Skip to content

add support for tile #317

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

Closed
Bahaatbb wants to merge 5 commits into from
Closed

add support for tile #317

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
fc14b6a
add supprot for tile
Bahaatbb Dec 29, 2023
c77a77b
Merge branch 'ml-explore:main' into tile-as-op
Bahaatbb Dec 31, 2023
583d373
Improved tile using broadcasting
Bahaatbb Dec 31, 2023
5a84e8b
Merge branch 'main' into tile-as-op
Bahaatbb Jan 3, 2024
d9a6b54
resolved some conflicts
Bahaatbb Jan 3, 2024
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
1 change: 1 addition & 0 deletions ACKNOWLEDGMENTS.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,7 @@ MLX was developed with contributions from the following individuals:
- Juarez Bochi: Fixed bug in cross attention.
- Justin Deschenaux: Sine, Cosine, arange, randint, truncated normal, bernoulli, lion optimizer, Dropout2d, linear and logistic regression python example.
- Diogo Da Cruz: Added `tri`, `tril`, `triu`, `tensordot` and safetensor support
- Bahaa Eddin Tabbakha: Added `repeat`, `tile`.

# Third-Party Software

Expand Down
1 change: 1 addition & 0 deletions docs/src/python/ops.rst
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -104,6 +104,7 @@ Operations
take_along_axis
tan
tanh
tile
tensordot
transpose
tri
Expand Down
37 changes: 37 additions & 0 deletions mlx/ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -753,6 +753,43 @@ array repeat(const array& arr, int repeats, StreamOrDevice s) {
return repeat(flatten(arr, s), repeats, 0, s);
}

/** Replicating array a specified number of times along each axis */
array tile(const array& a, std::vector<int> repeats, StreamOrDevice s) {
if (repeats.size() == 0) {
return a;
}

for (size_t i = 0; i < repeats.size(); i++) {
if (repeats[i] < 0) {
throw std::invalid_argument(
"[tile] Negative repeat value found at index " + std::to_string(i));

@angeloskath angeloskath Jan 3, 2024

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Extremely nitpicky, feel free to ignore. Using an ostringstream would be better as it would allow you to also provide the whole repeats vector. Something like:

std::ostringstream msg;
msg << "[tile] Repeats argument cannot contain negative values but "
    << repeats << " was provided.";
throw std::invalid_argument(msg.str());

}
}

int d = repeats.size();
array arr = copy(a, s);

@angeloskath angeloskath Jan 3, 2024

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copy is not needed here, you can simply do array arr = a;.


if (d < arr.ndim()) {
repeats.insert(repeats.begin(), arr.ndim() - d, 1);
} else if (d > arr.ndim()) {
arr = expand_dims(arr, std::vector<int>(d - arr.ndim(), 0), s);
}

for (size_t i = 0; i < arr.shape().size(); i++) {
if (repeats[i] != 1) {
arr = expand_dims(arr, i, s);
std::vector<int> new_shape(arr.shape());
new_shape[i] *= repeats[i];
arr = broadcast_to(arr, new_shape, s);
new_shape[i] = -1;
new_shape.erase(new_shape.begin() + i + 1);
arr = reshape(arr, new_shape, s);
}
}

@angeloskath angeloskath Jan 3, 2024

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is already much better than before, however again each of these reshapes will be an actual copy and there is no need to launch many. Basically this is like calling repeat() many times.

The better way to do it is to keep 3 vectors.

  • a reshape vector which will do all the expands at once (start with a.shape() and add 1s where needed)
  • a broadcast to vector which will have the number of repeats where 1s were added in the above
  • final reshape vector where the repeats are merged to the axis to their left

After that the implementation is as simple as the following:

auto out = reshape(a, expand_shape, s);
out = broadcast_to(out, repeat_shape, s);
out = reshape(out, final_shape, s);


return arr;
}

/** Pad an array with a constant value */
array pad(
const array& a,
Expand Down
3 changes: 3 additions & 0 deletions mlx/ops.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -218,6 +218,9 @@ array stack(const std::vector<array>& arrays, StreamOrDevice s = {});
array repeat(const array& arr, int repeats, int axis, StreamOrDevice s = {});
array repeat(const array& arr, int repeats, StreamOrDevice s = {});

/** Replicating array a specified number of times along each axis */
array tile(const array& arr, std::vector<int> repeats, StreamOrDevice s = {});

/** Permutes the dimensions according to the given axes. */
array transpose(const array& a, std::vector<int> axes, StreamOrDevice s = {});
inline array transpose(
Expand Down
23 changes: 23 additions & 0 deletions python/src/ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2439,6 +2439,29 @@ void init_ops(py::module_& m) {
Returns:
array: The resulting repeated array.
)pbdoc");
m.def(
"tile",
[](const array& array, std::vector<int> repeats, StreamOrDevice s) {
return tile(array, repeats, s);
},
"array"_a,
py::pos_only(),
"repeats"_a,
py::kw_only(),
"stream"_a = none,
R"pbdoc(
tile(array: array, repeats: array, *, stream: Union[None, Stream, Device] = None) -> array

Replicating array a specified number of times along each axis.

