Ais train optim DO NOT MERGE by constantinpape · Pull Request #1371 · computational-cell-analytics/micro-sam

constantinpape · 2026-09-08T12:39:30Z

This contains all the code for AIS optimization.

This is just for reference, we don't want to merge this.

Snapshot of the APG 2D/3D optimization experiments on this branch: library changes to the v2 automatic prompt generator and UniSAM2 decoder width, the campaign screening/training/benchmark scripts and their JSON configs, campaign notes and reviews, development check scripts, and the accompanying tests. Committed as-is so the full experimental state stays on this branch; the clean-up happens on a separate branch. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Every optimization tested in the 2026-09 APG campaigns was refuted under the generalization rule and is removed from the library, together with the learned prompt-selection machinery that PR #1340 had merged earlier: - micro_sam/v2/automatic_prompt_generation.py drops the learned multimask selector, filter and refinement gate (multimask_scorer, multimask_selection, score_filter, set_multimask_models, gate kwargs), the structural hooks (prompt_type, arbitration, fusion, recover_residual and their helpers), the label-free refinement gating rules (protect_neighbours, negative_scope, gate="isolated", isolated_fallback, touch_radius), the 3D learned candidate filter (return_metadata, VOLUME_CANDIDATE_FEATURE_NAMES, candidate_* options) and every experiment-only hook (prompts=, keep_trace, return_multimask_features). The plain second-round refinement (points / boxes / masks, 2D and the 3D anchor-slice variant) stays, as does the tiled generator. - micro_sam/v2/multimask_selection.py is deleted; the annotator no longer passes score_filter. - finetuning/v2/evaluation keeps the generic harness core (2D/3D manifests and runners, comparator, SLURM submitter, task registry, 3D case tools, pre-existing library benchmarks) stripped of the learned/structural plumbing; the 17 feature-specific screens/trainers/readers and 39 of 44 configs are removed. The 3D runner loses its trace-based recall attribution and anchor export. The stray test_apg_3d_tiling.py becomes check_apg_3d_tiling.py so bare pytest no longer collects it. - Tests for the removed features are deleted or trimmed; the two comparator test files are merged into test_apg_optimization.py. The full campaign state stays preserved on branch apg-optim-fable (356b76d). This commit starts a new implementation checksum epoch (f76ee7170ca77da882c0078dfaa5b301, seven files). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Every campaign note gets a banner stating that the mechanisms it describes were tested, refuted and removed on this branch, that the full state is preserved on apg-optim-fable (356b76d), and which sections describe kept functionality (the second-round refinement). A "Status on this branch" section at the end of each note maps the removed scripts, configs and library hooks. EXPERIMENTAL_SETUP.md records the reproducible set-up of the campaigns for future instance-segmentation (AIS) optimization: environment and SLURM presets, datasets and splits, checkpoints and the v4 staging recipe, the 2D subset manifests and the leak-free 3D crop manifests with their checksums, metrics and aggregation, parameter defaults and controls, acceptance and generalization gates, the timing-trial protocol, implementation checksum epochs (now f76ee7170ca77da882c0078dfaa5b301), the output-root layout and the steps to run an AIS campaign on the same set-up. The two tiled-3D-APG design notes move from the repository root into the notes directory unchanged. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…al_features torch_em 0.10.4 takes 'initial_features' as an argument of UNETR3D and sizes the decoder from it, so the joint/v4 checkpoints (32-wide decoders) load strictly through the plain kwargs forwarding UniSAM2 already had on dev. Verified against joint/v4 hvit_t geodesic best.pt: all 325 decoder keys match with no rebuild. micro_sam/v2/models/util.py returns to its dev version; the UniSAM2 width test stays. torch_em 0.10.1 built a 64-wide decoder regardless, so v4 checkpoints now require the newer torch_em (dependency floor to follow). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

submit_all_evaluations.py defaulted to the 'super' environment, which does not exist on grete, and to 3g.40gb slices for 3D jobs, which are scarce; it now activates 'new-stack' and puts 3D jobs on 1g.20gb (the (8, 512, 512) production crops peak around 6 GiB). The parameter_search.py array template activated 'super' as well and now uses the same environment; as a checksum file this starts implementation epoch e1903b1b3c1e4e3610c71e1d0bd81f1d (harness-only, results unaffected). EXPERIMENTAL_SETUP.md gains section 14 with the 2026-09-06 baseline reruns of the cleaned harness on joint/v4 geodesic defaults: the 2D subset benchmark (bit-identical to the recorded v4 controls), the 3D deep crops (holdout identical, primary identical on 56/57 crops) and the AIS/APG production defaults on nine test splits, each with its run directories, result files and reference numbers. The campaign notes point to it from their status sections. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Predict every manifest sample once, cache the (4, *spatial) decoder prediction with its labels, and run configurations, parameter sweeps, seed diagnostics and ground-truth oracles on the cache. Run directories follow the APG layout so the comparator reads them; the report applies the generalization gate. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…cessing The geodesic decoder predicts a distance magnitude that falls to zero along the boundary of every object it recognised, while a false foreground region carries no such dip. drop_instances_without_boundary_dip removes instances whose median boundary magnitude exceeds a threshold; flow_instance_segmentation exposes it as boundary_magnitude_max (off by default). The AIS benchmark mirrors it, the cached sweep scorer applies it, and the campaign notes record the diagnostics and prototypes that led to it. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…the A1 screens report_ais_sweep.py joins the per-dataset sweep tables of a grid over several manifests and scores every combination as a shared default: balanced mSA, change per dataset against the library defaults, the gate verdict and the ratio to each dataset's optimum. The notes record the Phase 1 diagnostics, the three prototypes and the epoch A1 filter screens. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

evaluate_automatic_segmentation.py --ais_params reads a benchmark-style configuration, resolves it against the library defaults (sparse and dense) and scores the test split with it; submit_all_evaluations.py passes --ais_params and --result_tag through. The sweep ranking accepts the deep 3d manifests. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…-aware The flow post-processing defaults of hvit_t become sigma 1.0, min_size 50 and the boundary-magnitude filter at 0.4 for images; volumes take min_size 200 and foreground 0.6 through a new 'sparse_volume' override table that default_postprocessing resolves by dimensionality. Against the registry values this gains +2.4 % balanced mSA on eleven 2d development datasets (9 up), +4.3 % on the 2d holdout and +22 % on the 3d LM crops with the joint/v4 geodesic checkpoint. The filter now derives every instance's boundary median from one sort over the boundary pixels, 18x faster with identical output. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…the filter test Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>