Repository navigation
BUG: Keep the real-CT test on the grid it shrinks to - #4
Merged
Merged
Conversation
The test shrinks its lung CT pair by 3 to stay light on a CPU, and asked for features at a voxel size of 1 mm. That voxel size is honoured since the fine stage resamples the images itself, so the model saw the pair at 1 mm again, over a hundred times the voxels. The test went from 7 s to 40 s on Linux and Windows and from 4 s to 1400-1500 s on macOS, where it timed out on main two runs out of three. A voxel size of 0 takes the grid as it is, the workload the test had before.
vboussot
force-pushed
the
test-real-ct-grid
branch
from
October 3, 2026 18:16
c94cf1d to
2724451
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
ImpactConvexAdam.RealLungCTCoarseInitThenAdamRefinestimes out on macOS onmain(two runs out of the last three; the third passed at 1454 s of the 1500 s allowed).The test shrinks its lung CT pair by 3 to stay light on a CPU and asked for features at a voxel size of 1 mm. Since #2 the fine stage resamples the images to that voxel size itself, so the model saw the shrunk pair (4.1 x 4.1 x 7.5 mm) at 1 mm again, about 126 times the voxels. Measured on Linux with the 0.1.6 wheel, the model's convolutions cost 2.44 s at 1 mm against 0.016 s on the grid as it is, for the same number of calls. In CI the test went from 7 s to 40 s on Linux and Windows, and from 3.8 s to 1400-1500 s on macOS.
A voxel size of 0 takes the grid as it is, which is the workload the test had before #2. The library is unchanged.