def test_extra_ops():
a = matrix("a")
a.tag.test_value = np.arange(6, dtype=config.floatX).reshape((3, 2))
out = aet_extra_ops.cumsum(a, axis=0)
fgraph = FunctionGraph([a], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
out = aet_extra_ops.cumprod(a, axis=1)
fgraph = FunctionGraph([a], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
out = aet_extra_ops.diff(a, n=2, axis=1)
fgraph = FunctionGraph([a], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
out = aet_extra_ops.repeat(a, (3, 3), axis=1)
fgraph = FunctionGraph([a], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
# This function also cannot take symbolic input.
c = aet.as_tensor(5)
out = aet_extra_ops.bartlett(c)
fgraph = FunctionGraph([], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
with pytest.raises(NotImplementedError):
out = aet_extra_ops.fill_diagonal(a, c)
fgraph = FunctionGraph([a], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
with pytest.raises(NotImplementedError):
out = aet_extra_ops.fill_diagonal_offset(a, c, c)
fgraph = FunctionGraph([a], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
with pytest.raises(NotImplementedError):
out = aet_extra_ops.Unique(axis=1)(a)
fgraph = FunctionGraph([a], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
indices = np.arange(np.product((3, 4)))
out = aet_extra_ops.unravel_index(indices, (3, 4), order="C")
fgraph = FunctionGraph([], out)
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs],
must_be_device_array=False)
multi_index = np.unravel_index(np.arange(np.product((3, 4))), (3, 4))
out = aet_extra_ops.ravel_multi_index(multi_index, (3, 4))
fgraph = FunctionGraph([], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs],
must_be_device_array=False)
# The inputs are "concrete", yet it still has problems?
out = aet_extra_ops.Unique()(aet.as_tensor(
np.arange(6, dtype=config.floatX).reshape((3, 2))))
fgraph = FunctionGraph([], [out])
compare_jax_and_py(fgraph, [])
def check(shape, index_ndim, order):
indices = np.arange(np.product(shape))
# test with scalars and higher-dimensional indices
if index_ndim == 0:
indices = indices[-1]
elif index_ndim == 2:
indices = indices[:, np.newaxis]
indices_symb = aesara.shared(indices)
# reference result
ref = np.unravel_index(indices, shape, order=order)
def fn(i, d):
return function([], unravel_index(i, d, order=order))
# shape given as a tuple
f_array_tuple = fn(indices, shape)
f_symb_tuple = fn(indices_symb, shape)
np.testing.assert_equal(ref, f_array_tuple())
np.testing.assert_equal(ref, f_symb_tuple())
# shape given as an array
shape_array = np.array(shape)
f_array_array = fn(indices, shape_array)
np.testing.assert_equal(ref, f_array_array())
# shape given as an Aesara variable
shape_symb = aesara.shared(shape_array)
f_array_symb = fn(indices, shape_symb)
np.testing.assert_equal(ref, f_array_symb())
# shape given as a Shape op (unravel_index will use get_vector_length
# to infer the number of dimensions)
indexed_array = aesara.shared(np.random.uniform(size=shape_array))
f_array_shape = fn(indices, indexed_array.shape)
np.testing.assert_equal(ref, f_array_shape())
# shape testing
self._compile_and_check(
[],
unravel_index(indices, shape_symb, order=order),
[],
UnravelIndex,
)
def test_extra_ops():
a = matrix("a")
a.tag.test_value = np.arange(6, dtype=config.floatX).reshape((3, 2))
out = at_extra_ops.cumsum(a, axis=0)
fgraph = FunctionGraph([a], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
out = at_extra_ops.cumprod(a, axis=1)
fgraph = FunctionGraph([a], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
out = at_extra_ops.diff(a, n=2, axis=1)
fgraph = FunctionGraph([a], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
out = at_extra_ops.repeat(a, (3, 3), axis=1)
fgraph = FunctionGraph([a], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
c = at.as_tensor(5)
with pytest.raises(NotImplementedError):
out = at_extra_ops.fill_diagonal(a, c)
fgraph = FunctionGraph([a], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
with pytest.raises(NotImplementedError):
out = at_extra_ops.fill_diagonal_offset(a, c, c)
fgraph = FunctionGraph([a], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
with pytest.raises(NotImplementedError):
out = at_extra_ops.Unique(axis=1)(a)
fgraph = FunctionGraph([a], [out])
compare_jax_and_py(fgraph, [get_test_value(i) for i in fgraph.inputs])
indices = np.arange(np.product((3, 4)))
out = at_extra_ops.unravel_index(indices, (3, 4), order="C")
fgraph = FunctionGraph([], out)
compare_jax_and_py(
fgraph, [get_test_value(i) for i in fgraph.inputs], must_be_device_array=False
)
def fn(i, d):
return function([], unravel_index(i, d, order=order))
def test_unravel_index(self):
def check(shape, index_ndim, order):
indices = np.arange(np.product(shape))
# test with scalars and higher-dimensional indices
if index_ndim == 0:
indices = indices[-1]
elif index_ndim == 2:
indices = indices[:, np.newaxis]
indices_symb = aesara.shared(indices)
# reference result
ref = np.unravel_index(indices, shape, order=order)
def fn(i, d):
return function([], unravel_index(i, d, order=order))
# shape given as a tuple
f_array_tuple = fn(indices, shape)
f_symb_tuple = fn(indices_symb, shape)
np.testing.assert_equal(ref, f_array_tuple())
np.testing.assert_equal(ref, f_symb_tuple())
# shape given as an array
shape_array = np.array(shape)
f_array_array = fn(indices, shape_array)
np.testing.assert_equal(ref, f_array_array())
# shape given as an Aesara variable
shape_symb = aesara.shared(shape_array)
f_array_symb = fn(indices, shape_symb)
np.testing.assert_equal(ref, f_array_symb())
# shape given as a Shape op (unravel_index will use get_vector_length
# to infer the number of dimensions)
indexed_array = aesara.shared(np.random.uniform(size=shape_array))
f_array_shape = fn(indices, indexed_array.shape)
np.testing.assert_equal(ref, f_array_shape())
# shape testing
self._compile_and_check(
[],
unravel_index(indices, shape_symb, order=order),
[],
UnravelIndex,
)
for order in ("C", "F"):
for index_ndim in (0, 1, 2):
check((3, ), index_ndim, order)
check((3, 4), index_ndim, order)
check((3, 4, 5), index_ndim, order)
# must specify ndim if length of dims is not fixed
with pytest.raises(ValueError):
unravel_index(ivector(), ivector())
# must provide integers
with pytest.raises(TypeError):
unravel_index(fvector(), (3, 4))
with pytest.raises(TypeError):
unravel_index((3, 4), (3.4, 3.2))
# dims must be a 1D sequence
with pytest.raises(TypeError):
unravel_index((3, 4), 3)
with pytest.raises(TypeError):
unravel_index((3, 4), ((3, 4), ))
