Port shared VisCy package changes from DynaCell v1.1 by alxndrkalinin · Pull Request #533 · Biohub/VisCy

alxndrkalinin · 2026-10-26T23:10:32Z

Summary

First of three stacked PRs that replace #494 (DynaCell v1.1, 2,626 files). This one ports the shared code: every change #494 makes outside applications/dynacell, taken file-for-file from dynacell-models (8e74a7e), including the package changes that Spotlight motivated. It is based on main.

PR-A (this PR): packages, root, CI, lock, cytoland/dynaclr fixes. 80 files.
PR-B: applications/dynacell without the Spotlight v2 experiments. Based on this branch.
PR-C: Spotlight v2 arms, eval buckets and tools. Based on PR-B.

dynacell-models itself is not touched; it stays the paper-cited snapshot.

What is ported, and from where

Each commit names its source commits/PRs on dynacell-models.

Area	Content	Source
viscy-utils losses	`SegAuxDice` + SAUNA/clDice maps, `SegConsistencyLoss`, `BackgroundLowPass`, `SpotlightLoss(return_components=)`	#498, #521, #524, #530
viscy-utils fg masks	CLAHE + 5³ median + Otsu masks; full-res Otsu; cubic GPU path (optional import); lazy torch	#498, #510
viscy-utils predict	`HCSPredictionWriter`: `z_reduction="center"`, `z_planes`, completion markers + ckpt/settings identity (`prediction_metadata.py`), refuse unsafe store writes, write-once async flush	`2b848a6`, #497, #511, #526, #531, #532
viscy-utils training	checkpoint guardrails + `MonitorHealthCheck` + `OptimizerHealthCheck` injected by `VisCyCLI` on every fit; `LinearLRCooldown`	`20e54db`, `daa7477`, `bc0f444`, #497, #528
viscy-utils misc	SSIM/MS-SSIM numerics under mixed precision; lazy `viscy_utils` exports	#484, `130e034`, `272051c`, `59343ad`, #509
viscy-data	`predict_z_planes` / `z_planes`, read-once predict (`cache_volume`), uint8 fg masks	#498, #526, #531, #532
viscy-models	per-axis CellDiff patches + `cond_channels`, Beta/cosine ODE time grids, MorphEm extractor, PatchGAN `preserve_z`, UNet3D `downsample_z` + 5-D check	`0126fe6`, #498, #520, #479, `c2d0481`, #484
viscy-transforms	no blanking of singleton axes in `BatchedRandAffined`; pointwise sampled normalizers; deterministic noise test	`9640537`, #532, #513
cytoland / dynaclr	Spotlight recipe comments; VCP tutorials install from GitHub (+ paired notebooks); dynaclr fixture fix (7 tests failing on main)	#498, #479, `d1ea492`
root	pytest-xdist + root `conftest.py` thread pin; strict markers registered at the root; `hf_demo` on the pytest pythonpath (used by PR-B's `build_demo_data_test.py`); CLAUDE.md trims	#512, #489, #523, `93ac234`

Edits relative to `dynacell-models`

git diff origin/dynacell-models HEAD -- <these 80 paths> touches only these five files:

.github/workflows/test.yml: no dynacell-models PR target; the dynacell job leaves out the eight test files that only arrive with PR-B; new test-utils job (ubuntu, Python 3.12 + 3.13). viscy-utils had no CI job on either branch, so most of its tests (losses, metrics, meta_utils, optimizer health, CLI) ran nowhere.
.github/workflows/lint.yml: unchanged from main (the dynacell-models target is dropped).
uv.lock: re-locked against this tree, whose applications/dynacell/pyproject.toml is still main's. uv lock --check exits 0. Against dynacell-models it drops ten dynacell-only packages and keeps cubic at 0.7.0; PR-B restores them.
pyproject.toml: also registers requires_data, which main's generate_grouped_eval_configs_test.py uses; with strict markers it would otherwise fail to collect. PR-B removes the registration with that test.
CLAUDE.md: the DynaCell job runbook (hung jobs, stall watchdog, wall limits) waits for PR-B, since it documents tools that only exist there.

Review-gate fixes on top of the port

Each fix is its own commit. These are the only lines that differ from dynacell-models in packages/, besides the five root files above.

Fix	Why	Source
`viscy-utils` declares `torchvision`	`evaluation.metrics` imports it at module level. The new `test-utils` job (viscy-utils alone) failed to collect 6 modules	CI run 1
`viscy_utils.callbacks` resolves its exports lazily	the health callbacks injected into every fit pulled in `OnlineEvalCallback` and with it scikit-learn (an extra). The new test blocks sklearn: it fails on the old `__init__` and passes now	Copilot
a fit with `trainer.callbacks: None` still gets `OptimizerHealthCheck`	it was only appended to an existing list	Copilot
the injected latest-weights checkpoint copies a step/time trigger	a step-cadence leaf would have had `last.ckpt` only at epoch ends. No current leaf combines a monitored checkpoint with a step trigger	/code-review
every monitored checkpoint gives up `save_last`, not only the first	a second monitored checkpoint kept writing `last.ckpt` next to the injected saver. No current leaf has two	/code-review
`equalize_clahe` returns zeros when the 0.1/99.9 percentiles are equal	a blank volume with a few hot pixels divided by zero (reproduced). The mask stays empty, as for a constant volume	/code-review
VCP tutorials also install `viscy-data` from GitHub	the writer reads `SlidingWindowDataset.z_planes`, which PyPI `viscy-data` 0.1.0 lacks (checked in its wheel)	/code-review
CLI injector lookup shared; `MonitorHealthCheck.check_finite` (unset anywhere) removed; AMP skip-counter message and fg-mask docstrings corrected	cleanup	/simplify, Copilot
pandas string inference is no longer switched off process-wide: `viscy_utils.evaluation.zarr_utils.write_anndata_zarr` converts obs/var Arrow-backed strings to object just before writing (reusing `append_to_anndata_zarr`'s conversion), and the embedding writer, `EmbeddingSnapshot`'s rewrite, `annotation.convert`, dynaclr's `split_embeddings` and 10 test writes use it. The viscy-utils and dynaclr conftest switches and the two runtime flips are gone; the cellanome scripts keep theirs	anndata 0.12.6 cannot write pandas 3's default strings to zarr, but the switches made every test run with that default off (dynaclr's `pytest_configure` did it for all packages in a repo-wide run), and two writers worked only after the embedding writer had flipped the option. New `test_zarr_utils.py` pins a round trip under the default, that plain `write_zarr` still fails, and that the writer leaves the option alone	review (`48347a7`, `1b2e227`)
dynacell CI job: main's `test_submit_benchmark_batch` (needs `/hpc`) and the four viscy-utils test files (now in `test-utils`) left out; `hf_demo` pytest path left to PR-B	CI run 1, /simplify

Findings judged out of scope for a port and left as follow-ups:

MonitorHealthCheck vs an LR cooldown to 0, or a monitor logged only every N epochs;
a leaf's own unmonitored ModelCheckpoint next to the injected one (Lightning rejects it at startup): no config has that shape;
OptimizerHealthCheck's per-step host syncs;
the full-resolution Otsu pass's memory/time, and a device-resident GPU chain;
three sorts per step in SegAuxDice;
MorphEmModel's revision=None default: dynacell's config pins it;
uint8 masks through interpolating CPU augmentations;
dropping dynaclr's testpaths;
compute_fg_masks still requiring compute_otsu.

Behavior changes that reach other apps

Every VisCyCLI fit (cytoland, dynaclr, dynacell) now gets an unmonitored latest-epoch ModelCheckpoint that owns last.ckpt (the monitored one gets save_last=False), MonitorHealthCheck(patience=10) and OptimizerHealthCheck. last.ckpt now always means "latest weights".
HCSPredictionWriter raises NotImplementedError when trainer.world_size > 1. Multi-GPU predict into one store (for example a cytoland demo left at default devices on a multi-GPU host) now refuses instead of racing. It also refuses writes into an existing store when a FOV outside the run has unmarked prediction channels, or when an array would grow.
Predict at num_workers=0 reads each (position, t) volume once, and normalizes it once when every transform is pointwise. The writer says the bytes are unchanged.
fg masks reach the batch as uint8 and are cast to float on the device. generate_fg_masks uses CLAHE + 5³ median + Otsu, and compute_otsu uses a full-resolution 5³ median + σ=5 Gaussian. otsu_grid_spacing is gone. Spotlight masks and thresholds that viscy preprocess writes change.
SSIM / MS-SSIM loss values change slightly: variance clamp, Cauchy–Schwarz bound, symmetric eps, MS-SSIM outside autocast.
BatchedRandAffined no longer blanks a singleton spatial axis. UNet3DBase.forward raises on non-5-D input. viscy_utils resolves its exports lazily.

