Soothsayer.Trainer (Soothsayer v0.6.1)

View Source

Training functionality for Soothsayer models.

Handles standard training and custom training with L1 regularization on specified layer weights (AR, trend, etc.).

Summary

Functions

compute_l1_penalty(params, layer_names)

Computes L1 penalty for specified layer kernels.

fit(network, x, y, epochs, config)

Trains a network on the provided data.

Functions

compute_l1_penalty(params, layer_names)

@spec compute_l1_penalty(Axon.ModelState.t(any(), any()), [String.t()]) ::
  Nx.Tensor.t()

Computes L1 penalty for specified layer kernels.

Parameters

  • params - An Axon.ModelState containing the model parameters.
  • layer_names - A list of layer names to include in the penalty.

Returns

A scalar tensor with the sum of absolute values of all kernel weights in the specified layers.

Examples

iex> penalty = Soothsayer.Trainer.compute_l1_penalty(params, ["ar_dense_out"])
#Nx.Tensor<f32 6.0>

fit(network, x, y, epochs, config)

@spec fit(
  Axon.t(),
  %{required(String.t()) => Nx.Tensor.t()},
  Nx.Tensor.t(),
  non_neg_integer(),
  map()
) ::
  Axon.ModelState.t(any(), any())

Trains a network on the provided data.

Parameters

  • network - An Axon neural network.
  • x - A map of input tensors.
  • y - A tensor of target values.
  • epochs - The number of training epochs.
  • config - A map containing training configuration:
    • :learning_rate - Learning rate for the optimizer.
    • :ar - Optional map with :regularization for AR L1 penalty.
    • :trend - Optional map with :regularization for trend L1 penalty.

Returns

The trained Axon.ModelState.

Examples

iex> config = %{learning_rate: 0.1}
iex> params = Soothsayer.Trainer.fit(network, x, y, 100, config)
%Axon.ModelState{...}