.uniform(size=(1, 3))
OpArgMngr.add_workload('concatenate', (a23, a13))
OpArgMngr.add_workload('concatenate', (a23, a13), axis=0)
OpArgMngr.add_workload('concatenate', (a23.T, a13.T), axis=1)
OpArgMngr.add_workload('concatenate', (a23.T, a13.T), axis=-1)
res = np.arange(2*3*7).reshape((2, 3, 7))
a0 = res[..., :4]
a1 = res[..., 4:6]
a2 = res[..., 6:]
OpArgMngr.add_workload('concatenate', (a0, a1, a2), axis=2)
OpArgMngr.add_workload('concatenate', (a0, a1, a2), axis=-1)
OpArgMngr.add_workload('concatenate', (a0.T, a1.T, a2.T), axis=0)
out = np.empty(4, np.float32)
OpArgMngr.add_workload('concatenate', (np.array([1, 2]), np.array([3, 4])), out=out)


def _add_workload_copy():
OpArgMngr.add_workload('copy', np.random.uniform(size=(4, 1)))
OpArgMngr.add_workload('copy', np.random.uniform(size=(2, 2)))
OpArgMngr.add_workload('copy', np.random.uniform(size=(2,2)))


def _add_workload_expand_dims():
OpArgMngr.add_workload('expand_dims', np.random.uniform(size=(4, 1)), -1)
for axis in range(-5, 4):
OpArgMngr.add_workload('expand_dims', np.empty((2, 3, 4, 5)), axis)


def _add_workload_split():
OpArgMngr.add_workload('split', np.random.uniform(size=(4, 1)), 2)
OpArgMngr.add_workload('split', np.arange(10), 2)
assertRaises(ValueError, np.split, np.arange(10), 3)


def _add_workload_squeeze():
OpArgMngr.add_workload('squeeze', np.random.uniform(size=(4, 1)))
OpArgMngr.add_workload('squeeze', np.random.uniform(size=(20, 10, 10, 1, 1)))
OpArgMngr.add_workload('squeeze', np.random.uniform(size=(20, 1, 10, 1, 20)))
OpArgMngr.add_workload('squeeze', np.random.uniform(size=(1, 1, 20, 10)))
OpArgMngr.add_workload('squeeze', np.array([[[1.5]]]))


def _add_workload_std():
OpArgMngr.add_workload('std', np.random.uniform(size=(4, 1)))
A = np.array([[1, 2, 3], [4, 5, 6]])
OpArgMngr.add_workload('std', A)
OpArgMngr.add_workload('std', A, 0)
OpArgMngr.add_workload('std', A, 1)
OpArgMngr.add_workload('std', np.array([1, -1, 1, -1]))
OpArgMngr.add_workload('std', np.array([1, -1, 1, -1]), ddof=1)
OpArgMngr.add_workload('std', np.array([1, -1, 1, -1]), ddof=2)
OpArgMngr.add_workload('std', np.arange(10), out=np.array(0.))


def _add_workload_swapaxes():
OpArgMngr.add_workload('swapaxes', np.random.uniform(size=(4, 1)), 0, 1)
OpArgMngr.add_workload('swapaxes', np.array([[[0, 1], [2, 3]], [[4, 5], [6, 7]]]), 0, 2)
a = np.arange(1*2*3*4).reshape(1, 2, 3, 4).copy()
b = a.copy()
# no AxisError defined in mxnet numpy
# assertRaises(np.AxisError, np.swapaxes, -5, 0)
for i in range(-4, 4):
for j in range(-4, 4):
for k, src in enumerate((a, b)):
OpArgMngr.add_workload('swapaxes', src, i, j)


def _add_workload_tensordot():
OpArgMngr.add_workload('tensordot', np.random.uniform(size=(4, 1)), np.random.uniform(size=(4, 1)))
OpArgMngr.add_workload('tensordot', np.random.uniform(size=(3, 0)), np.random.uniform(size=(0, 4)), (1, 0))
OpArgMngr.add_workload('tensordot', np.array(1), np.array(1), ([], []))