Conflicts with open PRs to main

dynacell-models contains none of these PRs' changes. Whichever merges second rebases.

fix(viscy-utils): fail fast when ckpt_path does not exist #505 (cli.py, test_cli.py) and lightning config for requeing jobs on a preempted slurm partition #478 (cli.py, uv.lock): this PR adds the guardrail injection to cli.py and raises the test_cli timeout.
DynaCLR misc: CSV metadata sidecar, data-path profiling, DDP logging #485 (meta_utils.py, trainer.py, test_meta_utils.py, CLAUDE.md, uv.lock): this PR rewrites the Otsu/fg-mask code in meta_utils.py and drops otsu_grid_spacing from trainer.py.
uv.lock: chore(deps): bump iohub to >=0.3.10 #482, DynaCLR misc: CSV metadata sidecar, data-path profiling, DDP logging #485, feat(dynaclr): read per-FOV tracks from in-plate tracks.geff #525 also re-lock. chore(deps): bump iohub to >=0.3.10 #482 and feat(dynaclr): read per-FOV tracks from in-plate tracks.geff #525 also edit applications/dynaclr/pyproject.toml, where this PR drops the dynaclr testpaths.

Tests

Local runs: pytest -n 8 on an A40 node, in this branch's own venv (uv sync --all-packages --all-extras --frozen from this lock: cubic 0.7.0, cellpose 4.2.1.1, torch 2.12.0). The root addopts deselect slow.

Suite	This branch	`main` (same venv)
`packages/` (tests + colocated `*_test.py`)	1233 passed, 1 xfailed, 0 failed (at `0b1cb979`)	916 passed, 3 failed (`test_cli` subprocess timeouts, fixed here), 1 xfailed
`packages/viscy-utils` in a viscy-utils-only venv (mirror of the `test-utils` job)	387 passed, 5 skipped, 0 failed (at `5467d3a9`)	n/a
cytoland + dynaclr + qc + airtable	398 passed, 4 skipped, 0 failed (at `0b1cb979`)	391 passed, 7 failed (dynaclr fixture, fixed here), 4 skipped (earlier measurement, dynacell-models venv)
main's `applications/dynacell`	723 passed, 13 failed, 79 skipped	722 passed, 14 failed, 79 skipped

Full repo, pytest -n 11 at 1b2e227e (11 cores, this branch's venv): 2163 passed, 12 failed, 83 skipped, 1 xfailed. All 12 are main's applications/dynacell (test_evaluation_metrics.py metric stubs resolving to None, and test_evaluation_pipeline.py::test_evaluate_model_reuses_cache_without_feature_metrics); the same 12 fail with the two pandas commits reverted.

The 13 dynacell failures are the same 13 on pristine main in the same venv (comm of the two FAILED sets: 0 only on this branch). They are main's own dynacell code against the re-locked environment (ImportError: MADScaler, metric functions resolved to None, a live-data walk test that finds 0 stores); PR-B replaces that code. The one extra failure on main, test_focus.py::test_per_cell_similarity_z_slab_restricts_and_changes_outcome, is an xdist ordering flake that passes in isolation.

Lint: uvx prek run on the changed files passes. CI: green on 0b1cb979 (Test, Lint, Documentation).

Review guide

Package code is about 10k added lines. The largest pieces are prediction_writer.py + prediction_metadata.py (~950 lines with ~1,150 lines of tests), the four loss modules (~1,160 lines with ~1,040 lines of tests) and meta_utils.py.

🤖 Generated with Claude Code

Port the Spotlight-motivated losses from dynacell-models so they live in the shared package before any application consumes them: SegAuxDice (seg_aux.py) with the SAUNA-map and soft-clDice weight maps (seg_aux_maps.py), the segmenter-consistency loss (seg_consistency.py), the Track H background low-pass target (background_lowpass.py), and SpotlightLoss(return_components=True). Defaults keep SpotlightLoss's existing behavior; no package imports an application. Source on dynacell-models: 40de793 (#498), 6c28a31 (#521), 0cbeda1 (#524), 050f553 (#530). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port the Spotlight fg-mask recipe from dynacell-models. generate_fg_masks now binarizes a CLAHE-equalized, 5^3-median-filtered volume at an Otsu threshold picked on that volume after a sigma=5 Gaussian; compute_otsu uses the same full-resolution smoothing, so the otsu_grid_spacing parameter is gone from generate_normalization_metadata and VisCyTrainer.preprocess. The median and Gaussian run through cubic on the GPU when cubic and a device are present, else SciPy; cubic stays an optional import. torch is imported lazily so that light callers of write_meta_field do not pay for it. Behavior change: Spotlight masks and Otsu thresholds written by `viscy preprocess` differ from main's. Source on dynacell-models: 40de793 (#498), d087d96 (#510). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…rite-once Port the HCSPredictionWriter and prediction_metadata work from dynacell-models: - z_reduction="center" and z_planes (slab-only predicts write only the planes they predicted, and record them); - per-FOV completion markers and the checkpoint/settings identity that predict resume and the eval caches key on (prediction_metadata.py); - refuse writes into an existing store when a FOV outside the run has unmarked prediction channels, or when the array would grow; - assemble each volume in memory and write it once on a background thread (bytes unchanged); - raise NotImplementedError when trainer.world_size > 1, instead of letting ranks race on one store. Source on dynacell-models: 2b848a6, 1be4466, 74fa18a (#497), 8e41498 (#511), 996adb4 (#526), 6c83fff (#531), 8e74a7e (#532). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port the training guardrails from dynacell-models. VisCyCLI now injects, on every `fit`: - a latest-weights ModelCheckpoint that owns last.ckpt (the monitored checkpoint gets save_last=False), so last.ckpt always means the latest epoch, inheriting the leaf's write cadence; - MonitorHealthCheck, which raises when the monitored metric is non-finite or bit-identical for 10 validation epochs; - OptimizerHealthCheck, which raises after 5 consecutive non-finite losses or on GradScaler collapse. Also adds the opt-in LinearLRCooldown callback and raises the CLI subprocess test timeout to 120 s (it timed out on main). Behavior change: this applies to cytoland, dynaclr and dynacell fits. Source on dynacell-models: 20e54db, daa7477, bc0f444, 74fa18a (#497), 92693ae, 2f9ef31 (#528). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port two viscy-utils changes from dynacell-models: - ssim_25d / ms_ssim_25d: clamp the variances, bound the covariance by Cauchy-Schwarz through a safe sqrt, floor the stability constants symmetrically, and run MS-SSIM outside autocast, so 16-mixed training cannot NaN the gradient. SSIM/MS-SSIM loss values change slightly for MixedLoss users. - viscy_utils/__init__ resolves its exports lazily (PEP 562), keeping torch and lightning off light import paths. Source on dynacell-models: a556bc9 (#484), 130e034, cee4e41, 272051c, 59343ad, 8b2413f (#509). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…asks Port the viscy-data side of the dynacell-models predict and Spotlight work: - HCSDataModule.predict_z_planes and SlidingWindowDataset z_planes / read_z_planes, so a 2D model predicts only its focus slab; - cache_volume: with num_workers == 0 predict reads each (position, t) volume once and, when every transform is pointwise, normalizes it once; - foreground masks are read in their stored dtype (uint8) and cast to float on the device in on_after_batch_transfer. Behavior change: every app's predict at num_workers=0 takes the read-once path. Source on dynacell-models: 40de793 (#498), 996adb4 (#526), 6c83fff (#531), 8e74a7e (#532). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…hannels Port from dynacell-models: - patch_size may be a per-axis (z, y, x) tuple through the patch embedding, transformer unpatchify and ViT bottleneck, which lets CELLDiffNet run Z-preserving at Z=1 for the 2D track (an int is still accepted); - CELLDiffNet(cond_channels=1) adds extra conditioning input channels (default keeps the old single-channel net). Source on dynacell-models: 0126fe6, 0617ed7, 40de793 (#498). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port the time_schedule option of the CellDiff ODE integrators from dynacell-models: uniform (default, unchanged), cosine, or a Beta(p, q) grid packed toward the data end, which is what lets the DynaCell-lite recipe sample in 14 velocity evaluations. scipy joins the celldiff extra for betaincinv. Source on dynacell-models: 19765e7 (#520). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port MorphEmModel from dynacell-models, the fifth deep-feature extractor of the DynaCell evaluation. transformers is imported lazily and the trust_remote_code model is pinned to a Hub commit. Source on dynacell-models: d3f6422 (#479), 4918d2e, 13cb336, c9c6914. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port preserve_z from dynacell-models, so the 3D PatchGAN runs on Z=1 inputs for pix2pix2d (default off, unchanged). Source on dynacell-models: c2d0481. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port from dynacell-models: - UNet3D(downsample_z=True): False keeps Z through the encoder, which is how FNet2D runs on single planes (default unchanged); - UNet3DBase.forward raises on non-5D input, and a comment flags the ConvTranspose checkerboard artifact. Source on dynacell-models: a556bc9 (#484), 83a1d69, 16b1d31. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port 9640537 from dynacell-models: a size-1 spatial axis (Z=1 for 2D models) gets an identity row in the affine, instead of being resampled off the volume and zeroed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

