Experiment

The experiment section of the configuration file contains the settings that are used to run the experiment.

experiment:
    batch_size: 64
    model:
        name: model-name
        args:
            arg1: value1
            arg2: value2
            ...
    training:
        enabled: true
        args:
            arg1: value1
            arg2: value2
            ...
    scores:
        bona_fide: /path/to/bona_fide_scores.txt
        morphed: /path/to/morphed_scores.txt
        metrics: /path/to/metrics.json
    metrics:
      - name: metric1
        args:
          arg1: value1
          arg2: value2
      - name: metric2
      - ...

While many of the settings vary according to the model used, there are some settings that are common to all models.

The batch_size setting specifies the batch size to use when training the model.

The model setting specifies the model to use. It has two keys: name, which contains the name of the model, and args, which contains the arguments that will be passed when creating the model.

Inside the model setting you can specify the checkpoint field, which contains the path to a checkpoint file. If this field is specified, the model will be loaded from the checkpoint file instead of being created from scratch.

The training setting specifies the training settings. It has two keys: enabled, which is a boolean that specifies whether to train the model, and args, which contains the arguments that will be passed when training the model. If the enabled field is set to false, the model will not be trained, and the model specified in the model setting will run only in inference mode.

Each model has its own set of training arguments. For instance, a neural network has a much more complicated set of training arguments than a simple linear model.

The scores setting specifies the paths to the files containing the scores for the bona fide and morphed images. The scores are the result of the model evaluation on the test set images. The scores.metrics field contains the path to a JSON file where the metrics for each testing group will be saved. This field is optional.

The metrics setting specifies the metrics to use to evaluate the model. It contains a list of metrics, each of which has a name and an optional args field.

Available models

There are two main models already available in Revelio: a neural network and a random guesser.

Neural network

The configuration of a neural network is quite complex, and it is described in the Neural network page.

Random guesser

The random guesser is a simple model that always returns a random score between 0 and 1.

The random guesser has no arguments.

Available metrics

There are several metrics already available in Revelio: a complete list can be found in the Metrics page.