Using your own data or built-in samples

Your own data

RAITAP can load your own data by pointing data.source to a local file or directory.

Supported inputs include:

• A directory of images such as .jpg, .png, .bmp, or .webp (walked recursively — nested ImageFolder-style layouts are supported)

• A single image file

• A CSV, TSV, or Parquet file

• A directory containing CSV, TSV, or Parquet files

Example (flat directory):

YAML

data:
  source: "./data/images" # a directory of images
  labels:
    source: "./data/labels.csv"
    id_column: "image"
    column: "label"

Python

from raitap.data import DataConfig, LabelsConfig

data = DataConfig(
    source="./data/images",  # a directory of images
    labels=LabelsConfig(
        source="./data/labels.csv",
        id_column="image",
        column="label",
    ),
)

with labels.csv rows like:

image,label
IM-0001.jpeg,0
IM-0002.jpeg,1

The ids are bare filenames (no path separators), so auto matches by stem. See How labels match to samples below.

Example (nested ImageFolder layout — data/test/<class>/<file>.jpg):

data/test/
├── NORMAL/IM-0001.jpeg
├── NORMAL/IM-0002.jpeg
└── PNEUMONIA/IM-0001.jpeg   # colliding stem with NORMAL/

YAML

data:
  source: "./data/test"
  labels:
    source: "./data/labels.csv"
    id_column: "image"
    column: "label"
    # id_strategy: "auto"   # default — relative paths auto-detected

Python

from raitap.data import DataConfig, IdStrategy, LabelsConfig

data = DataConfig(
    source="./data/test",
    labels=LabelsConfig(
        source="./data/labels.csv",
        id_column="image",
        column="label",
        # id_strategy=IdStrategy.auto,  # default — relative paths auto-detected
    ),
)

with labels.csv rows like:

image,label
NORMAL/IM-0001.jpeg,0
NORMAL/IM-0002.jpeg,0
PNEUMONIA/IM-0001.jpeg,1

The default labels.id_strategy: "auto" detects the path separators and matches by relative path (extension is stripped during comparison, so NORMAL/IM-0001.jpeg and NORMAL/IM-0001 both work). Sample order is sorted by relative posix path. See Configuration for the full id_strategy reference.

How labels match to samples

Labels come from the data.labels file, not from folder names. RAITAP does not infer the class from a parent directory the way torchvision ImageFolder does. Each labels row is tied to a sample two ways:

• By id — when labels.id_column is set, its value is matched against the discovered sample files.

• By order — when no id_column is set, labels align to samples by sorted file order (row 1 to the first file, and so on).

For id matching, id_strategy controls how both sides are normalised before the lookup. The same rule is applied to the sample paths (from disk) and the label ids (from the file):

id_strategy	strips	NORMAL/IM-0001.jpeg becomes	use when
relative_path	extension only	NORMAL/IM-0001	label ids carry the directory (manifest)
stem	directory + ext	IM-0001	flat dir, label ids are bare filenames
auto (default)	picks per id	depends on the id column	leave it; detects which form fits

auto switches to relative_path as soon as any label id contains / or \, otherwise stem. The manifest form (relative_path/auto) is the conventional one and the safe default. stem collapses ids that share a filename across subdirs (NORMAL/IM-0001.jpeg and PNEUMONIA/IM-0001.jpeg both become IM-0001), which raises a duplicate-id error.

Labels from directory structure

If your images are already organised into one folder per class (the torchvision ImageFolder convention), set labels.source: "directory" to use the folder names as labels. No labels file needed.

data/train/
├── NORMAL/IM-0001.jpeg      # label: NORMAL
├── NORMAL/IM-0002.jpeg
└── PNEUMONIA/IM-0001.jpeg   # label: PNEUMONIA

YAML

data:
  source: "./data/train"
  labels:
    source: "directory"

Python

from raitap.data import DIRECTORY_LABELS_SOURCE, DataConfig, LabelsConfig

data = DataConfig(
    source="./data/train",
    labels=LabelsConfig(source=DIRECTORY_LABELS_SOURCE),  # == "directory"
)

The class is each sample's top-level subdirectory; nesting within a class folder is fine. Class ids are assigned alphabetically (NORMAL to 0, PNEUMONIA to 1). id_column and id_strategy do not apply. If images sit directly under the source with no class subdirs, RAITAP warns and falls back to predictions as metric targets.

If you want to evaluate metrics against ground-truth labels, configure the optional data.labels block as described in Configuration.

Built-in samples

RAITAP also includes a few built-in sample datasets that can be referenced by name through data.source.

Available sample names (registered in src/raitap/data/samples.py) are:

• imagenet_samples — 4 ImageNet test images (tench, shih-tzu, golden retriever, tiger cat), with bundled labels.csv.

• isic2018 — small ISIC 2018 dermoscopy subset (no labels).

• malaria — small malaria thin-blood-smear subset (no labels).

• acas_xu_n1_1 — ACAS Xu N1_1 tabular sample (no labels).

• UdacitySelfDriving — small Udacity self-driving subset (no labels).

Example:

YAML

data:
  source: "imagenet_samples"

Python

from raitap.data import DataConfig

data = DataConfig(source="imagenet_samples")

Built-in samples are useful for quickly testing the pipeline without preparing your own dataset first. RAITAP downloads them to a local cache automatically when needed.