Common configuration patterns

This page gathers the conventions that are most important in day-to-day KonfAI usage.

dataset_filenames

dataset_filenames accepts several forms:

  • ./Dataset

  • ./Dataset:mha

  • ./Dataset:a:mha

  • ./Predictions/TRAIN_01/Dataset:i:mha

From konfai.data.data_manager.Data.get_data():

  • no suffix means default format mha

  • a appends cases

  • i keeps only the intersection of cases

This is especially useful during evaluation, where you often combine:

  • the ground-truth dataset

  • the prediction dataset

groups_src / groups_dest

Use groups_src to declare what exists on disk and groups_dest to declare how it should appear inside the workflow.

This allows patterns such as:

  • loading MASK from disk but not feeding it to the model

  • loading the same source group into multiple transformed destinations

  • renaming groups logically inside the workflow

outputs_criterions

Use outputs_criterions when you need:

  • multi-head supervision

  • multiple criteria per output

  • different targets for the same output

  • scheduler-weighted loss composition

The keys are model output paths. Always start from a working example and only then introduce custom output names.

outputs_dataset

Use outputs_dataset when you need to control:

  • which prediction is exported

  • how predictions are reduced across TTA or ensembles

  • which final transforms run before the file is written

  • which geometry should be reused for the output

subset and validation

Use:

  • subset to restrict the whole run to specific items

  • validation to carve out a validation split during training or to define a validation report during evaluation

Supported subset forms are inferred from the dataset loader code:

  • None

  • 0:10

  • ./Subset.txt

  • ~./Exclude.txt

  • [0, 1, 2]

  • ["CASE_001", "CASE_002"]

  • ["./SubsetA.txt", "./SubsetB.txt"]

Supported validation forms are inferred from the dataset loader code:

  • None

  • 0.2

  • 0:10

  • ./Validation.txt

  • [0, 1, 2]

  • ["CASE_001", "CASE_002"]

  • ["./ValidationA.txt", "./ValidationB.txt"]

Notes:

  • subset: None keeps the full dataset

  • validation: None keeps the full subset for training and disables the validation split

  • ~... exclusion applies to subset, not to validation

Local modules next to YAML files

When you want to extend KonfAI without packaging a new Python distribution, put Python files next to your YAML and use explicit classpaths:

Model:
  classpath: Model:UNetpp5

final_transforms:
  UnNormalize:UnNormalize: {}

This pattern is used heavily in examples/Synthesis.

Notes on inferred behavior

The following conventions are inferred directly from the code rather than being formally described elsewhere in the repository:

  • default|... fallback values

  • the precise semantics of dataset_filenames flags

  • the way ;accu; addresses patch-wise outputs before reassembly

These patterns are stable enough to document because they are visible in the config loader, patching logic, and shipped examples.

See also