def _add_workload_tile():
OpArgMngr.add_workload('tile', np.random.uniform(size=(4, 1)), 2)
a = np.array([0, 1, 2])
b = np.array([[1, 2], [3, 4]])
OpArgMngr.add_workload('tile', a, 2)
OpArgMngr.add_workload('tile', a, (2, 2))
OpArgMngr.add_workload('tile', a, (1, 2))
OpArgMngr.add_workload('tile', b, 2)
OpArgMngr.add_workload('tile', b, (2, 1))
OpArgMngr.add_workload('tile', b, (2, 2))
OpArgMngr.add_workload('tile', np.arange(5), 1)
OpArgMngr.add_workload('tile', np.array([[], []]), 2)
OpArgMngr.add_workload('tile', np.array([[[]]]), (3, 2, 5))
reps = [(2,), (1, 2), (2, 1), (2, 2), (2, 3, 2), (3, 2)]
shape = [(3,), (2, 3), (3, 4, 3), (3, 2, 3), (4, 3, 2, 4), (2, 2)]
for s in shape:
b = np.random.randint(0, 10, size=s)
for r in reps:
# RuntimeError to be tracked
# where s = (3, 4, 3), r = (2, 3, 2)
# OpArgMngr.add_workload('tile', b, r)
pass


def _add_workload_transpose():
OpArgMngr.add_workload('transpose', np.random.uniform(size=(4, 1)))
OpArgMngr.add_workload('transpose', np.array([[]]))
OpArgMngr.add_workload('transpose', np.array([[1, 2]]))
OpArgMngr.add_workload('transpose', np.array([[1, 2, 3], [4, 5, 6]]))
OpArgMngr.add_workload('transpose', np.array([[1, 2], [3, 4], [5, 6]]), (1, 0))
OpArgMngr.add_workload('transpose', np.array([[1, 2], [3, 4]]))


def _add_workload_linalg_norm():
OpArgMngr.add_workload('linalg.norm', np.random.uniform(size=(4, 1)))
for dt in ["double", "float32", "int64"]:
OpArgMngr.add_workload('linalg.norm', np.array([], dtype=dt))
OpArgMngr.add_workload('linalg.norm', np.array([np.array([]), np.array([])], dtype=dt))
# numerical error exceed the tolerance
if dt == "int64":
continue
for v in ([1, 2, 3, 4], [-1, -2, -3, -4], [-1, 2, -3, 4]):
OpArgMngr.add_workload('linalg.norm', np.array(v, dtype=dt))
A = np.array([[1, 2, 3], [4, 5, 6]], dtype=dt)
[OpArgMngr.add_workload('linalg.norm', A[:, k]) for k in range(A.shape[1])]
OpArgMngr.add_workload('linalg.norm', A, axis=0)
[OpArgMngr.add_workload('linalg.norm', A[k, :]) for k in range(A.shape[0])]
OpArgMngr.add_workload('linalg.norm', A, axis=1)
B = np.arange(1, 25).reshape(2, 3, 4).astype(dt)
for axis in itertools.combinations(range(-B.ndim, B.ndim), 2):
row_axis, col_axis = axis
if row_axis < 0:
row_axis += B.ndim
if col_axis < 0:
col_axis += B.ndim
if row_axis == col_axis:
# improper assertion behavior
# assertRaises(ValueError, np.linalg.norm, B, axis=axis)
pass
else:
OpArgMngr.add_workload('linalg.norm', B, axis=axis)
k_index = B.ndim - row_axis - col_axis
for k in range(B.shape[k_index]):
if row_axis < col_axis:
OpArgMngr.add_workload('linalg.norm', np.take(B[:], np.array(k), axis=k_index))
else:
OpArgMngr.add_workload('linalg.norm', np.take(B[:], np.array(k), axis=k_index).T)
A = np.arange(1, 25, dtype=dt).reshape(2, 3, 4)
OpArgMngr.add_workload('linalg.norm', A, ord=None, axis=None)
OpArgMngr.add_workload('linalg.norm', A, ord=None,axis=None, keepdims=True)
for k in range(A.ndim):
OpArgMngr.add_workload('linalg.norm', A, axis=k)
OpArgMngr.add_workload('linalg.norm', A, axis=k, keepdims=True)
for k in itertools.permutations(range(A.ndim), 2):
OpArgMngr.add_workload('linalg.norm', A, axis=k)
OpArgMngr.add_workload('linalg.norm', A, axis=k, keepdims=True)
OpArgMngr.add_workload('linalg.norm', np.array([[]], dtype=dt))
A = np.array([[1, 3], [5, 7]], dtype=dt)
OpArgMngr.add_workload('linalg.norm', A)
OpArgMngr.add_workload('linalg.norm', A, 'fro')
A = (1 / 10) * np.array([[1, 2, 3], [6, 0, 5], [3, 2, 1]], dtype=dt)
OpArgMngr.add_workload('linalg.norm', A)
OpArgMngr.add_workload('linalg.norm', A, 'fro')
for dt in [np.float16, np.float32, np.float64]:
OpArgMngr.add_workload('linalg.norm', np.array([[1, 0, 1], [0, 1, 1]], dtype=dt))
OpArgMngr.add_workload('linalg.norm', np.array([[1, 0, 1], [0, 1, 1]], dtype=dt), 'fro')


