Repository navigation
configs: ship PALETTE arms against the current setup, and build-test the paper configs - #142
Open
Hendrik-code wants to merge 1 commit into
Open
Hendrik-code wants to merge 1 commit into
Hendrik-code wants to merge 1 commit into
Conversation
…the paper configs `RandomPaletteGPU` has been on main since the palette restack, but no shipped config runs it, so there was no way to train the published PALETTE arms from this repo. The only configs that did live on the `palette` branch, 274 commits back and written against the pre-rename `auglab` schema (`probability`, `mix_in_out`, `ImageContrastV26_6_2GPUTransform`). Two configs, both translations of arms from that branch: * `transform_params_paper-palette.json` -- PALETTE at p=1.0 in the frame `transform_params_paper-synthseg.json` defines, with SynthSeg removed and every other key untouched. The two therefore differ in one augmentation, which makes PALETTE and SynthSeg comparable as synthesis methods rather than as two differently-tuned pipelines. * `transform_params_paper-palette-ours.json` -- the paper `Ours` config plus PALETTE at p=0.5, which is the `train050` arm. Every PALETTE parameter the `train050` arm set -- c_choices, s_choices, blur_sigmas, dark_threshold, n_kmeans_subsample, skip_parcellation_prob, skip_sub_parc_prob, alpha_magnitude_range, label_remap_prob, min_label_voxels, label_classes -- is byte-for-byte the registry default here, so both configs carry `p` and nothing else rather than restating values that would then be free to drift. `gpu_config_paths` now also covers `transform_params_paper-*`. Those configs were checked for JSON validity and nothing else, so neither the new ones nor `transform_params_paper-synthseg.json` -- the config the published `SynthSeg (post)` arm trained on -- was ever built or run by the suite. All five build and push a volume through. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This branch has not been deployed
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.
Why
RandomPaletteGPUhas been onmainsince the palette restack, and the composablePaletteSynthesisGPUbeside it, but no shipped config runs either — so there is no way to train a PALETTE arm from this repo. The only configs that did live on thepalettebranch, 274 commits behindmainand written against the pre-renameauglabschema (probabilityrather thanp,mix_in_out,ImageContrastV26_6_2GPUTransform).What
Two configs, both translations of arms from that branch to the current setup:
transform_params_paper-palette.json4b92045cp: 1.0, transfer augmentations offtransform_params_paper-palette-ours.json3a4eb987Oursconfig plus PALETTE atp: 0.5transform_params_paper-palette.jsonistransform_params_paper-synthseg.jsonwithRandomSynthSegGPUremoved andRandomPaletteGPUin its place, every other key untouched. The two configs therefore differ in exactly one augmentation, which makes PALETTE and SynthSeg comparable as synthesis methods rather than as two differently-tuned pipelines. The-oursfile is thetrain050arm.Both carry
pand nothing else. Every PALETTE parameter the oldtrain050arm set —c_choices,s_choices,blur_sigmas,dark_threshold,n_kmeans_subsample,skip_parcellation_prob,skip_sub_parc_prob,alpha_magnitude_range,label_remap_prob,min_label_voxels,label_classes— is byte-for-byte the registry default here, so restating them would only create something that can drift from the defaults.Both needed a
.gitignoreentry:*.jsonis blanket-ignored and the comment there says a new tracked config has to say so.A gap this found
gpu_config_paths()matchedtransform_params_gpu*only, sotransform_params_paper-*was covered by the JSON-validity check and nothing else. That includestransform_params_paper-synthseg.json— the config the publishedSynthSeg (post)benchmark arm trained on — which no test has ever built or run.The prefix list now includes
transform_params_paper-, so all five GPU configs build anAugTransformsGPUand push a volume through it:Tests
469 passed, 13 skipped, 1527 subtests passed— the subtest count is up because the two new configs andpaper-synthsegnow run throughtest_configs.py's build, determinism and shape checks.🤖 Generated with Claude Code