RadTrain
Python template · sinapsis_anomalib.templates.anomalib_train
Overview
Training implementation for Anomalib models.
Rad — inherited documentation
RAD Lightning Module for anomaly detection.
Implements Retrieval-based Anomaly Detection (RAD), a training-free framework that stores anomaly-free features in a multi-layer memory bank and detects anomalies through image-level retrieval followed by position-aware patch matching.
Args: backbone (str): Name of the ViT backbone from timm. Defaults to ``"vit_base_patch16_dinov3"``. layers (list[int] | None): Block indices (0-based) to extract features from. Defaults to ``[3, 6, 9, 11]``. pre_trained (bool): Whether to use pre-trained backbone weights. Defaults to ``True``. k_image (int): Number of nearest-neighbor training images retrieved per test image. The paper uses ``150`` for MVTec-AD, ``900`` for VisA and Real-IAD, and ``48`` for 3D-ADAM. Defaults to ``150``. use_positional_bank (bool): Enable position-aware patch matching. Defaults to ``True``. pos_radius (int): Spatial neighborhood radius (in patch units). The paper uses ``1`` for MVTec-AD and 3D-ADAM, ``2`` for VisA, and ``0`` for Real-IAD. Defaults to ``1``. max_ratio (float): Fraction of highest anomaly pixels pooled for image-level score. ``0`` means use max. Defaults to ``0.01``. layer_weights (list[float] | None): Per-layer score fusion weights. ``None`` means uniform. Defaults to ``None``. bank_dtype (torch.dtype | str | None): Storage dtype of the memory bank. ``None`` keeps the backbone dtype. ``torch.float16`` halves bank memory and checkpoint size and speeds up scoring, at the cost of a small perturbation of the scores. Defaults to ``None``. pre_processor (PreProcessor | bool): Pre-processor instance or flag. Defaults to ``True``. post_processor (PostProcessor | bool): Post-processor instance or flag. Defaults to ``True``. evaluator (Evaluator | bool): Evaluator instance or flag. Defaults to ``True``. visualizer (Visualizer | bool): Visualizer instance or flag. Defaults to ``True``.
Note: RAD is training-free: fitting only fills the memory bank. The bank is stored in the model state, so checkpoints grow with the size of the anomaly-free training set.
Example: >>> from anomalib.data import MVTecAD >>> from anomalib.models import Rad >>> from anomalib.engine import Engine
>>> datamodule = MVTecAD() >>> model = Rad( ... backbone="vit_base_patch16_dinov3", ... layers=[3, 6, 9, 11], ... k_image=150, ... )
>>> engine = Engine() >>> engine.fit(model=model, datamodule=datamodule)
Capabilities
Capabilities: not setInstallation
Install the published package using pip or uv.
pip install sinapsis-anomalib==1.0.0uv pip install sinapsis-anomalib==1.0.0uv add sinapsis-anomalib==1.0.0Configuration
Effective fields from the inspector schema, including nested Pydantic model fields.
Optional
Default: null
Optional
Default: null
Required
Default: null
Optional
Default: null
Optional
Default: 150
Optional
Default: "vit_base_patch16_dinov3"
Optional
Default: true
Optional
Default: 0.01
Optional
Default: null
Optional
Default: 1
Optional
Default: true
Optional
Default: true
Optional
Default: null
Optional
Default: true
Optional
Default: true
Optional
Default: true
Optional
Default: null
Path to checkpoint for resuming training. If None, starts training from scratch.
Optional
Default: null
Optional
Default: null
Optional
Default: null
(TrainerConfig): General trainer arguments. For more details see: https://lightning.ai/docs/pytorch/stable/common/trainer.html#trainer-flags rad_init backbone <class 'str'>: default: vit_base_patch16_dinov3 layers… Read full field description
(TrainerConfig): General trainer arguments. For more details see: https://lightning.ai/docs/pytorch/stable/common/trainer.html#trainer-flags rad_init backbone <class 'str'>: default: vit_base_patch16_dinov3 layers <class 'NoneType'>: default: None pre_trained <class 'bool'>: default: True k_image <class 'int'>: default: 150 use_positional_bank <class 'bool'>: default: True pos_radius <class 'int'>: default: 1 max_ratio <class 'float'>: default: 0.01 layer_weights <class 'NoneType'>: default: None bank_dtype typing.Union[torch.dtype, str, NoneType]: default: None pre_processor typing.Union[torch.nn.modules.module.Module, bool]: default: True post_processor typing.Union[torch.nn.modules.module.Module, bool]: default: True evaluator typing.Union[anomalib.metrics.evaluator.Evaluator, bool]: default: True visualizer typing.Union[anomalib.visualization.base.Visualizer, bool]: default: True
Optional
Default: null
The devices to use for training (e.g., 1, [0, 1], "auto"). Defaults to "auto".
