Repository navigation
Update reference outputs and make output comparison stricter - #565
Merged
Merged
Conversation
- Regenerate `test/reference_outputs` with the current master. The old ones had drifted: `resid_gaus.fits` differed by 2.9% (close to the CI tolerance of 3%), and the wavelet model images by 12.7%. - `compare_results` now compares all subdirectories recursively (the images in `wavelet/model`, `wavelet/residual` and `wavelet/background`, and the top-level `ch0_*` images, were not compared), and fails if a file is missing in one of the runs (this was only logged). - `tbdsf_process_image.py` takes the tolerance as `--rtol`, with a default of 0.001 for runs on the same kind of system as the one that generated the reference outputs (x86-64 Linux). CI and cibuildwheel use 0.03. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Collaborator
Author
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
Codex Review: Didn't find any major issues. Can't wait for the next one! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
darafferty
approved these changes
Nov 9, 2026
Collaborator
Author
Yes, definitely. I will try to make sure we update those files only once. But that will require me to bypass the "pipelines must succeed" rule. |
gmloose
added a commit
that referenced
this pull request
Nov 9, 2026
* Fix `inigaus_fbdsm` PA (#535) `fit_island()` converts the beam PA to the convention used by `gaus_2d()` by adding 90 deg. That beam is passed to `get_maxima()` and then to `mclean()`, which adds another 90 deg. The clean model is therefore rotated by 90 deg. * Fix wavelet output files (#548) **Bugs** The wavelet module incorrectly generated the `model_wavelets.fits` and `resid_wavelets.fits`. After Gaussian fitting, the residual image is further processed by wavelet Gaussian fitting. However, the implementation did not correctly account for the standard Gaussian model when saving the final outputs. - `model_wavelets.fits` contained only the additional model contributed by wavelet Gaussian fitting, rather than the complete source model. - `resid_wavelets.fits` contained the original image with only the wavelet model subtracted. So the standard Gaussian model was still present in this image, meaning it was not the actual residual. - This also caused incorrect results when wavelet Gaussian fitting was disabled (atrous_bdsm_do = False): the model image was 0 and the residual image was identical to the input image. **Fixed** - `model_wavelets.fits` contain the complete model (both standard and wavelet Gaussian components). - `resid_wavelets.fits` contain the final residual after subtraction of both model components. **Added** - `model_wavelets_component.fits` to preserve the wavelet-only model. Also updated the wavelet image filenames to distinguish wavelet components from remaining scale images when `atrous_sum = True`. * Update reference outputs and make output comparison stricter (#565) * Update reference outputs and make output comparison stricter - Regenerate `test/reference_outputs` with the current master. The old ones had drifted: `resid_gaus.fits` differed by 2.9% (close to the CI tolerance of 3%), and the wavelet model images by 12.7%. - `compare_results` now compares all subdirectories recursively (the images in `wavelet/model`, `wavelet/residual` and `wavelet/background`, and the top-level `ch0_*` images, were not compared), and fails if a file is missing in one of the runs (this was only logged). - `tbdsf_process_image.py` takes the tolerance as `--rtol`, with a default of 0.001 for runs on the same kind of system as the one that generated the reference outputs (x86-64 Linux). CI and cibuildwheel use 0.03. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Restore reference outputs from master --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * Update reference outputs --------- Co-authored-by: Aleksander Kurek <aleksander.kurek@uj.edu.pl> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
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
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.
The reference outputs of
test/tbdsf_process_image.pyhad drifted from whatmasterproduces, and the comparison didn't check all outputs. This PR is meant to be merged before #564 (replace_cbdsm), so that #564 can be checked against up-to-date references.Changes
test/reference_outputswith the currentmaster(4fc79a2) on x86-64 Linux, with Python 3.12, NumPy 2.4.6, SciPy 1.18.1 and Astropy 8.0.2 (_cbdsmfrom a cibuildwheel build). Compared with the old references:residual/resid_gaus.fitsdiffered by +2.9%, just under the CI tolerance of 3%.wavelet/model/*model*.fitsby -12.7%.compare_resultsnow recurses into subdirectories. Before, it only compared the files one level down. The 15 images inwavelet/model,wavelet/residualandwavelet/background, and the top-levelch0_*images, were never compared. With the old references, the recursive comparison fails even at 3%, on the wavelet model images.check_all_files_presentwas ignored, so it was only logged.tbdsf_process_image.pytakes the tolerance as--rtol. The default is 0.001, for runs on the same kind of system as the one that generated the references. CI (ci.yml) and the cibuildwheeltest-commanduse--rtol 0.03, as before. A tighter CI tolerance would be fragile: some fits in this test are ill-conditioned, so last-bit rounding differences (another architecture such as macOS arm64, other library versions) can change the PSF-variation maps by tens of percent.Verification
All on the machine that generated the references:
mastermaster_cbdsmreplacement)The last row shows the test still detects real changes.
Not changed: the image comparison uses
np.isclosewith its default absolute tolerance (1e-8). For near-empty images, such as some wavelet background maps, that tolerance dominates, so those images are hardly checked.🤖 Generated with Claude Code