Skip to content

Propagate static shape in MaxAndArgmax #206

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Feb 6, 2023
Merged

Propagate static shape in MaxAndArgmax #206

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 2 additions & 11 deletions pytensor/tensor/math.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -142,15 +142,10 @@ def get_params(self, node):
def make_node(self, x):
x = as_tensor_variable(x)

# We keep the original broadcastable flags for dimensions on which
# we do not perform the max / argmax.
# Keep the original shapes for axes on which we do not perform the max/argmax.
all_axes = set(self.axis)
inputs = [x]
out_shape = tuple(
1 if s == 1 else None
for i, s in enumerate(x.type.shape)
if i not in all_axes
)
out_shape = tuple(s for i, s in enumerate(x.type.shape) if i not in all_axes)
outputs = [
tensor(dtype=x.type.dtype, shape=out_shape, name="max"),
tensor(dtype="int64", shape=out_shape, name="argmax"),
Expand Down Expand Up @@ -1521,7 +1516,6 @@ def perform(self, node, inp, out):
output[0] = np.asarray(np.mean(input, dtype="float64", axis=axis))

def c_code(self, node, name, inames, onames, sub):

ret = super().c_code(node, name, inames, onames, sub)

if self.axis is not None:
Expand Down Expand Up @@ -1940,7 +1934,6 @@ def perform(self, node, inp, out):
z[0] = np.asarray(np.dot(x, y))

def grad(self, inp, grads):

x, y = inp
(gz,) = grads
xdim, ydim, gdim = x.type.ndim, y.type.ndim, gz.type.ndim
Expand Down Expand Up @@ -2631,7 +2624,6 @@ def L_op(self, inp, out, grads):
# this handles inputs with zeros, but only certain input shapes
return [grad_case_without_zeros]
else:

where_zeros = eq(prod_in, 0.0)
sum_where_zeros = sum(where_zeros, axis=self.axis)
groups_with_single_zero = eq(sum_where_zeros, 1).dimshuffle(new_dims)
Expand Down Expand Up @@ -2924,7 +2916,6 @@ def _get_output_shape(cls, x1, x2, shapes, validate=False):
)
return x2_shape[:-2] + x1_shape[-2:-1] + x2_shape[-1:]
else:

if validate:
from pytensor.tensor.random.basic import broadcast_shapes

Expand Down
114 changes: 73 additions & 41 deletions tests/tensor/test_math.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -771,10 +771,9 @@ def test_basic_1(self):
v = eval_outputs(max_and_argmax(n)[0].shape)
assert len(v) == 0

def test_basic_2(self):
data = random(2, 3)
n = as_tensor_variable(data)
for (axis, np_axis) in [
@pytest.mark.parametrize(
"axis,np_axis",
[
(-1, -1),
(0, 0),
(1, 1),
Expand All @@ -783,19 +782,28 @@ def test_basic_2(self):
([1, 0], None),
(NoneConst.clone(), None),
(constant(0), 0),
]:
v, i = eval_outputs(max_and_argmax(n, axis))
assert i.dtype == "int64"
assert np.all(v == np.max(data, np_axis))
assert np.all(i == np.argmax(data, np_axis))
v_shape = eval_outputs(max_and_argmax(n, axis)[0].shape)
assert tuple(v_shape) == np.max(data, np_axis).shape

def test_basic_2_float16(self):
# Test negative values and bigger range to make sure numpy don't do the argmax as on uint16
data = (random(20, 30).astype("float16") - 0.5) * 20
n = shared(data)
for (axis, np_axis) in [
],
)
def test_basic_2(self, axis, np_axis):
data = random(2, 3)
n = as_tensor_variable(data)
# Test shape propagates (static & eval)
vt, it = max_and_argmax(n, axis)
np_max, np_argm = np.max(data, np_axis), np.argmax(data, np_axis)
assert vt.type.shape == np_max.shape
assert it.type.shape == np_argm.shape
v_shape, i_shape = eval_outputs([vt.shape, it.shape])
assert tuple(v_shape) == vt.type.shape
assert tuple(i_shape) == it.type.shape
# Test values
v, i = eval_outputs([vt, it])
assert i.dtype == "int64"
assert np.all(v == np_max)
assert np.all(i == np_argm)