Args:
array (array): Input array.
repeats (array): The number of repetitions along each axis.
stream (Stream, optional): Stream or device. Defaults to ``None``.

Returns:
array: The resulting tiled array.
)pbdoc");
m.def(
"clip",
[](const array& a,
Expand Down
49 changes: 48 additions & 1 deletion python/tests/test_ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1518,7 +1518,10 @@ def test_linspace(self):

def test_repeat(self):
# Setup data for the tests
data = np.array([[[13, 3], [16, 6]], [[14, 4], [15, 5]], [[11, 1], [12, 2]]])
data = np.array(
[[[13, 3], [16, 6]], [[14, 4], [15, 5]], [[11, 1], [12, 2]]],
dtype=np.float32,
)
# Test repeat along axis 0
repeat_axis_0 = mx.repeat(mx.array(data), 2, axis=0)
expected_axis_0 = np.repeat(data, 2, axis=0)
Expand Down Expand Up @@ -1547,6 +1550,50 @@ def test_repeat(self):
expected_3 = np.repeat(data_3, 2, axis=0)
self.assertEqualArray(repeat_3, mx.array(expected_3))

def test_tile(self):
data = np.array(
[
[[6.0, 10.0], [1.0, 3.0]],
[[9.0, 5.0], [9.0, 2.0]],
[[5.0, 3.0], [3.0, 4.0]],
],
dtype=np.float32,
)

expected_1 = mx.tile(mx.array(data), [3])
tiled_1 = np.tile(data, [3])
self.assertEqualArray(expected_1, mx.array(tiled_1))

data_2 = np.array(
[
[[4.0, 4.0], [9.0, 7.0]],
[[7.0, 3.0], [6.0, 8.0]],
[[8.0, 3.0], [8.0, 7.0]],
],
dtype=np.float32,
)

expected_2 = mx.tile(mx.array(data_2), [1, 3, 2])
tiled_2 = np.tile(data_2, [1, 3, 2])
self.assertEqualArray(expected_2, mx.array(tiled_2))

expected_3 = mx.tile(mx.array(data_2), [3, 2, 2])
tiled_3 = np.tile(data_2, (3, 2, 2))
self.assertEqualArray(expected_3, mx.array(tiled_3))

expected_4 = mx.tile(mx.array(data_2), [3, 2, 2, 2])
tiled_4 = np.tile(data_2, (3, 2, 2, 2))
self.assertEqualArray(expected_4, mx.array(tiled_4))

expected_5 = mx.tile(mx.array(data_2), [1, 1, 1])
tiled_5 = np.tile(data_2, (1, 1, 1))
self.assertEqualArray(expected_5, mx.array(tiled_5))

expected_6 = mx.tile(mx.array(data_2), [1, 1, 1, 1])
tiled_6 = np.tile(data_2, (1, 1, 1, 1))
self.assertEqualArray(expected_6, mx.array(tiled_6))


def test_tensordot(self):
x = mx.arange(60.0).reshape(3, 4, 5)
y = mx.arange(24.0).reshape(4, 3, 2)
Expand Down
76 changes: 75 additions & 1 deletion tests/ops_tests.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2279,6 +2279,80 @@ TEST_CASE("test repeat") {
CHECK_THROWS_AS(repeat(data_3, -3, 0), std::invalid_argument);
}

TEST_CASE("test tile") {
auto data =
array({6., 10., 1., 3., 9., 5., 9., 2., 5., 3., 3., 4.}, {3, 2, 2});
auto expected_1 = array(
{6., 10., 6., 10., 6., 10., 1., 3., 1., 3., 1., 3.,
9., 5., 9., 5., 9., 5., 9., 2., 9., 2., 9., 2.,
5., 3., 5., 3., 5., 3., 3., 4., 3., 4., 3., 4.},
{3, 2, 6});

auto tiled_1 = tile(data, {3});

auto data_2 =
array({4., 4., 9., 7., 7., 3., 6., 8., 8., 3., 8., 7.}, {3, 2, 2});

auto expected_2 = array(
{4., 4., 4., 4., 9., 7., 9., 7., 4., 4., 4., 4., 9., 7., 9., 7., 4., 4.,
4., 4., 9., 7., 9., 7., 7., 3., 7., 3., 6., 8., 6., 8., 7., 3., 7., 3.,
6., 8., 6., 8., 7., 3., 7., 3., 6., 8., 6., 8., 8., 3., 8., 3., 8., 7.,
8., 7., 8., 3., 8., 3., 8., 7., 8., 7., 8., 3., 8., 3., 8., 7., 8., 7.},
{3, 6, 4});

auto tiled_2 = tile(data_2, {3, 2});