Optional
Default: "auto"
The maximum number of epochs to train for. Defaults to 5.
Optional
Default: 5
The minimum number of epochs to train for. Defaults to 1.
Optional
Default: 1
The hardware accelerator to use. Defaults to "auto".
Optional
Default: "cpu"
Optional
Default: null
Optional
Default: null
Optional
Default: null
Optional
Default: null
Optional
Default: null
Optional
Default: null
A descriptive name for the dataset.
Required
Default: null
The root directory where the dataset is located.
Optional
Default: null
A random seed for reproducibility in data splitting.
Optional
Default: null
Optional
Default: null
Optional
Default: null
Path to the directory containing normal images.
Required
Default: null
The number of subprocesses to use for data loading. Defaults to 8.
Optional
Default: 8
Path to the directory for abnormal images.
Optional
Default: null
Optional
Default: "from_test"
The number of samples per batch for evaluation. Defaults to 32.
Optional
Default: 32
Optional
Default: null
Optional
Default: "from_dir"
Optional
Default: 0.5
Optional
Default: 0.2
The number of samples per batch for training. Defaults to 32.
Optional
Default: 32
Optional
Default: 0.2
Configuration example
Generated from documented defaults. Replace any placeholders and complete required nested settings before running.
templates:
- template_name: "RadTrain"
class_name: "RadTrain"
attributes:
logger: null
rad_init:
layers: null
k_image: 150
backbone: "vit_base_patch16_dinov3"
evaluator: true
max_ratio: 0.01
bank_dtype: null
pos_radius: 1
visualizer: true
pre_trained: true
layer_weights: null
pre_processor: true
post_processor: true
use_positional_bank: true
callbacks: null
ckpt_path: null
threshold: null
train_root: null
trainer_args:
devices: "auto"
max_epochs: 5
min_epochs: 1
accelerator: "cpu"
image_metrics: null
normalization: null
pixel_metrics: null
logger_configs: null
callback_configs: null
folder_attributes:
name: "<required>"
root: null
seed: null
mask_dir: null
extensions: null
normal_dir: "<required>"
num_workers: 8
abnormal_dir: null
val_split_mode: "from_test"
eval_batch_size: 32
normal_test_dir: null
test_split_mode: "from_dir"
val_split_ratio: 0.5
test_split_ratio: 0.2
train_batch_size: 32
normal_split_ratio: 0.2
View original docstring example
The source example references CfaTrain, not RadTrain.
agent:
name: my_test_agent
templates:
- template_name: InputTemplate
class_name: InputTemplate
attributes: {}
- template_name: CfaTrain
class_name: CfaTrain
template_input: InputTemplate
attributes:
folder_attributes_config_path: 'path/to/config.yaml'
generic_key: 'my_generic_key'
callbacks: null
normalization: null
threshold: null
image_metrics: null
pixel_metrics: null
logger: null
callback_configs: null
logger_configs: null
ckpt_path: null
trainer_args:
devices: "0"
accelerator: gpu
max_epochs: null
cfa_init:
backbone: wide_resnet50_2
gamma_c: 1
gamma_d: 1
num_nearest_neighbors: 3
num_hard_negative_features: 3
radius: 1.0e-05
For a full list of options use the sinapsis cli: sinapsis info --all-template-names.
If you want to see all available models, please visit:
https://anomalib.readthedocs.io/en/v1.2.0/markdown/guides/reference/models/image/index.html.Inheritance
sinapsis_anomalib.templates.anomalib_train.AnomalibTrainEnvironment variables
Environment definitions belong to their canonical Python modules.