anomalib-tiled-ensemble

# 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. ## Running it ```bash python tools/tiled_ensemble/train.py --config tools/tiled_ensemble/ens_config.yaml python tools/tiled_ensemble/eval.py --config tools/tiled_ensemble/ens_config.yaml \ --root results/Padim/MVTecAD/bottle/v0 ``` `train.py` runs `TrainTiledEnsemble().run()` which includes evaluation after training; `eval.py` runs `EvalTiledEnsemble` to **re-run** evaluation against an existing results directory (`--root`) — use it only when you want to evaluate again without retraining. ## Config structure Start from `tools/tiled_ensemble/ens_config.yaml` and adjust the fields you need: ```yaml seed: 42 accelerator: "cuda" # or "cpu" default_root_dir: "results" tiling: image_size: [256, 256] # size the full image is resized to before tiling tile_size: [128, 128] # size of each tile stride: 128 # tile stride; stride < tile_size gives overlapping tiles normalization_stage: image thresholding_stage: image data: class_path: anomalib.data.MVTecAD init_args: root: ./datasets/MVTecAD category: bottle train_batch_size: 32 eval_batch_size: 32 num_workers: 8 val_split_mode: from_test test_split_mode: from_dir SeamSmoothing: apply: False sigma: 2 width: 0.1 TrainModels: model: class_path: Padim ``` Key fields: - `tiling.tile_size` / `tiling.stride` — the core tiling geometry; `stride < tile_size` produces overlap that `SeamSmoothing` then blends. - `data.class_path` — any **image** `anomalib.data.*` datamodule that yields `ImageBatch` (see `anomalib-training` / `anomalib-adding-a-datamodule`). Video and depth datamodules are **not supported** — the tiled collater uses `ImageBatch.collate` internally. - `TrainModels.model.class_path` — the model class trained per tile; must be a standard **image** model that only requires `batch.image` as input. Models requiring additional inputs (e.g. CFM which needs `point_cloud`/`depth_map`) are not compatible with the tiled collater. Video models are also not compatible. - `SeamSmoothing.apply` — when `True`, applies Gaussian blending at tile boundaries. This is most useful when tiles overlap (`stride < tile_size`), but can also smooth hard boundaries between non-overlapping tiles. Set `False` to skip if seam artifacts are not visible. For a worked reference invocation with a full config, see `tests/integration/pipelines/test_tiled_ensemble.py`. ## Gotchas - The pipeline trains **one model instance per tile position**, not one shared model — total training cost scales with the number of tiles, not just image count. Budget accordingly before scaling up `tile_size`/`stride` combinations. - `accelerator: cuda` with multiple visible GPUs triggers `ParallelRunner`, which trains multiple tile jobs concurrently across devices — set `accelerator: cpu` (or restrict visible devices) for deterministic single-process runs while debugging a config. - Eval (`eval.py`) needs `--root` pointing at the exact output directory produced by the matching training run; it does not re-derive this automatically. - `data.init_args` **must** include `val_split_mode` and `test_split_mode` — the pipeline reads these directly from the config before datamodule defaults are applied, and will raise `KeyError` if missing. ## Reviewer / self-check - [ ] `tiling.tile_size`/`stride` chosen relative to `tiling.image_size` (stride ≤ tile_size). - [ ] `data.class_path` and `TrainModels.model.class_path` both resolve to real, exported classes. - [ ] `SeamSmoothing.apply` is intentional given whether tiles overlap. - [ ] Training run completed and its `results/...` path is used correctly as `eval.py --root`.