auto expected_3 = array(
{4., 4., 4., 4., 9., 7., 9., 7., 4., 4., 4., 4., 9., 7., 9., 7., 7., 3.,
7., 3., 6., 8., 6., 8., 7., 3., 7., 3., 6., 8., 6., 8., 8., 3., 8., 3.,
8., 7., 8., 7., 8., 3., 8., 3., 8., 7., 8., 7., 4., 4., 4., 4., 9., 7.,
9., 7., 4., 4., 4., 4., 9., 7., 9., 7., 7., 3., 7., 3., 6., 8., 6., 8.,
7., 3., 7., 3., 6., 8., 6., 8., 8., 3., 8., 3., 8., 7., 8., 7., 8., 3.,
8., 3., 8., 7., 8., 7., 4., 4., 4., 4., 9., 7., 9., 7., 4., 4., 4., 4.,
9., 7., 9., 7., 7., 3., 7., 3., 6., 8., 6., 8., 7., 3., 7., 3., 6., 8.,
6., 8., 8., 3., 8., 3., 8., 7., 8., 7., 8., 3., 8., 3., 8., 7., 8., 7.},
{9, 4, 4});

auto tiled_3 = tile(data_2, {3, 2, 2});

auto expected_4 = array(
{4., 4., 4., 4., 9., 7., 9., 7., 4., 4., 4., 4., 9., 7., 9., 7., 7., 3.,
7., 3., 6., 8., 6., 8., 7., 3., 7., 3., 6., 8., 6., 8., 8., 3., 8., 3.,
8., 7., 8., 7., 8., 3., 8., 3., 8., 7., 8., 7., 4., 4., 4., 4., 9., 7.,
9., 7., 4., 4., 4., 4., 9., 7., 9., 7., 7., 3., 7., 3., 6., 8., 6., 8.,
7., 3., 7., 3., 6., 8., 6., 8., 8., 3., 8., 3., 8., 7., 8., 7., 8., 3.,
8., 3., 8., 7., 8., 7., 4., 4., 4., 4., 9., 7., 9., 7., 4., 4., 4., 4.,
9., 7., 9., 7., 7., 3., 7., 3., 6., 8., 6., 8., 7., 3., 7., 3., 6., 8.,
6., 8., 8., 3., 8., 3., 8., 7., 8., 7., 8., 3., 8., 3., 8., 7., 8., 7.,
4., 4., 4., 4., 9., 7., 9., 7., 4., 4., 4., 4., 9., 7., 9., 7., 7., 3.,
7., 3., 6., 8., 6., 8., 7., 3., 7., 3., 6., 8., 6., 8., 8., 3., 8., 3.,
8., 7., 8., 7., 8., 3., 8., 3., 8., 7., 8., 7., 4., 4., 4., 4., 9., 7.,
9., 7., 4., 4., 4., 4., 9., 7., 9., 7., 7., 3., 7., 3., 6., 8., 6., 8.,
7., 3., 7., 3., 6., 8., 6., 8., 8., 3., 8., 3., 8., 7., 8., 7., 8., 3.,
8., 3., 8., 7., 8., 7., 4., 4., 4., 4., 9., 7., 9., 7., 4., 4., 4., 4.,
9., 7., 9., 7., 7., 3., 7., 3., 6., 8., 6., 8., 7., 3., 7., 3., 6., 8.,
6., 8., 8., 3., 8., 3., 8., 7., 8., 7., 8., 3., 8., 3., 8., 7., 8., 7.},
{3, 6, 4, 4});

auto tiled_4 = tile(data_2, {3, 2, 2, 2});

auto expected_5 =
array({4., 4., 9., 7., 7., 3., 6., 8., 8., 3., 8., 7.}, {3, 2, 2});
auto tiled_5 = tile(data_2, {1, 1, 1});

auto expected_6 =
array({4., 4., 9., 7., 7., 3., 6., 8., 8., 3., 8., 7.}, {1, 3, 2, 2});
auto tiled_6 = tile(data_2, {1, 1, 1, 1});

CHECK(array_equal(expected_1, tiled_1).item<bool>());
CHECK(array_equal(expected_2, tiled_2).item<bool>());
CHECK(array_equal({expected_3}, {tiled_3}).item<bool>());
CHECK(array_equal({expected_4}, {tiled_4}).item<bool>());
CHECK(array_equal({expected_5}, {tiled_5}).item<bool>());
CHECK(array_equal({expected_6}, {tiled_6}).item<bool>());

CHECK_THROWS_AS(tile(data_2, {-1, 2}), std::invalid_argument);
}
TEST_CASE("tensordot") {
auto x = reshape(arange(60.), {3, 4, 5});
auto y = reshape(arange(24.), {4, 3, 2});
Expand Down Expand Up @@ -2314,4 +2388,4 @@ TEST_CASE("tensordot") {
65370.},
{3, 6});
CHECK(array_equal(z, expected).item<bool>());
}
}