NormalizeSampled and MinMaxSampled set pointwise = True, which lets the read-once predict path in viscy-data normalize a whole volume once instead of once per window. Source on dynacell-models: 8e74a7e (#532). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ministic Port 3409560 (#513) from dynacell-models. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port dfcf9a5 (#512) from dynacell-models: pytest-xdist joins the test group, and a root conftest.py pins OMP/MKL/OpenBLAS to one thread under each xdist worker. Without the pin every worker's torch sizes its pool to all cores, which made `-n 11` 10x slower than it should be and timed out subprocess tests. CLAUDE.md documents the usage. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port the marker half of c50300f (#489) from dynacell-models. strict_markers = true turns an unregistered marker into a collection error, so a typo'd marker can no longer silently decorate nothing. hpc_integration moves from the cytoland and dynaclr conftests to the root: an addinivalue_line registration only holds when that conftest is in the invocation's ancestry. dynaclr drops its own testpaths so the root ini applies. requires_data is registered too: main's applications/dynacell/tools/generate_grouped_eval_configs_test.py uses it, and strict markers would otherwise drop that module at collection. (dynacell-models no longer uses the marker; the dynacell PR that follows removes the registration with it.) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port the root half of 7b7cb43 (#523) from dynacell-models so the HF demo's tests import its app module. The directory already exists on main. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port d1ea492 from dynacell-models. The orchestrated linear-classifier fixture joined on the first fov_name component, which failed 7 tests on main once pytest collects dynaclr's colocated *_test.py. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port the cytoland recipe comment fixes from 40de793 (#498): the preprocess flags take explicit values (`--compute_otsu true --compute_fg_masks true`), fg_threshold and nonzero_threshold are inert while fg_mask_key is set, and the Otsu-centred target is what makes 0 the FG/BG boundary. Comments only. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port the tutorial fix from d3f6422 (#479): cytoland and viscy-utils are not on PyPI yet, so the tutorials install them from the GitHub monorepo; their siblings resolve from PyPI. Adds the jupytext-paired notebooks (no outputs). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port 93ac234 from dynacell-models: drop the repo-layout tree, setup and common-command blocks, lint/docstring rules (pyproject.toml is the source) and the generic coding-philosophy sections. The DynaCell job runbook from the same file arrives with the dynacell PR, since it documents tools that only exist there. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

viscy-utils had no CI job on main or on dynacell-models: four of its test files ran inside the dynacell job and the rest (losses, metrics, meta_utils, optimizer health, CLI, lazy imports) ran nowhere. Mirror the viscy-data job on ubuntu-latest for Python 3.12 and 3.13, and make the gate job wait for it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Port the test-dynacell-configs job from dynacell-models, minus the eight test files that only arrive with the dynacell PR: sync viscy-utils' test group with dynacell, and add test_submit_benchmark_batch, test_engine, test_pipeline_cache, feature_metrics_test and the viscy-utils prediction-writer, prediction-metadata, monitor-health and LR-cooldown tests. The dynacell-models PR target is not carried over. Source on dynacell-models: 33d2fee, 74fa18a (#497), 2d1ed32 (#506), 4450efd (#518), 2798dee (#519), 2f9ef31 (#528). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Start from dynacell-models' uv.lock and re-resolve with `uv lock` against this tree, whose applications/dynacell/pyproject.toml is still main's. Against dynacell-models' lock this drops the ten dynacell-only packages (dinov3, etils, ftfy, jsonpath-rw, mlcroissant, pandas-stubs, ply, rdflib, submitit, termcolor) and resolves cubic back to 0.7.0; the dynacell PR restores them. Against main's lock it adds pytest-xdist and moves boto3/botocore 1.43.0 -> 1.43.20, cellpose 4.1.1 -> 4.2.1.1, pyarrow 19.0.1 -> 24.0.0, s3transfer and tracksdata; torch and lightning are unchanged. `uv lock --check` exits 0 (main's own lock does not). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Copilot

🟡 Changes recommended

Unresolved training-guard and segmenter-dtype defects can leave fits unprotected or interrupt otherwise progressing training.

3 open issues

What changed in this PR

Ports shared DynaCell v1.1 infrastructure into VisCy’s packages, providing the foundation for the two subsequent application-focused PRs.

Changes:

Adds training guardrails, segmentation-consistency loss, numerical stability improvements, and lighter imports.
Extends prediction data handling, singleton-axis transforms, model geometry, and feature extraction.
Expands regression coverage and CI, centralizes pytest configuration, and updates application examples.

File	Description
`pyproject.toml`	Adds parallel testing support and centralized marker configuration.
`packages/viscy-utils/tests/test_spotlight_loss.py`	Tests returned loss components.
`packages/viscy-utils/tests/test_smooth_median.py`	Tests median-filter backend agreement.
`packages/viscy-utils/tests/test_smooth_gaussian.py`	Tests Gaussian smoothing and threshold behavior.
`packages/viscy-utils/tests/test_seg_consistency_loss.py`	Tests consistency loss, projection, and gradients.
`packages/viscy-utils/tests/test_prediction_metadata.py`	Tests checkpoint hashing and marker identity.
`packages/viscy-utils/tests/test_meta_utils.py`	Updates foreground-mask and Otsu coverage.
`packages/viscy-utils/tests/test_lazy_imports.py`	Checks lightweight imports and public exports.
`packages/viscy-utils/tests/test_cli.py`	Increases subprocess timeout for cold imports.
`packages/viscy-utils/src/viscy_utils/trainer.py`	Removes Otsu sampling-spacing configuration.
`packages/viscy-utils/src/viscy_utils/losses/spotlight.py`	Optionally returns individual loss terms.
`packages/viscy-utils/src/viscy_utils/losses/seg_consistency.py`	Adds frozen-segmenter consistency loss.
`packages/viscy-utils/src/viscy_utils/losses/__init__.py`	Exports additional loss modules.
`packages/viscy-utils/src/viscy_utils/evaluation/metrics.py`	Stabilizes SSIM values and gradients.
`packages/viscy-utils/src/viscy_utils/cli.py`	Injects checkpoint and optimizer guardrails.
`packages/viscy-utils/src/viscy_utils/callbacks/optimizer_health.py`	Adds optimizer-health monitoring.
`packages/viscy-utils/src/viscy_utils/callbacks/optimizer_health_test.py`	Tests optimizer failures and guard injection.
`packages/viscy-utils/src/viscy_utils/callbacks/monitor_health.py`	Detects non-finite or frozen checkpoint metrics.
`packages/viscy-utils/src/viscy_utils/callbacks/linear_lr_cooldown.py`	Adds step-based learning-rate cooldown.
`packages/viscy-utils/src/viscy_utils/callbacks/linear_lr_cooldown_test.py`	Tests cooldown scheduling and resume behavior.
`packages/viscy-utils/src/viscy_utils/callbacks/__init__.py`	Exports new training callbacks.
`packages/viscy-utils/src/viscy_utils/__init__.py`	Resolves public exports lazily.
`packages/viscy-transforms/tests/test_noise.py`	Seeds noise tests and corrects expected variance.
`packages/viscy-transforms/tests/test_affine.py`	Covers singleton-axis affine transforms.
`packages/viscy-transforms/src/viscy_transforms/_normalize.py`	Marks sampled normalizers as pointwise.
`packages/viscy-transforms/src/viscy_transforms/_affine.py`	Prevents singleton-axis data blanking.
`packages/viscy-models/tests/test_unet/test_unet3d.py`	Tests optional Z downsampling.
`packages/viscy-models/tests/test_unet/test_unet3d_base.py`	Tests rejection of non-5-D inputs.
`packages/viscy-models/tests/test_gan/test_patchgan3d.py`	Tests Z-preserving discriminator behavior.
`packages/viscy-models/tests/test_foundation/test_morphem.py`	Tests MorphEm loading and feature contracts.
`packages/viscy-models/tests/test_celldiff/test_unet_vit_3d.py`	Updates patch-validation error expectations.
`packages/viscy-models/tests/test_celldiff/test_transport.py`	Tests cosine and Beta sampling grids.
`packages/viscy-models/tests/test_celldiff/test_celldiff_net.py`	Updates patch-validation error expectations.
`packages/viscy-models/src/viscy_models/unet/unet3d.py`	Makes Z downsampling configurable.
`packages/viscy-models/src/viscy_models/unet/unet3d_base.py`	Validates input rank and documents existing artifacts.
`packages/viscy-models/src/viscy_models/gan/patchgan3d.py`	Adds Z-preserving convolution geometry.
`packages/viscy-models/src/viscy_models/foundation/morphem.py`	Adds the MorphEm feature-extraction wrapper.
`packages/viscy-models/src/viscy_models/foundation/__init__.py`	Exports MorphEm.
`packages/viscy-models/src/viscy_models/celldiff/vit_bottleneck.py`	Supports per-axis patch dimensions.
`packages/viscy-models/src/viscy_models/celldiff/unet_vit_3d.py`	Accepts per-axis patch configuration.
`packages/viscy-models/src/viscy_models/celldiff/modules/transport/transport.py`	Exposes ODE time schedules.
`packages/viscy-models/src/viscy_models/celldiff/modules/transport/integrators.py`	Constructs and validates nonuniform time grids.
`packages/viscy-models/src/viscy_models/celldiff/modules/transformer.py`	Handles non-cubic patch reconstruction.
`packages/viscy-models/src/viscy_models/celldiff/modules/patch_embed_3d.py`	Normalizes and embeds per-axis patches.
`packages/viscy-models/src/viscy_models/celldiff/celldiff_net.py`	Adds configurable conditioning channels and patches.
`packages/viscy-models/src/viscy_models/celldiff/celldiff_2d_test.py`	Tests 2-D geometry and conditioning channels.
`packages/viscy-models/src/viscy_models/__init__.py`	Adds the package-level MorphEm export.
`packages/viscy-models/pyproject.toml`	Adds SciPy to CellDiff dependencies.
`packages/viscy-data/tests/test_hcs.py`	Tests mask dtypes and cached normalization.
`packages/viscy-data/src/viscy_data/sliding_window.py`	Adds plane selection and volume caching.
`packages/viscy-data/src/viscy_data/hcs.py`	Integrates prediction options and device-side mask casting.
`packages/viscy-data/src/viscy_data/foreground_masks.py`	Preserves stored mask dtype during reads.
`conftest.py`	Limits numerical-library threads in parallel workers.
`CLAUDE.md`	Trims guidance and documents parallel testing.
`applications/dynaclr/tests/conftest.py`	Removes locally registered test marker.
`applications/dynaclr/src/dynaclr/evaluation/linear_classifiers/orchestrated_test.py`	Corrects annotation fixture join keys.
`applications/dynaclr/pyproject.toml`	Removes local pytest test-path configuration.
`applications/cytoland/tests/conftest.py`	Removes locally registered test marker.
`applications/cytoland/examples/vcp_tutorials/quick_start.py`	Updates installation instructions to use GitHub.
`applications/cytoland/examples/vcp_tutorials/neuromast.py`	Updates installation instructions to use GitHub.
`applications/cytoland/examples/vcp_tutorials/hek293t.py`	Updates installation instructions to use GitHub.
`applications/cytoland/examples/configs/vscyto3d/train_spotlight.yml`	Corrects preprocessing command comments.
`applications/cytoland/examples/configs/recipes/modes/spotlight.yml`	Clarifies normalization and mask-related comments.
`.github/workflows/test.yml`	Adds utility-package CI and regression coverage.

