Overview

InfoProvider is an abstract class that provides convenient information log for different classes (Dataset, Model, Strategy, etc.).

Classes must implement the abstract method info to output important information that should be saved in results (e.g. names) as a Python dictionary. For example, this mechanisms ensures that strategies uses the key "strategy" and provides a strategy name in the output. An example of JSON output is the following:

{
    "strategy": {
        # info:
        "name": "Cumulative",

        # additional info from specific strategies 
        "model": "IsolationForest",
        "replay_size": 50
    }
}

Moreover, optional information, such as hyperparameters defining a specific configuration of models, strategies, etc can be provided through the additional_info() method.

This information can be passed to a JsonOutputWriter as a list of InfoProvider objects to save information about a pyCLAD execution as a JSON output file.

Code Example: Dataset

class ConceptsDataset(Dataset):
    ...

    def name(self) -> str:
        return self._name

    def additional_info(self):
        return {"train_concepts_no": len(self._train_concepts), "test_concepts_no": len(self._test_concepts)}

Code Example: Model

The Model class implements InfoProvider to log useful information about specific models.

All models have a name and can provide additional information (e.g. number of estimators in IsolationForest) through the additional_info() method:

class IsolationForestAdapter(Model):
    ...

    def name(self) -> str:
        return "IsolationForest"

    def additional_info(self):
        return {"n_estimators": self.n_estimators, "contamination": self.contamination}

Code Example: Strategy

The Strategy class implements InfoProvider to log useful information about specific strategies.

All strategies have a name which is included in the info method:

class Strategy(InfoProvider):
    """Base class for all continual learning strategies."""

    @abc.abstractmethod
    def name(self) -> str: ...

    def additional_info(self) -> Dict:
        return {}

    def info(self) -> Dict[str, Any]:
        return {"strategy": {"name": self.name(), **self.additional_info()}}

Strategies can provide additional information (e.g. replay buffer size) through the additional_info() method:

class ReplayOnlyStrategy(ConceptIncrementalStrategy, ConceptAwareStrategy):
    ...

    def name(self) -> str:
        return "ReplayOnly"

    def additional_info(self) -> Dict:
        return {"replay_size": len(self._buffer.data())}