Nasty.Statistics.Model behaviour (Nasty v0.3.0)

View Source

Behaviour for statistical models in Nasty.

All statistical models (HMM, PCFG, CRF, etc.) implement this behaviour, providing a consistent interface for training, prediction, and persistence.

Model Lifecycle

# Training
model = MyModel.new(opts)
model = MyModel.train(model, training_data)
:ok = MyModel.save(model, "path/to/model.bin")

# Loading and prediction
{:ok, model} = MyModel.load("path/to/model.bin")
predictions = MyModel.predict(model, input_data)

Callbacks

  • train/2 - Train the model on annotated data
  • predict/2 - Make predictions on new data
  • save/2 - Serialize model to disk
  • load/1 - Deserialize model from disk
  • metadata/1 - Get model metadata (version, accuracy, etc.)

Summary

Types

input_data()
metadata()
model()
options()
predictions()
training_data()

Callbacks

load(t)

Load a serialized model from disk.

metadata(model)

Get model metadata (version, training info, etc.).

predict(model, input_data, options)

Make predictions on new input data.

save(model, t)

Serialize and save the model to disk.

train(model, training_data, options)

Train the model on annotated training data.

Functions

deserialize(binary)

Helper function to deserialize a model from binary format.

serialize(model, metadata)

Helper function to serialize a model to binary format.

Types

input_data()

@type input_data() :: term()

metadata()

@type metadata() :: %{
  :version => String.t(),
  :trained_at => DateTime.t(),
  optional(:accuracy) => float(),
  optional(:training_size) => pos_integer(),
  optional(atom()) => term()
}

model()

@type model() :: struct()

options()

@type options() :: keyword()

predictions()

@type predictions() :: term()

training_data()

@type training_data() :: list()

Callbacks

load(t)

@callback load(Path.t()) :: {:ok, model()} | {:error, term()}

Load a serialized model from disk.

Parameters

  • path - File path to load from

Returns

  • {:ok, model} - Successfully loaded model
  • {:error, reason} - Load failed

metadata(model)

@callback metadata(model()) :: metadata()

Get model metadata (version, training info, etc.).

Parameters

  • model - The model

Returns

  • Metadata map with version, accuracy, training time, etc.

predict(model, input_data, options)

@callback predict(model(), input_data(), options()) ::
  {:ok, predictions()} | {:error, term()}

Make predictions on new input data.

Parameters

  • model - Trained model
  • input_data - Data to predict on
  • opts - Prediction options

Returns

  • {:ok, predictions} - Predicted labels/structures
  • {:error, reason} - Prediction failed

save(model, t)

@callback save(model(), Path.t()) :: :ok | {:error, term()}

Serialize and save the model to disk.

Parameters

  • model - Model to save
  • path - File path for saving

Returns

  • :ok - Successfully saved
  • {:error, reason} - Save failed

train(model, training_data, options)

@callback train(model(), training_data(), options()) :: {:ok, model()} | {:error, term()}

Train the model on annotated training data.

Parameters

  • model - The model struct to train
  • training_data - Annotated training examples
  • opts - Training options (learning rate, iterations, etc.)

Returns

  • {:ok, trained_model} - Successfully trained model
  • {:error, reason} - Training failed

Functions

deserialize(binary)

@spec deserialize(binary()) :: {:ok, model(), metadata()} | {:error, term()}

Helper function to deserialize a model from binary format.

Validates version compatibility and extracts model data.

serialize(model, metadata)

@spec serialize(model(), metadata()) :: binary()

Helper function to serialize a model to binary format.

Uses Erlang's term_to_binary for efficient serialization. Includes versioning and compression.