_{Review effort: Balanced}

viscy_utils.evaluation.metrics imports torchvision.ops at module level, and losses.MixedLoss imports that module, but viscy-utils never listed torchvision. Installs only worked because another workspace member pulled it in. The new test-utils CI job syncs viscy-utils alone and failed to collect six test modules with ModuleNotFoundError. The lock gains only the dependency edge; no version moves. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Two corrections to the test-dynacell-configs job: - main's tests/test_submit_benchmark_batch.py resolves leaves under /hpc and fails on a runner (5 PermissionErrors in CI). The dynacell-models version that runs there arrives with the DynaCell PR, which adds it back. - the four viscy-utils test files (prediction writer, prediction metadata, monitor health, LR cooldown) now run in the test-utils job on both Python versions, so the dynacell job no longer repeats them or syncs viscy-utils' test group for them. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ectors _inject_checkpoint_guardrails and _inject_optimizer_health_guard each repeated the same walk from the config to fit's trainer.callbacks plus the already-injected check; the `else config` branch of that walk could never run behind the `subcommand != "fit"` return. Both now call _fit_callbacks_without. _callback_field collapses its two identical branches, and the monitored checkpoint's init_args are kept from the search loop instead of being looked up twice. No behavior change. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Brings in #533's fix that keeps scikit-learn off the import path of the health callbacks every fit gets. No conflicts. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Brings in, via the DynaCell branch, #533's fix that keeps scikit-learn off the import path of the health callbacks every fit gets. No conflicts. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Copilot

🔵 Needs a closer look

Shared training, prediction, and numerical changes require maintainer integration review, including GPU-dependent behavior not independently validated here.

No open issues 🎉

2 resolved since last review

_{Review effort: Balanced}

The tutorials install viscy-utils from the GitHub monorepo and let the siblings resolve from PyPI. The writer they predict with now reads SlidingWindowDataset.z_planes, which the PyPI viscy-data 0.1.0 does not have (checked in its wheel), so predict would fail with AttributeError in on_predict_start. Install viscy-data from GitHub too, in the .py sources and their paired notebooks. Found by /code-review on #533. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… zero foreground_mask_from_volume guards a constant volume, but a blank volume with a few hot pixels (under 0.2% of voxels) has equal 0.1/99.9 percentiles too, and equalize_clahe then divided by zero and fed NaN to CLAHE (RuntimeWarning: divide by zero, reproduced on a (20, 64, 64) volume). equalize_clahe now returns zeros there, which the mask's existing flat-histogram guard turns into an empty mask, as for a constant volume. Volumes with any contrast are unchanged. Found by /code-review on #533. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…igger The injected latest-weights ModelCheckpoint takes over save_last from the monitored one but copied only every_n_epochs (defaulting to 1). A leaf that checkpoints every N steps, typical for max_steps runs with long epochs, would get last.ckpt only at epoch boundaries, so a preempted run resumed from much older weights. The injected callback now copies the monitored callback's every_n_train_steps or train_time_interval when set; epoch-cadence leaves are unchanged. No current leaf combines a monitored checkpoint with a step trigger. Found by /code-review on #533. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Brings in #533's /code-review fixes: viscy-data installed from GitHub in the VCP tutorials, the zero-range CLAHE guard, and the step-cadence inheritance of the latest-weights checkpoint. No conflicts. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Brings in #534's /code-review fixes (the walk skips only PAPER_KEY models the grouped generator does not register; the audit docstring; the seg-consistency engine and focus tests, identical to the copies this branch already added) and, through it, #533's fixes. Resolution: generate_grouped_eval_configs_test.py keeps dynacell-models' tests, including the Spotlight-arm skip test, plus #534's two scope-filter tests. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Copilot

🔵 Needs a closer look

Cross-application training defaults, prediction contracts, and mixed-precision numerics require final human integration review.

No open issues 🎉

_{Review effort: Balanced}

The guardrail turned save_last off only on the first monitored ModelCheckpoint. A leaf with two monitored checkpoints (say loss and a validation metric) kept a second last.ckpt writer next to the injected latest-weights callback, so last.ckpt would alternate between "latest" and "that monitor's cadence". All monitored callbacks now give it up; the first still names the monitor, directory and cadence. No current leaf has two monitored checkpoints. Found by /code-review on #533. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Brings in #533's fix that takes last.ckpt from every monitored checkpoint, not only the first. No conflicts. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Brings in #534's shared foreign-model predicate (walk and instance-AP audit), the single model-dir acceptance rule, the keep-raising rule for campaign models that do not parse, and #533's save_last fix. Resolutions: - generate_grouped_eval_configs.py: dynacell-models' version (Spotlight registrations and _SPOTLIGHT_ARM_MODELS skip) plus #534's _match_model_variant and is_foreign_model_dir in the walk; - evaluation/paths.py and tests/test_paths.py: dynacell-models' text, which names the Spotlight leaf generator this branch carries. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Copilot

🔵 Needs a closer look

The cross-application training guardrails, prediction changes, and precision-sensitive numerical updates warrant final human validation.

No open issues 🎉

_{Review effort: Balanced}

…inference off anndata 0.12.6 (capped below 0.12.9, which needs pandas<3) cannot write pandas 3's default Arrow-backed strings to zarr. The workaround was global: write_embedding_dataset set pd.options.future.infer_string = False in the caller's process, and the test conftest set it for the whole session. Tests therefore never ran under the configuration users have, and the other writers (EmbeddingSnapshot's rewrite, annotation convert) worked only after the embedding writer had flipped the option. write_anndata_zarr converts obs/var Arrow strings (columns, categories, index) to object just before writing, reusing the conversion append_to_anndata_zarr already had; all three writers and the tests use it, and both global switches are gone. New tests pin a round trip under the default, that plain write_zarr still fails (so the TODO is dropped deliberately once anndata 0.13 lands), and that the embedding writer leaves the option alone. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MzpgExfYMKCgogfnhcSgDc

dynaclr's pytest_configure set pd.options.future.infer_string = False for the whole pytest session, so in a repo-wide run every package's tests ran with pandas 3's default off (it hid the viscy-utils guard added in the previous commit). split_embeddings flipped it in the caller's process too. Both go; the eleven test writes and the split writer use viscy_utils' write_anndata_zarr instead. The cellanome embedding scripts keep their own switch: each owns its process and has no test to check a change. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MzpgExfYMKCgogfnhcSgDc

edyoshikun

LGTM! it fixes a lot of issues even for DynaCLR. Ill address #485 later

#533 was squash-merged as 4a4ac9f while this branch carried its commits only up to 0b1cb97, so the merge base is pre-#533 main and four files conflicted: - .github/workflows/test.yml, pyproject.toml, uv.lock: main's copy equals #533 at 0b1cb97 and this branch only adds to it, so this branch's side is kept (uv lock --check passes); - applications/dynaclr/tests/conftest.py: unchanged here since 0b1cb97, so main's side is taken (the pandas infer_string fix). No file changed on both sides after 0b1cb97, and every merged file equals this branch's copy where it changed the file, else main's. Full suite (pytest -n 11): 5269 passed, 0 failed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MzpgExfYMKCgogfnhcSgDc

