Anomalib Training

Training a Model on a Dataset

Training in anomalib always goes through anomalib.engine.Engine, which wraps a Lightning Trainer.

Python API — standard benchmark dataset (MVTecAD)

from anomalib.data import MVTecAD
from anomalib.models import Patchcore
from anomalib.engine import Engine

datamodule = MVTecAD(root="./datasets/MVTecAD", category="bottle", train_batch_size=32)
model = Patchcore()
engine = Engine()                      # any Lightning Trainer kwarg can go here
engine.fit(model=model, datamodule=datamodule)
results = engine.test(model=model, datamodule=datamodule)

Engine(**kwargs) forwards unknown kwargs straight to the underlying Lightning Trainer (accelerator, devices, strategy, max_epochs, logger, callbacks, enable_checkpointing, val_check_interval, barebones, ...) — there is no separate "Trainer config object" to build.

Key Engine methods: fit(model, datamodule=...), train(...) (fit + test in one call), test(model=None, datamodule=...), predict(model=None, datamodule=..., dataset=..., data_path=...). If model/datamodule are omitted from test/predict, the engine reuses the ones passed to fit.

Python API — custom data with the Folder datamodule

Use Folder whenever your data is laid out as root/normal_dir/*, root/abnormal_dir/*, and optionally root/mask_dir/* (segmentation masks) — no new dataset code needed:

from anomalib.data import Folder
from anomalib.models import Padim
from anomalib.engine import Engine

datamodule = Folder(
    name="custom",             # required — used as the datamodule's display name
    root="./datasets/custom",
    normal_dir="good",         # required
    abnormal_dir="defect",     # optional: enables anomalous test/eval samples
    mask_dir="mask",           # optional: enables pixel-level (segmentation) evaluation
    train_batch_size=32,
    eval_batch_size=32,
    num_workers=8,
)
model = Padim()
engine = Engine()
engine.fit(model=model, datamodule=datamodule)

If you only have normal training images and unlabeled/normal-only test images, omit abnormal_dir and mask_dir — Folder will still produce a valid train/test split via test_split_ratio (fraction of normal training images held out for testing). Use normal_test_dir if you have a separate directory of normal images specifically for the test set.

CLI

# Standard dataset, defaults
anomalib train --model Patchcore --data anomalib.data.MVTecAD

# Override a datamodule field
anomalib train --model Patchcore --data anomalib.data.MVTecAD --data.category transistor

# Override a trainer field (use a gradient-trained model like Stfpm where max_epochs is meaningful)
anomalib train --model anomalib.models.Stfpm --data anomalib.data.MVTecAD --trainer.max_epochs 3

# Custom Folder dataset from the CLI
anomalib train --model Padim --data anomalib.data.Folder \
  --data.name custom --data.root ./datasets/custom \
  --data.normal_dir good --data.abnormal_dir defect

# From a config file (jsonargparse; combine with any of the above overrides)
anomalib train --config path/to/config.yaml

CLI wiring lives in src/anomalib/cli/cli.py (AnomalibCLI); model/data classes are exposed as jsonargparse subclass arguments, so --model/--data accept either a short name (Padim) or a full class path (anomalib.models.Padim, anomalib.data.Folder).

Choosing accelerator / devices

Pass standard Lightning kwargs to Engine(...): accelerator="gpu"|"cpu"|"xpu", devices=1. For Intel XPU specifically, use SingleXPUStrategy/XPUAccelerator from anomalib.engine:

from anomalib.engine import Engine, SingleXPUStrategy, XPUAccelerator
engine = Engine(strategy=SingleXPUStrategy(), accelerator=XPUAccelerator())

Gotchas

• Not every model trains via gradient descent — training-free models (e.g. Padim, Patchcore) still go through engine.fit(...); Engine/Trainer handles the single "epoch" needed to build their memory bank. You don't need special-case code for this. Note that these models override max_epochs via their trainer_arguments property (e.g. max_epochs=1), so any max_epochs you pass to Engine will be overwritten for such models.
• Folder's mask_dir is what switches evaluation from image-level (classification) to pixel-level (segmentation) metrics — only pass it if you actually have per-pixel ground-truth masks.
• Results (checkpoints, logs, images) are written under Engine's default_root_dir ("results" by default), nested by model/datamodule/category — check there first when debugging a run.

Reviewer / self-check

• Engine(...) receives Trainer overrides as plain kwargs, not a hand-built Trainer object.
• Folder-dataset training specifies normal_dir and, if applicable, abnormal_dir/mask_dir matching the actual on-disk layout.
• CLI invocations use anomalib.data.<Class> / anomalib.models.<Class> paths that are actually exported (see anomalib-adding-a-model / anomalib-adding-a-datamodule for how exports work).