MODELS.md

Machine Learning Model

This section is for developers and data scientists who want to create new model types for time series and other data changing over time.

This interface allows for users to pick and chose what is best for their use cases and makes it easy for developers to create new ways of deriving value from their data.

Model authorship follows the following guidelines to deliver consistent and high quality implementations.

Model Guidelines

• A new model must conform to the Model interface
• A new model class should call super.__init__() in their __init__ function to register themselves and inherit from the Model class
• A new model class must be defined in a new file
• The model must be able to serialize and deserialize the training state in JSON format
• Methods that are not implemented must raise raise NotImplemented() an exception
• The main Description header must reference the ArXiv publication that motivates the new model implementation
• Unit tests include both training and inference
• Follow the recommended CodeStyle

Let's say you've written a new model that fits a gaussian to input data.

Gaussian Model Example

"""My new Gaussian model

# Reference:
- [ArXiv article](
    https://arxiv.org/xxxxx)
"""

# gaussian.py

class GaussianModel(Model):
    """
    Time-series Gaussian model
    """
    TYPE = 'gaussian'

    SCHEMA = Model.SCHEMA.extend({
        Optional('new_param'): Any(None, "auto", All(int, Range(min=1))),
    })

    def __init__(self, settings, state=None):
        super().__init__(settings, state)

        settings = self.validate(settings)

    @property
    def type(self):
        return self.TYPE

    def train(
        self,
        bucket,
        from_date,
        to_date="now",
        train_size=0.67,
        batch_size=256,
        num_epochs=100,
        num_cpus=1,
        num_gpus=0,
        max_evals=None,
        progress_cb=None,
        incremental=False,
        windows=[],
    ):
        """
        Train model
        """
        return {
            "loss": loss_value,
        }

Development

• Add new Python code to the loudml/loudml directory
• Add new Python file test_model.py to the loudml/tests directory
• Run make rpm to package the new version

Data Formats

Your model can query data using a Bucket interface and a time range. Returned data can be stored into a NumPy array or other structures you need in your model implementation.