Configuration¶

CUVIS.AI uses Hydra composition and checked-in YAML files under configs/ as the current source of truth for pipeline, data, training, trainrun, and plugin configuration.

Directory Layout¶

configs/
├── data/
├── pipeline/
│   ├── anomaly/
│   └── sam3/
├── plugins/
├── training/
└── trainrun/

Configuration Types¶

Config type	Location	Purpose
Pipeline	`configs/pipeline/`	Node graph structure and connections
Data	`configs/data/`	Dataset paths, splits, and loader settings
Training	`configs/training/`	Trainer, optimizer, scheduler, callbacks
Trainrun	`configs/trainrun/`	Composed experiment definition
Plugins	`configs/plugins/`	Plugin manifests and registry

Current Pipeline Shape¶

Use the current schema keys only:

metadata:
  name: RX_Statistical
  description: RX anomaly detector
  author: cuvis.ai

nodes:
  - name: data_node
    class_name: cuvis_ai.node.data.LentilsAnomalyDataNode
    hparams:
      normal_class_ids: [0, 1]

  - name: normalizer
    class_name: cuvis_ai.node.normalization.MinMaxNormalizer
    hparams:
      eps: 1.0e-06
      use_running_stats: true

connections:
  - source: data_node.outputs.cube
    target: normalizer.inputs.data

Current rules:

Use class_name, not class.
Use hparams, not params.
Use source and target, not from and to.

Trainrun Composition¶

Trainruns compose pipeline, data, and training groups:

# @package _global_
defaults:
  - /pipeline/anomaly/rx@pipeline: rx_statistical
  - /data@data: lentils
  - /training@training: default
  - _self_

name: rx_lentils_demo
output_dir: ./outputs/${name}
metric_nodes:
  - metrics

Typical command-line overrides:

uv run python train.py training.trainer.max_epochs=50 data.batch_size=4

Shipped Pipeline Families¶

Family	Directory	Notes
RX	`configs/pipeline/anomaly/rx/`	Baseline anomaly workflows
Deep SVDD	`configs/pipeline/anomaly/deep_svdd/`	Deep one-class workflows
AdaCLIP	`configs/pipeline/anomaly/adaclip/`	CLIP-based anomaly workflows
SAM3	`configs/pipeline/sam3/`	Text, bbox, mask, and segment-everything tracking

Plugin Manifests¶

Current checked-in manifests:

configs/plugins/registry.yaml
configs/plugins/adaclip.yaml
configs/plugins/bytetrack.yaml
configs/plugins/deepeiou.yaml
configs/plugins/detr.yaml
configs/plugins/sam3.yaml
configs/plugins/trackeval.yaml
configs/plugins/ultralytics.yaml

Use the narrow manifest for a workflow when possible; use registry.yaml only when you intentionally want the full registry.