* feat(viscy-utils): add the segmentation-aware and background losses Port the Spotlight-motivated losses from dynacell-models so they live in the shared package before any application consumes them: SegAuxDice (seg_aux.py) with the SAUNA-map and soft-clDice weight maps (seg_aux_maps.py), the segmenter-consistency loss (seg_consistency.py), the Track H background low-pass target (background_lowpass.py), and SpotlightLoss(return_components=True). Defaults keep SpotlightLoss's existing behavior; no package imports an application. Source on dynacell-models: 40de793 (#498), 6c28a31 (#521), 0cbeda1 (#524), 050f553 (#530). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(viscy-utils): build foreground masks with CLAHE, median and Otsu Port the Spotlight fg-mask recipe from dynacell-models. generate_fg_masks now binarizes a CLAHE-equalized, 5^3-median-filtered volume at an Otsu threshold picked on that volume after a sigma=5 Gaussian; compute_otsu uses the same full-resolution smoothing, so the otsu_grid_spacing parameter is gone from generate_normalization_metadata and VisCyTrainer.preprocess. The median and Gaussian run through cubic on the GPU when cubic and a device are present, else SciPy; cubic stays an optional import. torch is imported lazily so that light callers of write_meta_field do not pay for it. Behavior change: Spotlight masks and Otsu thresholds written by `viscy preprocess` differ from main's. Source on dynacell-models: 40de793 (#498), d087d96 (#510). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(viscy-utils): make prediction writes resumable, identified and write-once Port the HCSPredictionWriter and prediction_metadata work from dynacell-models: - z_reduction="center" and z_planes (slab-only predicts write only the planes they predicted, and record them); - per-FOV completion markers and the checkpoint/settings identity that predict resume and the eval caches key on (prediction_metadata.py); - refuse writes into an existing store when a FOV outside the run has unmarked prediction channels, or when the array would grow; - assemble each volume in memory and write it once on a background thread (bytes unchanged); - raise NotImplementedError when trainer.world_size > 1, instead of letting ranks race on one store. Source on dynacell-models: 2b848a6, 1be4466, 74fa18a (#497), 8e41498 (#511), 996adb4 (#526), 6c83fff (#531), 8e74a7e (#532). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(viscy-utils): guard every fit with checkpoint and health callbacks Port the training guardrails from dynacell-models. VisCyCLI now injects, on every `fit`: - a latest-weights ModelCheckpoint that owns last.ckpt (the monitored checkpoint gets save_last=False), so last.ckpt always means the latest epoch, inheriting the leaf's write cadence; - MonitorHealthCheck, which raises when the monitored metric is non-finite or bit-identical for 10 validation epochs; - OptimizerHealthCheck, which raises after 5 consecutive non-finite losses or on GradScaler collapse. Also adds the opt-in LinearLRCooldown callback and raises the CLI subprocess test timeout to 120 s (it timed out on main). Behavior change: this applies to cytoland, dynaclr and dynacell fits. Source on dynacell-models: 20e54db, daa7477, bc0f444, 74fa18a (#497), 92693ae, 2f9ef31 (#528). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(viscy-utils): keep SSIM finite under mixed precision; lazy exports Port two viscy-utils changes from dynacell-models: - ssim_25d / ms_ssim_25d: clamp the variances, bound the covariance by Cauchy-Schwarz through a safe sqrt, floor the stability constants symmetrically, and run MS-SSIM outside autocast, so 16-mixed training cannot NaN the gradient. SSIM/MS-SSIM loss values change slightly for MixedLoss users. - viscy_utils/__init__ resolves its exports lazily (PEP 562), keeping torch and lightning off light import paths. Source on dynacell-models: a556bc9 (#484), 130e034, cee4e41, 272051c, 59343ad, 8b2413f (#509). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(viscy-data): slab-only and read-once predict, uint8 foreground masks Port the viscy-data side of the dynacell-models predict and Spotlight work: - HCSDataModule.predict_z_planes and SlidingWindowDataset z_planes / read_z_planes, so a 2D model predicts only its focus slab; - cache_volume: with num_workers == 0 predict reads each (position, t) volume once and, when every transform is pointwise, normalizes it once; - foreground masks are read in their stored dtype (uint8) and cast to float on the device in on_after_batch_transfer. Behavior change: every app's predict at num_workers=0 takes the read-once path. Source on dynacell-models: 40de793 (#498), 996adb4 (#526), 6c83fff (#531), 8e74a7e (#532). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(viscy-models): per-axis CellDiff patches and mask-conditioning channels Port from dynacell-models: - patch_size may be a per-axis (z, y, x) tuple through the patch embedding, transformer unpatchify and ViT bottleneck, which lets CELLDiffNet run Z-preserving at Z=1 for the 2D track (an int is still accepted); - CELLDiffNet(cond_channels=1) adds extra conditioning input channels (default keeps the old single-channel net). Source on dynacell-models: 0126fe6, 0617ed7, 40de793 (#498). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * perf(viscy-models): non-uniform ODE time grids for CellDiff sampling Port the time_schedule option of the CellDiff ODE integrators from dynacell-models: uniform (default, unchanged), cosine, or a Beta(p, q) grid packed toward the data end, which is what lets the DynaCell-lite recipe sample in 14 velocity evaluations. scipy joins the celldiff extra for betaincinv. Source on dynacell-models: 19765e7 (#520). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(viscy-models): add the MorphEm feature extractor Port MorphEmModel from dynacell-models, the fifth deep-feature extractor of the DynaCell evaluation. transformers is imported lazily and the trust_remote_code model is pinned to a Hub commit. Source on dynacell-models: d3f6422 (#479), 4918d2e, 13cb336, c9c6914. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(viscy-models): let the PatchGAN discriminator preserve the Z axis Port preserve_z from dynacell-models, so the 3D PatchGAN runs on Z=1 inputs for pix2pix2d (default off, unchanged). Source on dynacell-models: c2d0481. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(viscy-models): UNet3D downsample_z for FNet-2D; reject non-5D input Port from dynacell-models: - UNet3D(downsample_z=True): False keeps Z through the encoder, which is how FNet2D runs on single planes (default unchanged); - UNet3DBase.forward raises on non-5D input, and a comment flags the ConvTranspose checkerboard artifact. Source on dynacell-models: a556bc9 (#484), 83a1d69, 16b1d31. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(viscy-transforms): stop BatchedRandAffined blanking singleton axes Port 9640537 from dynacell-models: a size-1 spatial axis (Z=1 for 2D models) gets an identity row in the affine, instead of being resampled off the volume and zeroed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * perf(viscy-transforms): mark the sampled normalizers pointwise NormalizeSampled and MinMaxSampled set pointwise = True, which lets the read-once predict path in viscy-data normalize a whole volume once instead of once per window. Source on dynacell-models: 8e74a7e (#532). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(viscy-transforms): make the Gaussian-noise statistics test deterministic Port 3409560 (#513) from dynacell-models. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * build: parallel test runs with pytest-xdist Port dfcf9a5 (#512) from dynacell-models: pytest-xdist joins the test group, and a root conftest.py pins OMP/MKL/OpenBLAS to one thread under each xdist worker. Without the pin every worker's torch sizes its pool to all cores, which made `-n 11` 10x slower than it should be and timed out subprocess tests. CLAUDE.md documents the usage. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test: register pytest markers at the root and make them strict Port the marker half of c50300f (#489) from dynacell-models. strict_markers = true turns an unregistered marker into a collection error, so a typo'd marker can no longer silently decorate nothing. hpc_integration moves from the cytoland and dynaclr conftests to the root: an addinivalue_line registration only holds when that conftest is in the invocation's ancestry. dynaclr drops its own testpaths so the root ini applies. requires_data is registered too: main's applications/dynacell/tools/generate_grouped_eval_configs_test.py uses it, and strict markers would otherwise drop that module at collection. (dynacell-models no longer uses the marker; the dynacell PR that follows removes the registration with it.) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * build: put the dynacell hf_demo on the pytest pythonpath Port the root half of 7b7cb43 (#523) from dynacell-models so the HF demo's tests import its app module. The directory already exists on main. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(dynaclr): join annotations on the full fov_name Port d1ea492 from dynacell-models. The orchestrated linear-classifier fixture joined on the first fov_name component, which failed 7 tests on main once pytest collects dynaclr's colocated *_test.py. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(cytoland): correct the Spotlight recipe comments Port the cytoland recipe comment fixes from 40de793 (#498): the preprocess flags take explicit values (`--compute_otsu true --compute_fg_masks true`), fg_threshold and nonzero_threshold are inert while fg_mask_key is set, and the Otsu-centred target is what makes 0 the FG/BG boundary. Comments only. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(cytoland): install the VCP tutorials' unpublished deps from GitHub Port the tutorial fix from d3f6422 (#479): cytoland and viscy-utils are not on PyPI yet, so the tutorials install them from the GitHub monorepo; their siblings resolve from PyPI. Adds the jupytext-paired notebooks (no outputs). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs: cut CLAUDE.md content the code or global memory already states Port 93ac234 from dynacell-models: drop the repo-layout tree, setup and common-command blocks, lint/docstring rules (pyproject.toml is the source) and the generic coding-philosophy sections. The DynaCell job runbook from the same file arrives with the dynacell PR, since it documents tools that only exist there. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * ci: run the viscy-utils tests in their own job viscy-utils had no CI job on main or on dynacell-models: four of its test files ran inside the dynacell job and the rest (losses, metrics, meta_utils, optimizer health, CLI, lazy imports) ran nowhere. Mirror the viscy-data job on ubuntu-latest for Python 3.12 and 3.13, and make the gate job wait for it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * ci(dynacell): run the launcher, engine and prediction-writer suites Port the test-dynacell-configs job from dynacell-models, minus the eight test files that only arrive with the dynacell PR: sync viscy-utils' test group with dynacell, and add test_submit_benchmark_batch, test_engine, test_pipeline_cache, feature_metrics_test and the viscy-utils prediction-writer, prediction-metadata, monitor-health and LR-cooldown tests. The dynacell-models PR target is not carried over. Source on dynacell-models: 33d2fee, 74fa18a (#497), 2d1ed32 (#506), 4450efd (#518), 2798dee (#519), 2f9ef31 (#528). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * build: re-lock against the ported packages and main's dynacell Start from dynacell-models' uv.lock and re-resolve with `uv lock` against this tree, whose applications/dynacell/pyproject.toml is still main's. Against dynacell-models' lock this drops the ten dynacell-only packages (dinov3, etils, ftfy, jsonpath-rw, mlcroissant, pandas-stubs, ply, rdflib, submitit, termcolor) and resolves cubic back to 0.7.0; the dynacell PR restores them. Against main's lock it adds pytest-xdist and moves boto3/botocore 1.43.0 -> 1.43.20, cellpose 4.1.1 -> 4.2.1.1, pyarrow 19.0.1 -> 24.0.0, s3transfer and tracksdata; torch and lightning are unchanged. `uv lock --check` exits 0 (main's own lock does not). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(dynacell): DynaCell v1.1 benchmark configs from dynacell-models Port every DynaCell config path of dynacell-models verbatim: the model leaves (baselines, A549/joint/HEK/brightfield, the 2D track with its seed replicates and warmup-matched refit, the temporal and patch-vs-augmentation ablations, CELL-Diff-R2 and the DynaCell-lite leaves), the grouped and instance-AP eval buckets of the paper models, the shared overlays, train/predict sets and launcher profiles, and the deletion of 155 stale config files. The leaves are generated; review the generators and guard tests in the code commits, not each leaf. Spotlight v2 arm leaves, their four fit overlays and their 231 spotlight_v2_* eval buckets are left out; they follow in their own PR. