Skip to content

🐍 Geometric Computer Vision Library for Spatial AI

fix(feature): ALIKED compiles with torch.compile when the image size changes (compiled 2.1x on CUDA) (#6495)

Dmytro Mishkinbc98050aa4bc
diff --git a/changelog.d/6495.fixed.md b/changelog.d/6495.fixed.md
new file mode 100644
index 000000000..64b1242ab
--- /dev/null
+++ b/changelog.d/6495.fixed.md
@@ -0,0 +1 @@
+`ALIKED` compiles with `torch.compile` when the image size changes. The second size made Dynamo trace the sizes symbolically, and the padding of `extract_dense_map`, a modulo of the image size, reached the offset bound of `DeformableConv2d`, which inductor could not lower: the call raised `BackendCompilerFailed` (`TensorReferenceAnalysis has no attribute 'python_mod'`). The padding is now computed without a modulo and has the same value, so eager results are unchanged.
diff --git a/kornia/feature/aliked/aliked.py b/kornia/feature/aliked/aliked.py
index fc6520867..159f3e6c1 100644
--- a/kornia/feature/aliked/aliked.py
+++ b/kornia/feature/aliked/aliked.py
@@ -422,8 +422,10 @@ class InputPadder:
def __init__(self, h: int, w: int, divis_by: int = 8) -> None:
self.ht = h
self.wd = w
- pad_ht = (((h // divis_by) + 1) * divis_by - h) % divis_by
- pad_wd = (((w // divis_by) + 1) * divis_by - w) % divis_by
+ # The distance to the next multiple, written without `%`: under dynamic-shape torch.compile a modulo of the
+ # image size reached the float offset bound of DeformableConv2d, which inductor (torch 2.14) cannot lower.
+ pad_ht = (h + divis_by - 1) // divis_by * divis_by - h
+ pad_wd = (w + divis_by - 1) // divis_by * divis_by - w
self._pad = [pad_wd // 2, pad_wd - pad_wd // 2, pad_ht // 2, pad_ht - pad_ht // 2]
def pad(self, x: torch.Tensor) -> torch.Tensor:
diff --git a/tests/feature/test_aliked.py b/tests/feature/test_aliked.py
index d3463f6e9..e9492c054 100644
--- a/tests/feature/test_aliked.py
+++ b/tests/feature/test_aliked.py
@@ -24,7 +24,7 @@ import torch.nn.functional as F
import kornia
from kornia.feature import laf_is_filled
from kornia.feature.aliked import ALIKED, ALIKEDFeatures
-from kornia.feature.aliked.aliked import SDDH, _affine_from_cov, get_patches
+from kornia.feature.aliked.aliked import SDDH, InputPadder, _affine_from_cov, get_patches
from kornia.feature.aliked.deform_conv2d import _bilinear_sample, _grid_sample_bilinear, deform_conv2d
from testing.base import (
@@ -438,6 +438,48 @@ class TestALIKED(BaseTester):
output = aliked(torch.rand(0, 3, 64, 64, device=device, dtype=dtype))
assert output == []
+ def test_input_padder_pads_to_the_next_multiple(self):
+ """The pad reaches the next multiple of ``divis_by``, split evenly with the extra pixel at the end."""
+ for size in range(70):
+ for divis_by in (1, 8, 32):
+ left, right, top, bottom = InputPadder(size, size + 3, divis_by)._pad
+ for pad, extent in ((top + bottom, size), (left + right, size + 3)):
+ assert (extent + pad) % divis_by == 0
+ assert 0 <= pad < divis_by
+ assert bottom - top in (0, 1)
+ assert right - left in (0, 1)
+
+ def test_dynamo_dense_map_with_changing_sizes(self, dtype, device, torch_optimizer):
+ """extract_dense_map compiles in one graph and matches eager, also once the image size changes.
+
+ A second size makes Dynamo trace the sizes symbolically. The padding used to be a modulo of the image size,
+ which reached the float offset bound of DeformableConv2d, and inductor (torch 2.14) failed to lower it:
+ ``TensorReferenceAnalysis has no attribute 'python_mod'``.
+ """
+ if device.type == "mps" and dtype == torch.float64:
+ pytest.skip("MPS does not support float64")
+ torch.manual_seed(0)
+ aliked = ALIKED(model_name="aliked-t16").to(device, dtype).eval()
+ op = torch_optimizer(aliked.extract_dense_map, fullgraph=True)
+ for hw in ((64, 96), (96, 72), (80, 112)):
+ inp = torch.rand(1, 3, *hw, device=device, dtype=dtype)
+ with torch.no_grad():
+ expected, actual = aliked.extract_dense_map(inp), op(inp)
+ for e, a in zip(expected, actual):
+ self.assert_close(a, e)
+
+ def test_dynamo_model_with_changing_sizes(self, dtype, device, torch_optimizer):
+ """The compiled model, with the graph breaks of its keypoint selection, runs on changing image sizes."""
+ if device.type == "mps" and dtype == torch.float64:
+ pytest.skip("MPS does not support float64")
+ aliked = ALIKED(model_name="aliked-t16").to(device, dtype).eval()
+ op = torch_optimizer(aliked)
+ for hw in ((64, 96), (96, 72)):
+ with torch.no_grad():
+ output = op(torch.rand(1, 3, *hw, device=device, dtype=dtype))
+ assert len(output) == 1
+ assert isinstance(output[0], ALIKEDFeatures)
+
def test_smoke(self, dtype, device):
aliked = ALIKED().to(device, dtype).eval()
inp = torch.rand(1, 3, 64, 64, device=device, dtype=dtype)