def _add_workload_trace():
OpArgMngr.add_workload('trace', np.random.uniform(size=(4, 1)))
OpArgMngr.add_workload('trace', np.random.uniform(size=(3, 2)))


def _add_workload_tril():
OpArgMngr.add_workload('tril', np.random.uniform(size=(4, 1)))
for dt in ['float16', 'float32', 'float64', 'int32', 'int64', 'int8', 'uint8']:
OpArgMngr.add_workload('tril', np.ones((2, 2), dtype=dt))
a = np.array([
[[1, 1], [1, 1]],
[[1, 1], [1, 0]],
[[1, 1], [0, 0]],
], dtype=dt)
OpArgMngr.add_workload('tril', a)
arr = np.array([[1, 1, np.inf],
[1, 1, 1],
[np.inf, 1, 1]])
OpArgMngr.add_workload('tril', arr)
OpArgMngr.add_workload('tril', np.zeros((3, 3), dtype=dt))


def _add_workload_einsum():
chars = 'abcdefghij'
sizes = [2, 3, 4, 5, 4, 3, 2, 6, 5, 4]
Expand Down Expand Up @@ -120,6 +316,15 @@ def _add_workload_einsum():
OpArgMngr.add_workload('einsum', subscripts, *args, optimize=optimize)


def _add_workload_expm1():
OpArgMngr.add_workload('expm1', np.random.uniform(size=(4, 1)))
OpArgMngr.add_workload('expm1', np.random.uniform(size=(1, 1)))
OpArgMngr.add_workload('expm1', np.array([np.inf]))
OpArgMngr.add_workload('expm1', np.array([-np.inf]))
OpArgMngr.add_workload('expm1', np.array([0.]))
OpArgMngr.add_workload('expm1', np.array([-0.]))


def _add_workload_argmax():
OpArgMngr.add_workload('argmax', np.random.uniform(size=(4, 5, 6, 7, 8)), 0)
OpArgMngr.add_workload('argmax', np.random.uniform(size=(4, 5, 6, 7, 8)), 1)
Expand Down Expand Up @@ -179,15 +384,6 @@ def _add_workload_clip():
OpArgMngr.add_workload('clip', np.array([-2., np.nan, 0.5, 3., 0.25, np.nan]), -1, 1)


def _add_workload_concatenate(array_pool):
OpArgMngr.add_workload('concatenate', [array_pool['4x1'], array_pool['4x1']])
OpArgMngr.add_workload('concatenate', [array_pool['4x1'], array_pool['4x1']], axis=1)


def _add_workload_copy(array_pool):
OpArgMngr.add_workload('copy', array_pool['4x1'])


def _add_workload_cumsum():
for ctype in _DTYPES:
OpArgMngr.add_workload('cumsum', np.array([1, 2, 10, 11, 6, 5, 4], dtype=ctype))
Expand Down Expand Up @@ -220,10 +416,6 @@ def _add_workload_dot():
OpArgMngr.add_workload('dot', np.random.normal(size=(1024, 16)), np.random.normal(size=(16, 32)))


def _add_workload_expand_dims(array_pool):
OpArgMngr.add_workload('expand_dims', array_pool['4x1'], -1)


def _add_workload_fix():
OpArgMngr.add_workload('fix', np.array([[1.0, 1.1, 1.5, 1.8], [-1.0, -1.1, -1.5, -1.8]]))
OpArgMngr.add_workload('fix', np.array([3.14]))
Expand Down Expand Up @@ -341,22 +533,10 @@ def _add_workload_roll():
OpArgMngr.add_workload('roll', np.array([]), 1)


def _add_workload_split(array_pool):
OpArgMngr.add_workload('split', array_pool['4x1'], 2)


def _add_workload_squeeze(array_pool):
OpArgMngr.add_workload('squeeze', array_pool['4x1'])


def _add_workload_stack(array_pool):
OpArgMngr.add_workload('stack', [array_pool['4x1']] * 2)