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(dynacell): DynaCell v1.1 application code from dynacell-models Port every DynaCell code path of dynacell-models verbatim (226 non-merge commits since 2026-06-16; PRs #479-#530): the evaluation pipeline (cpdino/MorphEm/focus-aware 2D, focus2d slab metrics, foreground-limited metrics, MicroMS3IM, scale-invariant pixel metrics, CP shared space, provenance and cache invalidation), the paths registry, preprocessing/registry/croissant/distribution, the launcher and stall watchdog, the lite benchmark and CELL-Diff fast inference, the 2D track, the HF demo, docs, and the engine, celldiff_wrapper and mask_conditioning modules exactly as on dynacell-models. The three LightningModules keep their Spotlight training hooks (seg_aux, seg_consistency, mask_mode, target_bg_lowpass). Every hook defaults off, the bit-identity tests guard that, and only Spotlight leaves turn them on; two DynaCell-paper 2D checkpoints store seg_aux_weight in their hparams, so a hook-free engine could not resume them. mask_conditioning.py comes along because engine.py and celldiff_wrapper.py import it. The files that import the Spotlight leaf generator are trimmed in the commits that follow. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * build(dynacell): lock the dynacell eval, croissant and preprocess deps Take dynacell-models' uv.lock verbatim, which this tree's pyprojects now match (`uv lock --check` exits 0). Against the shared-packages lock it moves cubic 0.7.0 -> 0.9.0a4 and adds dinov3 (pinned commit), etils, ftfy, jsonpath-rw, mlcroissant, pandas-stubs, ply, rdflib, submitit and termcolor. A fresh `uv lock` would instead float dinov3 to its current HEAD and bump mlcroissant/pandas-stubs, so the verbatim lock keeps the versions the paper evaluations ran with. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test: drop the requires_data marker registration Its only user was main's generate_grouped_eval_configs_test.py, which the DynaCell port replaced: the live-data tests there are now marked slow. With strict markers, nothing needs the registration. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * ci(dynacell): run the release-path, provenance and watchdog suites Restore the eight dynacell test files the shared-packages PR left out of the test-dynacell-configs job, now that they exist: test_evaluation_microssim_calibration, segmentation_test, test_distribution, test_croissant, test_main_cli_routing, test_evaluation_fid_provenance, provenance_test and watch_stalled_jobs_test. The job now matches dynacell-models'. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs: add the runbook for jobs that hang while holding an allocation Port the DynaCell job runbook of the root CLAUDE.md from dynacell-models (6193b9c, bcbfbe6, c4c5ed0, 74fa18a #497, e1d75cf): how to tell a finished-but-hung job from healthy compute by the CPU-time derivative, the stall watchdog (tools/watch_stalled_jobs.py), per-family predict and fit wall limits, and the forensics to capture before cancelling. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * refactor(dynacell): scope the grouped eval walk to registered model dirs The grouped generator imported ARMS from the Spotlight-v2 leaf generator, which is not part of this PR, to skip the Spotlight arms that share the canonical prediction tree. Drop that import and the Spotlight model names from _CODE_TO_PAPER, _DETERMINISTIC_MODELS and _CELLDIFF_MODELS, keeping the four 2D seed replicates the 2D appendix scores. The shared tree still holds the arms' stores, and parse_zarr_name raises on an unknown model token (e.g. the CellDiff "variant" segaux_iterative). walk_predictions now skips any model dir outside _REGISTERED_MODEL_DIRS before parsing and prints how many stores it skipped. The test that pinned the Spotlight skip becomes one that pins the unregistered-dir skip and its count. evaluation/paths.py keeps its Spotlight names: without them canonical_model_name would silently resolve an arm's run dir onto its baseline. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(dynacell): drop the Spotlight leaf entries from the config guards The Spotlight arm leaves are not part of this PR, so: - test_benchmark_config_composition: drop the eight v1 *_spotlight train leaves, their iPSC predict leaves and data-hparam entries, the _UNMIGRATED_MODELS filter that only they needed, and the Spotlight members of _WIDE_GPU_TRAIN_LEAVES and _LONG_WALL_TRAIN_LEAVES (the 2D pix2pix seed1 replicate keeps its wide-GPU opt-out); - test_paths: drop the two tests that import the Spotlight leaf generator or depend on the grouped generator registering the arms. The canonical_model_name / resolve_model tests of the arm names stay, since evaluation/paths.py keeps those names. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(dynacell): describe the 2D seed1 leaves without the Spotlight generator The 40 leaves of the four 2D baseline seed replicates were emitted by the Spotlight-v2 leaf generator, which is not part of this PR, and their headers told the reader to edit that generator. Say instead what they are for: the second draw scored by the *__2d_focus2d_seed1 buckets. Comments only; the configs are unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(dynacell): drop the Spotlight eval generator from the generator lists evaluation/README.md and the hand-written-bucket guard's comment in test_evaluate_compose.py named generate_spotlight_v2_eval_configs.py among the generators that write grouped buckets; it is not part of this PR. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(viscy-utils): declare torchvision, which evaluation.metrics imports viscy_utils.evaluation.metrics imports torchvision.ops at module level, and losses.MixedLoss imports that module, but viscy-utils never listed torchvision. Installs only worked because another workspace member pulled it in. The new test-utils CI job syncs viscy-utils alone and failed to collect six test modules with ModuleNotFoundError. The lock gains only the dependency edge; no version moves. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * ci(dynacell): keep main's batch-launcher test and the writer tests out Two corrections to the test-dynacell-configs job: - main's tests/test_submit_benchmark_batch.py resolves leaves under /hpc and fails on a runner (5 PermissionErrors in CI). The dynacell-models version that runs there arrives with the DynaCell PR, which adds it back. - the four viscy-utils test files (prediction writer, prediction metadata, monitor health, LR cooldown) now run in the test-utils job on both Python versions, so the dynacell job no longer repeats them or syncs viscy-utils' test group for them. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * refactor(viscy-utils): share the fit-callbacks lookup between CLI injectors _inject_checkpoint_guardrails and _inject_optimizer_health_guard each repeated the same walk from the config to fit's trainer.callbacks plus the already-injected check; the `else config` branch of that walk could never run behind the `subcommand != "fit"` return. Both now call _fit_callbacks_without. _callback_field collapses its two identical branches, and the monitored checkpoint's init_args are kept from the search loop instead of being looked up twice. No behavior change. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * refactor(viscy-utils): drop MonitorHealthCheck's unused check_finite flag Nothing sets check_finite (the CLI injects monitor and patience only, and no config or test passes it), so the callback always checks finiteness; it now does so with math.isfinite instead of building a tensor for one float. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(viscy-utils): stop describing fg masks as built from Otsu thresholds generate_fg_masks does not read the stored otsu_threshold; it binarizes a CLAHE-equalized, median-filtered volume. Its summary line and VisCyTrainer.preprocess's compute_fg_masks entry still said "from Otsu thresholds". Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * build: leave the hf_demo pytest path to the DynaCell PR Nothing on this branch imports from applications/dynacell/examples/ hf_demo: its only test, build_demo_data_test.py, arrives with the DynaCell PR, which adds the pythonpath entry back with it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(viscy-utils): guard a fit that sets no callbacks Trainer.callbacks defaults to None, and the CLI injectors returned early unless they found a list, so a fit that set no callbacks ran without OptimizerHealthCheck, the guard meant for every fit. The shared lookup now gives such a fit an empty list to append to (None and [] mean the same to Trainer). The checkpoint guardrails still need a monitored ModelCheckpoint and remain a no-op there. Reported by Copilot on #533. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(viscy-utils): say what the AMP skip counter actually counts The counter resets only when GradScaler grows its scale, so successful optimizer steps can sit between the reductions it counts; the error message claimed every step was skipped "in a row". Describe it as reductions without growth in between, which is what it detects: the backward overflowing faster than growth_interval clean steps can restore the scale. Behavior unchanged. Reported by Copilot on #533. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(viscy-utils): resolve viscy_utils.callbacks names lazily VisCyCLI now injects MonitorHealthCheck and OptimizerHealthCheck into every fit by class path through viscy_utils.callbacks, whose __init__ eagerly imported OnlineEvalCallback and with it scikit-learn, which only the eval extra installs. A fit on a base install would therefore fail before training. The package now resolves its exports lazily (PEP 562), like viscy_utils itself. The new test blocks sklearn and imports the health callbacks: the eager __init__ fails it with ModuleNotFoundError, the lazy one passes. Reported by Copilot on #533. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(cytoland): install viscy-data from GitHub in the VCP tutorials The tutorials install viscy-utils from the GitHub monorepo and let the siblings resolve from PyPI. The writer they predict with now reads SlidingWindowDataset.z_planes, which the PyPI viscy-data 0.1.0 does not have (checked in its wheel), so predict would fail with AttributeError in on_predict_start. Install viscy-data from GitHub too, in the .py sources and their paired notebooks. Found by /code-review on #533. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(viscy-utils): return an empty CLAHE result when the clip range is zero foreground_mask_from_volume guards a constant volume, but a blank volume with a few hot pixels (under 0.2% of voxels) has equal 0.1/99.9 percentiles too, and equalize_clahe then divided by zero and fed NaN to CLAHE (RuntimeWarning: divide by zero, reproduced on a (20, 64, 64) volume). equalize_clahe now returns zeros there, which the mask's existing flat-histogram guard turns into an empty mask, as for a constant volume. Volumes with any contrast are unchanged. Found by /code-review on #533. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(viscy-utils): keep a step-cadence leaf's last.ckpt on its step trigger The injected latest-weights ModelCheckpoint takes over save_last from the monitored one but copied only every_n_epochs (defaulting to 1). A leaf that checkpoints every N steps, typical for max_steps runs with long epochs, would get last.ckpt only at epoch boundaries, so a preempted run resumed from much older weights. The injected callback now copies the monitored callback's every_n_train_steps or train_time_interval when set; epoch-cadence leaves are unchanged. No current leaf combines a monitored checkpoint with a step trigger. Found by /code-review on #533. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(dynacell): skip only PAPER_KEY-known arms in the grouped eval walk The scope filter skipped every model dir outside the generator's registry and only printed a count. That also silenced the loud "unknown model code-name" a new, never-registered model used to raise, so its stores would drop out of every paper bucket without an error. Now the walk skips only dirs that paths.PAPER_KEY knows but this generator does not register (the Spotlight arms on this branch); anything unknown to both still raises in parse_zarr_name, as on dynacell-models. Tests pin both sides. Found by /code-review on #534. