Configuration

Four files, four roots

A KonfAI config is YAML that says which Python objects to build and how they connect. Each workflow has one file with one root key; the model workflows share a workspace named by train_name, so use the same train_name in the files of one experiment.

How YAML becomes objects

Each key is a constructor argument: KonfAI reads the class's signature, takes each argument from the key of the same name (snake_case) and builds nested objects the same way. A missing key takes its default (konfai <COMMAND> --init writes every default into the file); null means None and turns a feature off; a required key must be written and the error names it; a value of the wrong shape is refused with its path; a key nothing reads is reported with the closest valid name. Running a config rewrites it with its resolved defaults.

The dataset

Every workflow reads its data through a Dataset: block. dataset_filenames is a list of path, path:format or path:flag:format entries, as the templates write them:

• ./Dataset:mha, ./DicomDataset:dicom, ./OmeDataset:omezarr;
• the flag a adds a dataset's cases to the others (union), i keeps only the cases present in all (intersection): ./Predictions/TRAIN_01/Dataset:i:mha;
• a manifest of explicit entries, ./cases.yml:omezarr, when the cases do not sit in the standard layout.

groups_src names the groups read from disk; each gives one or more tensors in groups_dest, each with its own transforms; is_input: true marks the model's inputs, the others are targets. transforms is case preprocessing, replayed on regions when streaming; patch_transforms runs on each patch. No transforms means the group reaches the model as stored.

Trainer:
  Dataset:
    dataset_filenames: [./Dataset:mha]
    groups_src:
      CT:
        groups_dest:
          CT:
            transforms:
              Standardize: {mean: None, std: None}
            is_input: true
      SEG:
        groups_dest:
          SEG: {}
    Patch:
      patch_size: [128, 128, 128]

subset and validation take the same selectors: a share of the cases, a list of names, a case-list file, or a start:end slice.

Patches

Patch.patch_size is the patch the model sees (default [128, 128, 128]); a 0 lets KonfAI size that axis, starting from the whole axis and, on a GPU out-of-memory at the first step, restarting with the fewest equal patches that fit. overlap defaults to 20 %. extend_slice adds neighbouring slices as channels for 2.5-D models (with patch_size[0] == 1). memory_budget decides whether the dataset stays in RAM or streams.

Models by classpath

A KonfAI model is a graph of named modules, either a Python class or the same graph written as a .yml. Pick it with classpath: segmentation.UNet.UNet for a Python class of the catalogue, default|UNet.yml for the YAML catalogue, UNet.yml for a file next to the config; its arguments go in a section named after the class. A module's path joins the names of its parents with :, and the config attaches losses and metrics to those paths under outputs_criterions (the segmentation example puts cross entropy on UNetBlock_0:Head:Conv, the logits, and Dice on UNetBlock_0:Head:Softmax, the probabilities). A model with no loss is refused. KonfAI also instantiates regular PyTorch and MONAI components named by classpath (monai.networks.nets:UNet, with the monai extra).

The exhaustive field tables: training, prediction, evaluation, transform.