def _add_workload_std(array_pool):
OpArgMngr.add_workload('std', array_pool['4x1'])


def _add_workload_sum():
# OpArgMngr.add_workload('sum', np.ones(101, dtype=bool))
OpArgMngr.add_workload('sum', np.arange(1, 10).reshape((3, 3)), axis=1, keepdims=True)
Expand All @@ -383,10 +563,6 @@ def _add_workload_sum():
# OpArgMngr.add_workload('sum', np.array([0.2]), initial=0.1)


def _add_workload_swapaxes(array_pool):
OpArgMngr.add_workload('swapaxes', array_pool['4x1'], 0, 1)


def _add_workload_take():
OpArgMngr.add_workload('take', np.array([[1, 2], [3, 4]], dtype=int), np.array([], int))
for mode in ['wrap', 'clip']:
Expand All @@ -403,19 +579,6 @@ def _add_workload_take():
OpArgMngr.add_workload('take', x, np.array([3], int), axis=0, mode='wrap')


def _add_workload_tensordot(array_pool):
OpArgMngr.add_workload('tensordot', array_pool['4x1'], array_pool['4x1'])


def _add_workload_tile(array_pool):
OpArgMngr.add_workload('tile', array_pool['4x1'], 2)
OpArgMngr.add_workload('tile', np.array([[[]]]), (3, 2, 5))


def _add_workload_transpose(array_pool):
OpArgMngr.add_workload('transpose', array_pool['4x1'])


def _add_workload_unique():
OpArgMngr.add_workload('unique', np.array([5, 7, 1, 2, 1, 5, 7]*10), True, True, True)
OpArgMngr.add_workload('unique', np.array([]), True, True, True)
Expand Down Expand Up @@ -683,10 +846,6 @@ def _add_workload_log10(array_pool):
OpArgMngr.add_workload('log10', array_pool['4x1'])


def _add_workload_expm1(array_pool):
OpArgMngr.add_workload('expm1', array_pool['4x1'])


def _add_workload_sqrt():
OpArgMngr.add_workload('sqrt', np.array([1, np.PZERO, np.NZERO, np.inf, np.nan]))

Expand Down Expand Up @@ -788,11 +947,11 @@ def _prepare_workloads():
_add_workload_broadcast_to()
_add_workload_clip()
_add_workload_concatenate(array_pool)
_add_workload_copy(array_pool)
_add_workload_copy()
_add_workload_cumsum()
_add_workload_ravel()
_add_workload_dot()
_add_workload_expand_dims(array_pool)
_add_workload_expand_dims()
_add_workload_fix()
_add_workload_flip()
_add_workload_max(array_pool)
Expand All @@ -804,19 +963,22 @@ def _prepare_workloads():
_add_workload_reshape()
_add_workload_rint(array_pool)
_add_workload_roll()
_add_workload_split(array_pool)
_add_workload_squeeze(array_pool)
_add_workload_split()
_add_workload_squeeze()
_add_workload_stack(array_pool)
_add_workload_std(array_pool)
_add_workload_std()
_add_workload_sum()
_add_workload_swapaxes(array_pool)
_add_workload_swapaxes()
_add_workload_take()
_add_workload_tensordot(array_pool)
_add_workload_tile(array_pool)
_add_workload_transpose(array_pool)
_add_workload_tensordot()
_add_workload_tile()
_add_workload_transpose()
_add_workload_unique()
_add_workload_var(array_pool)
_add_workload_zeros_like(array_pool)
_add_workload_linalg_norm()
_add_workload_trace()
_add_workload_tril()
_add_workload_outer()
_add_workload_meshgrid()
_add_workload_einsum()
Expand Down Expand Up @@ -845,7 +1007,7 @@ def _prepare_workloads():
_add_workload_log2(array_pool)
_add_workload_log1p()
_add_workload_log10(array_pool)
_add_workload_expm1(array_pool)
_add_workload_expm1()
_add_workload_sqrt()
_add_workload_square()
_add_workload_cbrt()
Expand Down Expand Up @@ -890,6 +1052,8 @@ def _check_interoperability_helper(op_name, *args, **kwargs):
onp_op = getattr(getattr(_np, strs[0]), strs[1])
else:
assert False
if not is_op_runnable():
return
out = onp_op(*args, **kwargs)
expected_out = _get_numpy_op_output(onp_op, *args, **kwargs)
if isinstance(out, (tuple, list)):
Expand Down

0 comments on commit c1d02ce

Please sign in to comment.