Skip to content

Fix conv input gradient shape with uneven or large padding #4662

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
zcbenz merged 1 commit into from
Oct 10, 2026
Merged

Fix conv input gradient shape with uneven or large padding #4662

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
14 changes: 11 additions & 3 deletions mlx/primitives.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1503,7 +1503,7 @@ std::vector<array> Convolution::vjp(

int64_t in_size = dilate_size(in.shape(1 + i), input_dilation_[i]);
int64_t out_size = dilate_size(cotan.shape(1 + i), kernel_strides_[i]);
padding_hi[i] = safe_cast(in_size - out_size + padding_hi_[i], "conv");
padding_hi[i] = safe_cast(in_size - out_size + padding_lo_[i], "conv");
}

// Check for negative padding
Expand All @@ -1515,13 +1515,21 @@ std::vector<array> Convolution::vjp(
has_neg_padding |= (pd < 0);
}

// Negative padding is cropped from the output below
std::vector<int> conv_padding_lo = padding_lo;
std::vector<int> conv_padding_hi = padding_hi;
for (int i = 0; i < conv_padding_lo.size(); ++i) {
conv_padding_lo[i] = std::max(conv_padding_lo[i], 0);
conv_padding_hi[i] = std::max(conv_padding_hi[i], 0);
}

auto wt_trans = group_transpose(wt, 0, 1, -1);
auto grad = conv_general(
/* const array& input = */ cotan,
/* const array& weight = */ wt_trans,
/* std::vector<int> stride = */ input_dilation_,
/* std::vector<int> padding_lo = */ padding_lo,
/* std::vector<int> padding_hi = */ padding_hi,
/* std::vector<int> padding_lo = */ conv_padding_lo,
/* std::vector<int> padding_hi = */ conv_padding_hi,
/* std::vector<int> kernel_dilation = */ kernel_dilation_,
/* std::vector<int> input_dilation = */ kernel_strides_,
/* int groups = */ groups_,
Expand Down
29 changes: 29 additions & 0 deletions python/tests/test_conv.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1185,6 +1185,35 @@ def loss_fn(kernel, inputs, strides, groups):
)
self.assertEqual(grads.shape, k_shape)

def test_padding_input_grad(self):
def conv_loss(x, w, stride, padding):
out = mx.conv_general(x, w, stride=stride, padding=padding)
return mx.sum(out * out)

def padded_loss(x, w, stride, padding):
lo, hi = padding
pad_width = [(0, 0)] + list(zip(lo, hi)) + [(0, 0)]
out = mx.conv_general(mx.pad(x, pad_width), w, stride=stride)
return mx.sum(out * out)

for in_shape, w_shape, stride, padding in [
((1, 4, 1), (1, 3, 1), (1,), ([0], [1])),
((1, 4, 1), (1, 3, 1), (2,), ([0], [1])),
((1, 4, 1), (1, 3, 1), (1,), ([1], [0])),
((2, 6, 6, 3), (4, 3, 3, 3), (2, 2), ([0, 1], [1, 2])),
((2, 5, 5, 5, 2), (3, 2, 2, 2, 2), (2, 1, 2), ([0, 0, 1], [1, 1, 0])),
((1, 4, 1), (1, 2, 1), (1,), ([3], [3])),
((1, 4, 1), (1, 2, 1), (1,), ([0], [3])),
((1, 4, 1), (1, 2, 1), (2,), ([3], [0])),
((2, 5, 6, 3), (4, 2, 3, 3), (1, 2), ([4, 0], [1, 5])),
]:
x = mx.random.normal(in_shape)
w = mx.random.normal(w_shape)
expected = mx.grad(padded_loss)(x, w, stride, padding)
grad = mx.grad(conv_loss)(x, w, stride, padding)
self.assertEqual(grad.shape, x.shape)
self.assertTrue(mx.allclose(grad, expected, atol=1e-4, rtol=1e-4))

@unittest.skipIf(mx.cuda.is_available() and "CI" in os.environ, "flaky in CI")
def test_conv_1d_with_2d(self):
x = mx.random.uniform(shape=(2, 10, 16))
Expand Down
Loading