Dataset Structure

Default Dataset Setup

The AUCMEDI AutoML expects a fixed dataset structure if run on default parameters.

A custom named data directory (in this example aucmedi.data) holds the two required image directories: training for the training process and test for the prediction process.

aucmedi.data/
├── training/                     # Required for training
│   ├── class_a/
│   │   ├── img_x.png
│   │   └── ...
│   ├── class_b/                  # Subdirectory for each class
│   │   ├── img_y.png
│   │   └── ...
│   ├── class_c/
│   │   ├── img_z.png             # Image names have to be unique
│   │   └── ...                   # between subdirectories
│   └── ...
├── test/                         # Required for prediction
│   ├── unknown_img_n.png
│   └── ...
├── model/                        # Will be created by training
├── xai/                          # Will be created by prediction (optional)
├── evaluation/                   # Will be created by evaluation
└── preds.csv                     # Will be created by prediction

The dataset structure is by default in the working directory for CLI or is mounted as volume into the container for Docker.

Dataset I/O Parameters

The dataset structure can be customized with the following parameters:

AutoML Mode	I/O	Parameter	Default
Training	Input	path_imagedir	'training'
Training	Output	path_modeldir	'model'
Prediction	Input	path_imagedir	'test'
Prediction	Input	path_modeldir	'model'
Prediction	Output	path_pred	'preds.csv'
Prediction	Output	xai_directory	'xai'
Evaluation	Input	path_imagedir	'training'
Evaluation	Input	path_pred	'preds.csv'
Evaluation	Output	path_evaldir	'evaluation'

CSV Dataset (multi-label)

Instead of utilizing the directory interface for passing the training data class annotations via subdirectories, it is also possible to pass them as CSV.

This is required for setup a multi-label classification pipeline.

The default dataset structure then would look like this:

aucmedi.data/
├── training/                     # Required for training
│   ├── img_x.png
│   ├── img_y.png                 # Single directory with all training images
│   ├── img_z.png
│   └── ...
├── test/                         # Required for prediction
│    ├── unknown_img_a.png
│    ├── unknown_img_b.png
│    └── ...
└── annotations.csv               # CSV annotations for training

The CLI command for multi-label CSV data would then look like this:

aucmedi training --path_gt annotations.csv --ohe

One-hot encoded CSV Example

More information for CSV files can be found here: IO_CSV Interface