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(dynacell): say the instance-AP audit does not share the walk's arm skip audit_prediction_coverage claimed to apply the same skip filters as walk_predictions. It applies the same test-set filter, but on purpose not the walk's skip of PAPER_KEY models the grouped generator leaves unregistered: listing every unregistered model is what it is for. Found by /code-review on #534. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(dynacell): carry the seg-consistency engine and focus-parity tests test_seg_consistency_training.py exercises the seg_consistency hook of DynacellUNet and DynacellGAN with a stand-in segmenter, and test_seg_consistency_focus.py pins the loss's focus-plane rule to dynacell.evaluation.focus. Both test code this PR ships and import nothing from the Spotlight PR, so they belong here rather than with the Spotlight arms. Verbatim from dynacell-models. Found by /code-review on #534. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(viscy-utils): take last.ckpt from every monitored checkpoint The guardrail turned save_last off only on the first monitored ModelCheckpoint. A leaf with two monitored checkpoints (say loss and a validation metric) kept a second last.ckpt writer next to the injected latest-weights callback, so last.ckpt would alternate between "latest" and "that monitor's cadence". All monitored callbacks now give it up; the first still names the monitor, directory and cadence. No current leaf has two monitored checkpoints. Found by /code-review on #533. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * refactor(dynacell): one model-dir acceptance rule for parse and walk filter _REGISTERED_MODEL_DIRS restated, as a separate set, the rule _split_model_variant uses to accept a dir name, so the two could drift. _match_model_variant is now that rule (None when the dir does not parse); _split_model_variant wraps it and raises the same errors as before, and the walk filter calls it. No behavior change. Found by /code-review on #534. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(dynacell): share the foreign-model skip between the walk and the audit walk_predictions skipped the Spotlight arms' dirs on the shared tree, but the instance-AP coverage audit still parsed them, so on this branch it reported 468 coverage errors on the live tree instead of dynacell-models' 102, and its docstring claimed it applied the walk's filters. Both now call one predicate, is_foreign_model_dir: a dir whose model paths.PAPER_KEY knows but the grouped generator does not parse. Measured on the live tree, the audit's error set is identical to dynacell-models' (102 entries; the remainder are pre-existing). The new audit test holds one registered, one foreign and one unknown store and expects exactly the unknown one. Found by /code-review on #534 (findings 1, 9, 10). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(dynacell): keep raising on campaign models the generator cannot parse pix2pix2d_unetvit_wmatch{,_seed1} are in _CODE_TO_PAPER (and in PAPER_KEY) but not in _DETERMINISTIC_MODELS, so their stores do not parse. On dynacell-models the walk raises on them and the audit lists them. The foreign-model predicate would have skipped them as another campaign's arms. A model the generator itself lists in _CODE_TO_PAPER is not foreign: the walk raises and the audit flags it, as on dynacell-models. With this, the audit's error set on the live tree is identical to dynacell-models' (102 entries, compared path by path). Registering wmatch properly is a separate decision (its comment says "evaluate in focused sibling buckets", and adding it to _DETERMINISTIC_MODELS would fold it into the canonical buckets), so it is left as on dynacell-models. Found by /code-review on #534 (findings 2, 7). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(dynacell): stop pointing at the Spotlight leaf generator from this branch evaluation/paths.py and tests/test_paths.py cited tools/generate_spotlight_v2_leaves.py, which this branch does not have, and test_paths claimed a pin to it that only exists with the Spotlight v2 configs. Say where the generator lives instead. The Spotlight model names stay: without them canonical_model_name would resolve an arm onto its baseline. Comments only. Found by /code-review on #534 (finding 8). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(dynacell): parse the warmup-matched pix2pix2d stores in the grouped walk _CODE_TO_PAPER declares pix2pix2d_unetvit_wmatch{,_seed1} a parse-only registration "same rule as fnet2d_voxelmatched", but unlike that model they were never added to _DETERMINISTIC_MODELS. Their 56 stores on the live tree therefore raised "unknown model code-name", which alone stops walk_predictions. Register them like fnet2d_voxelmatched: they now parse, and, as for every parse-only ablation, they are scored in their focused sibling buckets (the committed *__2d_focus2d_pix2pix2d_wmatch leaves), not by regenerating the canonical ones. Fixes part of FOLLOWUPS #1 (live-tree walk crash). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(dynacell): skip the listed checkpoint re-predicts in the grouped walk The live tree holds second-checkpoint predictions of registered models beside their canonical stores, under train-set dirs paths.py does not know: fnet3d_paper/ipsc_step{8400,11k}, fnet2d_voxelmatched/ipsc_step200k, fcmae_vscyto{2d_scratch_samplematched,3d_scratch}/ipsc_final and fcmae_vscyto3d_scratch/ipsc_repredict (34 stores). key_from_prediction_store raises on them ("unknown train_set token 'ipsc_final'"), which stopped walk_predictions and filled the instance-AP audit. They come from the voxel-matched FNet-2D and sample-matched UNeXt2-2D ablations, are scored only by hand-written experiment buckets, and no committed config references them, so they are not canonical inputs. Each (model dir, train-set dir) pair is now listed in _CHECKPOINT_REPREDICTS with its campaign; the walk skips and counts them and the audit skips them. A train-set dir that is neither canonical nor listed still raises. Fixes part of FOLLOWUPS #1 (live-tree walk crash). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(dynacell): name the P1 segcons replicate dirs as foreign in the walk nucleus/pix2pix2d_unetvit_segcons_{s1,ctrl_s1,x10} (12 stores, Spotlight P1 follow-ups of 2026-10-19) sit in the canonical tree but are in neither this generator's registry nor paths.PAPER_KEY, so the PAPER_KEY-based foreign test cannot see them and they raised "unknown model code-name". They are now listed by name in _FOREIGN_MODEL_DIRS and skipped like the PAPER_KEY arms; the skip message no longer says "PAPER_KEY". Any other unknown dir still raises. The new audit test holds one store of each kind on the live tree (registered, checkpoint re-predict, named foreign, wmatch, unknown train set) and expects only the unknown one to be flagged. Fixes the rest of FOLLOWUPS #1 (live-tree walk crash). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(dynacell): list checkpoint re-predicts per organelle _CHECKPOINT_REPREDICTS matched (model dir, train-set dir) in every organelle, so a future mito/fnet3d_paper/ipsc_step8400 store would have been skipped as a known re-predict instead of raising, though only the ER one exists. Entries are now (organelle, model, train set) triples for exactly the stores on the live tree; a listed token in another organelle raises "unknown train_set token" again (new test, fails before this change). The audit test now uses the nucleus step-11k store, since the step-8400 one is ER-only. Found by /code-review on the FOLLOWUPS #1 fix. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * refactor(dynacell): one skip rule for the grouped walk and the audit walk_predictions and audit_prediction_coverage each called is_foreign_model_dir and is_checkpoint_repredict inline, so a third skip category added to one could silently miss the other and split them again. Both now call skip_reason(zarr_path), which names the category; the walk counts per category. No behavior change. Found by /code-review on the FOLLOWUPS #1 fix. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(dynacell): guard the grouped registries against the walk crash class Two guards on the hand-kept registries behind the live-tree crash: - every _CODE_TO_PAPER model must parse. At a04b3eb this lists pix2pix2d_unetvit_wmatch{,_seed1}, the omission that stopped walk_predictions; - the name-based skip lists stay disjoint from what the campaign registers: no _FOREIGN_MODEL_DIRS entry in _CODE_TO_PAPER or the parse tuples (it would be dropped silently, since the foreign check runs first), and no re-predict token among the canonical train sets. Found by /code-review on the FOLLOWUPS #1 fix. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(dynacell): reflow the checkpoint re-predict comment The organelle-scoping edit left a broken line wrap in the _CHECKPOINT_REPREDICTS comment. Comment only. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(dynacell): keep the wmatch refit out of canonical and instance-AP leaves Registering pix2pix2d_unetvit_wmatch{,_seed1} for parsing (97c30a1) fixed the live-tree walk, but a default regeneration would then also fold their 56 stores into the canonical {organelle}_{train_set} buckets and the instance-AP leaves, against DON'T-FOLD: the committed *__2d_focus2d_pix2pix2d_wmatch buckets are where they are scored. They are now listed in _SIBLING_ONLY_MODELS: walk_predictions still returns them, so sibling buckets can be built from it, but main() buckets canonical_predictions() (the walk minus that list) and the instance-AP in_scope() drops them. New tests: canonical_predictions leaves them out, in_scope rejects them. MEASURED on the live tree, both generators run the default way (no --out-root) emit 0 wmatch blocks, and the canonical pool (628 stores) is identical, store by store, to what the pre-fix rules parse, so a default regeneration writes exactly what a04b3eb would have written had its walk not crashed. Found by /code-review on the FOLLOWUPS #1 fix (finding 2). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(dynacell): retitle the paper and describe DynaCell as the broader framework The camera-ready title is now "Evaluating Dynamic 3D Virtual Staining of Live Cells with DynaCell": DynaCell names a broader framework for image-based profiling of cell-state dynamics, and this paper's evaluation framework is part of it. Update the title in CITATION.cff (software and preferred citation), the README and both HF card BibTeX entries, and the Croissant citeAs string; rewrite the CITATION.cff abstract and the README tagline to describe the evaluation framework as part of DynaCell. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015kuht3npKtb2o5E6h19xbA * docs(dynacell): call the contribution the DynaCell evaluation suite Mirror paper 1c85e84: "This evaluation framework is part of DynaCell, a broader framework" repeated the noun. The paper's contribution is now the DynaCell evaluation suite, and DynaCell stays the broader framework. The CITATION.cff abstract also says "a set of baseline models", so "suite" means one thing. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015kuht3npKtb2o5E6h19xbA * docs(dynacell): retitle to The DynaCell Evaluation Suite The paper's title is now "The DynaCell Evaluation Suite for Dynamic 3D Virtual Staining of Live Cells" (Overleaf b73f25f). Updates CITATION.cff, the README and HF card BibTeX, and the Croissant citeAs. test_croissant.py: 17 passed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015kuht3npKtb2o5E6h19xbA * chore(dynacell): delete finished one-shot tools Each did a one-time job that is done, and nothing else imports them: - build_migration_manifest, run_migration, restandardize_config_paths (+ tests): the canonical-path migration was applied 2026-07-13; a rebuilt manifest on 07-24 had nothing left to move; - retokenize_deconv_predictions (+ test): the 24 a549__deconv / joint__legacy_deconvgt renames were applied 2026-07-22; paths.py still reads both tokens for the dirs kept on purpose; - scan_nan_microssim_caches (+ test): one audit for the cubic 0.9.0a1 NaN MicroMS3IM caches that #506 fixed; - add_instance_ap_to_ablation_leaves: one codemod, already applied; - smoke_joint_leaf: an April CPU smoke the config-composition tests now cover. The provenance builders (cpdino seg masks, HEK stores, A549 dual fuse, random-init checkpoints) stay. README entries and one bake_best_ckpt docstring line follow. Full suite: 5251 passed, 0 failed (the 18 fewer are the two deleted tests/ files; the 14 colocated tool tests were never collected by the root run). Raised in review by edyoshikun. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MzpgExfYMKCgogfnhcSgDc --------- Co-authored-by: Alexandr Kalinin <alxndrkalinin@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

