Models

class models.WaveNet(input_channels, output_channels, horizon, hidden_channels=64, skip_channels=64, dense_units=128, n_layers=7, n_blocks=1, dilation=2)

Implements WaveNet architecture for time series forecasting. Inherits from pytorch Module. Vector forecasts are made via a fully-connected layer.

References:
Arguments:
  • input_channels (int): Number of covariates in input time series.
  • output_channels (int): Number of target time series.
  • horizon (int): Number of time steps to forecast.
  • hidden_channels (int): Number of channels in convolutional hidden layers.
  • skip_channels (int): Number of channels in convolutional layers for skip connections.
  • dense_units (int): Number of hidden units in final dense layer.
  • n_layers (int): Number of layers per Wavenet block (determines receptive field size).
  • n_blocks (int): Number of Wavenet blocks.
  • dilation (int): Dilation factor for temporal convolution.

Inititalize variables.

decode(inputs: <sphinx.ext.autodoc.importer._MockObject object at 0x7f7e27095da0>)

Returns forecasts based on embedding vectors.

Arguments:
  • inputs: embedding vectors to generate forecasts for
encode(inputs: <sphinx.ext.autodoc.importer._MockObject object at 0x7f7e26fcd908>)

Returns embedding vectors.

Arguments:
  • inputs: time series input to make forecasts for
forward(inputs)

Forward function.

n_parameters

Returns the number of model parameters.

receptive_field_size

Returns the length of the receptive field.