Skip to content

Cuvis.AI Documentation

TrainRun Schema

cubert-hyperspectral/cuvis-ai

TrainRun Configuration Schema¶

Trainruns are the top-level experiment config. They compose pipeline, data, training, and optional plugin concerns into one runnable unit.

Core Shape¶

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

name: rx_demo
output_dir: ./outputs/${name}
loss_nodes: []
metric_nodes:
  - metrics
freeze_nodes: []
unfreeze_nodes: []

Required Fields¶

Field	Meaning
`name`	Experiment identifier
`pipeline`	Composed pipeline config
`data`	Data config: module, splits, params (see Data)
`training`	Training config
`output_dir`	Output root

Common Optional Fields¶

Field	Meaning
`loss_nodes`	Loss node names for gradient training
`metric_nodes`	Metric node names to log/evaluate
`freeze_nodes`	Node names frozen at startup
`unfreeze_nodes`	Node names unfrozen for later phases
`tags`	Metadata for run tracking

Data¶

data is a DataConfig: which DataModule to load, how it is split, and module params.

Field	Meaning
`data_module`	Registered module name (e.g. `cu3s`, `cu3s_multi`, `tiff_paired`, `npz_multi`)
`splits`	A selector split (`DataSplitConfig`) or a `splits_path` to a committed `splits.json`. Omit for a module that owns its split.
`batch_size` / `num_workers`	DataLoader options
`params`	Module-specific arguments (e.g. `cu3s_file_path`, `annotation_json_path`; `universe_csv` for `cu3s_multi` and `npz_multi`)

A selector split assigns samples to stages by identity. A universe_csv (a universe.csv) supplies an explicit sample universe; cu3s_multi and npz_multi both read one (only tiff_paired enumerates from disk). See Data Splits for the full model (universe, selectors, baking).

data:
  data_module: npz_multi
  batch_size: 4
  splits:                       # inline selectors, or: splits_path: splits/dinomaly.json
    train:
      - { kind: file_indices, source: X.cu3s, ids: [0, 2, 3] }
    val:
      - { kind: file_indices, source: X.cu3s, ids: [1, 5] }
  params:
    universe_csv: outputs/npz_local/universe.csv

Current Patterns¶

Statistical Workflow¶

defaults:
  - /pipeline/anomaly/rx@pipeline: rx_statistical
  - /data@data: lentils
  - /training@training: default_statistical
  - _self_

name: rx_statistical_demo
metric_nodes:
  - metrics

SAM3 Workflow¶

defaults:
  - /pipeline/sam3@pipeline: sam3_text_propagation
  - /data@data: tracking_cap_and_car
  - /training@training: default
  - _self_

name: sam3_text_demo
output_dir: ./outputs/${name}