@pytest.mark.parametrize(
"axis,np_axis",
[
(-1, -1),
(0, 0),
(1, 1),
Expand All @@ -804,13 +812,25 @@ def test_basic_2_float16(self):
([1, 0], None),
(NoneConst.clone(), None),
(constant(0), 0),
]:
v, i = eval_outputs(max_and_argmax(n, axis), (MaxAndArgmax,))
assert i.dtype == "int64"
assert np.all(v == np.max(data, np_axis))
assert np.all(i == np.argmax(data, np_axis))
v_shape = eval_outputs(max_and_argmax(n, axis)[0].shape)
assert tuple(v_shape) == np.max(data, np_axis).shape
],
)
def test_basic_2_float16(self, axis, np_axis):
# Test negative values and bigger range to make sure numpy don't do the argmax as on uint16
data = (random(20, 30).astype("float16") - 0.5) * 20
n = as_tensor_variable(data)
# Test shape propagates (static & eval)
vt, it = max_and_argmax(n, axis)
np_max, np_argm = np.max(data, np_axis), np.argmax(data, np_axis)
assert vt.type.shape == np_max.shape
assert it.type.shape == np_argm.shape
v_shape, i_shape = eval_outputs([vt.shape, it.shape])
assert tuple(v_shape) == vt.type.shape
assert tuple(i_shape) == it.type.shape
# Test values
v, i = eval_outputs([vt, it])
assert i.dtype == "int64"
assert np.all(v == np_max)
assert np.all(i == np_argm)

def test_basic_2_invalid(self):
n = as_tensor_variable(random(2, 3))
Expand Down Expand Up @@ -840,23 +860,33 @@ def test_basic_2_valid_neg(self):
v = eval_outputs(max_and_argmax(n, -2)[0].shape)
assert v == (3)

def test_basic_3(self):
data = random(2, 3, 4)
n = as_tensor_variable(data)
for (axis, np_axis) in [
@pytest.mark.parametrize(
"axis,np_axis",
[
(-1, -1),
(0, 0),
(1, 1),
(None, None),
([0, 1, 2], None),
([1, 2, 0], None),
]:
v, i = eval_outputs(max_and_argmax(n, axis))
assert i.dtype == "int64"
assert np.all(v == np.max(data, np_axis))
assert np.all(i == np.argmax(data, np_axis))
v = eval_outputs(max_and_argmax(n, axis)[0].shape)
assert tuple(v) == np.max(data, np_axis).shape
],
)
def test_basic_3(self, axis, np_axis):
data = random(2, 3, 4)
n = as_tensor_variable(data)
# Test shape propagates (static & eval)
vt, it = max_and_argmax(n, axis)
np_max, np_argm = np.max(data, np_axis), np.argmax(data, np_axis)
assert vt.type.shape == np_max.shape
assert it.type.shape == np_argm.shape
v_shape, i_shape = eval_outputs([vt.shape, it.shape])
assert tuple(v_shape) == vt.type.shape
assert tuple(i_shape) == it.type.shape
# Test values
v, i = eval_outputs([vt, it])
assert i.dtype == "int64"
assert np.all(v == np_max)
assert np.all(i == np_argm)

def test_arg_grad(self):
# The test checks that the gradient of argmax(x).sum() is 0
Expand Down Expand Up @@ -948,17 +978,19 @@ def test_preserve_broadcastable(self):
# Ensure the original broadcastable flags are preserved by Max/Argmax.
x = matrix().dimshuffle("x", 0, "x", 1, "x")
y = x.max(axis=1)
assert y.type.shape == (1, 1, None, 1)
assert y.type.broadcastable == (True, True, False, True)

def test_multiple_axes(self):
data = np.arange(24).reshape(3, 2, 4)
x = as_tensor_variable(data)
v, i = eval_outputs(max_and_argmax(x, [1, -1]))
vt, it = max_and_argmax(x, [1, -1])
assert vt.type.shape == it.type.shape == (3,)
v, i = eval_outputs([vt, it])
assert np.all(v == np.array([7, 15, 23]))
assert np.all(i == np.array([7, 7, 7]))

v = eval_outputs(max_and_argmax(x, [1, -1])[0].shape)
assert tuple(v) == np.max(data, (1, -1)).shape
v = eval_outputs(vt.shape)
assert tuple(v) == vt.type.shape

def test_zero_shape(self):
x = matrix()
Expand All @@ -972,8 +1004,8 @@ def test_zero_shape(self):
def test_numpy_input(self):
ar = np.array([1, 2, 3])
max_at, argmax_at = max_and_argmax(ar, axis=None)
assert max_at.eval(), 3
assert argmax_at.eval(), 2
assert max_at.eval() == 3
assert argmax_at.eval() == 2


class TestArgminArgmax:
Expand Down