#534 was squash-merged into main as 7286fb2. This branch stacked on #534 but stopped six commits short of its final head (d7986fd): the paper retitle, the #533 squash merge, and the deletion of the finished one-shot tools. Merging that head first keeps the real ancestry, so the following merge of main resolves against #533's squash commit instead of the pre-split base. None of these six commits touches a file this branch changes, and no file here references a deleted tool. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01APoGhoKfkCho9Z1VMS5jsG

alxndrkalinin and others added 24 commits October 26, 2026 15:47

feat(viscy-models): let the PatchGAN discriminator preserve the Z axis

1bf1821

Port preserve_z from dynacell-models, so the 3D PatchGAN runs on Z=1 inputs for pix2pix2d (default off, unchanged). Source on dynacell-models: c2d0481. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

test(viscy-transforms): make the Gaussian-noise statistics test deter…

860d0f8

…ministic Port 3409560 (#513) from dynacell-models. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

build: put the dynacell hf_demo on the pytest pythonpath

cd3ed0f

Port the root half of 7b7cb43 (#523) from dynacell-models so the HF demo's tests import its app module. The directory already exists on main. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

alxndrkalinin requested a balanced review from Copilot October 26, 2026 23:10

Copilot started reviewing on behalf of alxndrkalinin October 26, 2026 23:11

View session

Copilot AI reviewed Oct 26, 2026

View reviewed changes

Comment thread packages/viscy-utils/src/viscy_utils/callbacks/optimizer_health.py Resolved

Comment thread packages/viscy-utils/src/viscy_utils/cli.py Resolved

Comment thread packages/viscy-utils/src/viscy_utils/losses/seg_consistency.py Resolved

alxndrkalinin and others added 3 commits October 26, 2026 16:25

alxndrkalinin changed the title ~~Port shared VisCy package changes from dynacell-models (DynaCell v1.1, 1/3)~~ Port shared VisCy package changes from DynaCell v1.1 Oct 27, 2026

alxndrkalinin requested a balanced review from Copilot October 27, 2026 00:39

Copilot started reviewing on behalf of alxndrkalinin October 27, 2026 00:40

View session

Copilot AI reviewed Oct 27, 2026

View reviewed changes

alxndrkalinin and others added 3 commits October 26, 2026 18:11

alxndrkalinin requested a balanced review from Copilot October 27, 2026 01:33

Copilot started reviewing on behalf of alxndrkalinin October 27, 2026 01:34

View session

Copilot AI reviewed Oct 27, 2026

View reviewed changes

alxndrkalinin mentioned this pull request Oct 27, 2026

DynaCell v1.1 #494

Closed

alxndrkalinin requested a balanced review from Copilot October 27, 2026 02:04

Copilot started reviewing on behalf of alxndrkalinin October 27, 2026 02:05

View session

Copilot AI reviewed Oct 27, 2026

View reviewed changes

alxndrkalinin requested a review from edyoshikun October 28, 2026 17:55

alxndrkalinin and others added 2 commits October 30, 2026 13:34

edyoshikun approved these changes Nov 3, 2026

View reviewed changes

alxndrkalinin merged commit 4a4ac9f into main Nov 3, 2026
30 checks passed

alxndrkalinin deleted the viscy-packages-from-dynacell-v1.1 branch November 3, 2026 18:45

Navigation Menu

Repository navigation

Port shared VisCy package changes from DynaCell v1.1 - #533

Port shared VisCy package changes from DynaCell v1.1#533
alxndrkalinin merged 39 commits into
mainfrom
viscy-packages-from-dynacell-v1.1

alxndrkalinin commented Oct 26, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

edyoshikun left a comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

alxndrkalinin commented Oct 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

What is ported, and from where

Edits relative to dynacell-models

Review-gate fixes on top of the port

Behavior changes that reach other apps

Conflicts with open PRs to main

Tests

Review guide

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🔵 Needs a closer look

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🔵 Needs a closer look

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🔵 Needs a closer look

Uh oh!

edyoshikun left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

alxndrkalinin commented Oct 26, 2026 •

edited

Loading

Edits relative to `dynacell-models`