PathForge default configurations

These self-contained templates use placeholder paths under /path/to/.... Update the annotation, slide, artifact, project, and Optuna storage paths before running them.

Foundation-model names follow LazySlide's runtime identifiers:

• h-optimus-1 = HOptimus1
• uni2 = UNI2
• virchow2 = Virchow2
• gpfm = GenBioPathFM

The benchmark templates evaluate the requested 224-pixel tiles at 20× and 10× (tile_mpp: [0.5, 1.0]), with Macenko and no stain normalization. MIL tasks use ABMIL, DSMIL, and TransMIL. Slide retrieval does not train a prediction model, so losses and MIL architectures do not apply to that task.

ABMIL, DSMIL, and TransMIL are adapter-backed models. Their usable task/output combinations depend on the installed TorchMIL or MIL-Lab version and compatible constructor kwargs. The templates satisfy PathForge's configuration schema, but users must confirm that the selected upstream implementation emits the output shape required by regression or the chosen survival formulation.

Files prefixed with benchmark_ enumerate task-supported fixed pipeline grids. Files prefixed with optimize_ use Optuna ranges plus categorical pipeline choices. The parallel templates show generated SLURM workflows; parallel optimization needs a shared PostgreSQL database and must not use SQLite across nodes. Set execution.slides_per_shard in either parallel template to process multiple slides sequentially in each feature-extraction array task. Larger values reduce the number of submitted tasks while increasing each task's runtime; the default value of 1 preserves one task per slide.

The current Optuna policy is MIL-training-specific. The optimize_slide_retrieval.yaml file documents the intended retrieval search space, but it is not executable until PathForge gains a retrieval objective adapter. Use benchmark_slide_retrieval.yaml for supported retrieval grid search today.

Create and submit a distributed plan with:

pathforge execution plan --config default_config/parallel_benchmark.yaml --output work/benchmark
bash work/benchmark/slurm/submit.sh

Validate any edited template with:

python -c "from pathforge.config.config import Config; Config.from_yaml('CONFIG.yaml')"