Using the Tiled Ensemble Pipeline

The tiled-ensemble pipeline splits each image into overlapping tiles, trains/evaluates a separate model instance per tile position, then merges tile predictions (with optional seam smoothing) back into a full-image anomaly map. Use it for high-resolution images where a single model can't see fine detail at a manageable input size.

Code locations

• src/anomalib/pipelines/tiled_ensemble/train_pipeline.py — TrainTiledEnsemble: composes the job graph (per-tile training, per-tile prediction, merge, seam smoothing, statistics) and picks SerialRunner or ParallelRunner based on the configured accelerator and available CUDA devices.
• src/anomalib/pipelines/tiled_ensemble/test_pipeline.py — EvalTiledEnsemble: runs inference/evaluation for an already-trained ensemble.
• src/anomalib/pipelines/tiled_ensemble/components/ — individual job implementations (model training, prediction, merging, smoothing, metrics).
• src/anomalib/pipelines/tiled_ensemble/components/utils/ensemble_engine.py — TiledEnsembleEngine, an Engine subclass that customizes per-tile checkpoint/workspace naming.