Core Module Documentation

class neuralprophet.components.trend.base.Trend(config, id_list, quantiles, num_trends_modelled, n_forecasts, device)

abstract add_regularization()

add regularization to loss

abstract property get_trend_deltas

trend deltas for regularization.

update if trend is modelled differently

class neuralprophet.components.trend.linear.GlobalLinearTrend(config, id_list, quantiles, num_trends_modelled, n_forecasts, device)

forward(t, meta)

Computes trend based on model configuration.

Parameters

• t (torch.Tensor float) – normalized time, dim: (batch, n_forecasts)

• meta (dict) –

  Metadata about the all the samples of the model input batch. Contains the following:

  • df_name (list, str), time series ID corresponding to each sample of the input batch.

Returns

Trend component, same dimensions as input t

Return type

torch.Tensor

class neuralprophet.components.trend.linear.LinearTrend(config, id_list, quantiles, num_trends_modelled, n_forecasts, device)

add_regularization()

add regularization to loss

property get_trend_deltas

trend deltas for regularization.

update if trend is modelled differently

class neuralprophet.components.trend.linear.LocalLinearTrend(config, id_list, quantiles, num_trends_modelled, n_forecasts, device)

forward(t, meta)

Computes trend based on model configuration.

Parameters

• t (torch.Tensor float) – normalized time, dim: (batch, n_forecasts)

• meta (dict) –

  Metadata about the all the samples of the model input batch. Contains the following:

  • df_name (list, str), time series ID corresponding to each sample of the input batch.

Returns

Trend component, same dimensions as input t

Return type

torch.Tensor

class neuralprophet.components.trend.piecewise_linear.GlobalPiecewiseLinearTrend(config, id_list, quantiles, num_trends_modelled, n_forecasts, device)

compute_k_t(current_segment, past_next_changepoint, meta_name_tensor_one_hot=None)

This method overrides the method from the PiecewiseLinear class.

compute_m_t(current_segment, past_next_changepoint, meta_name_tensor_one_hot=None)

This method overrides the method from the PiecewiseLinear class.

compute_trend(t, k_t, m_t, meta_name_tensor_one_hot=None)

This method overrides the method from the PiecewiseLinear class.

class neuralprophet.components.trend.piecewise_linear.LocalPiecewiseLinearTrend(config, id_list, quantiles, num_trends_modelled, n_forecasts, device)

compute_k_t(current_segment, past_next_changepoint, meta_name_tensor_one_hot)

This method overrides the method from the PiecewiseLinear class.

compute_m_t(current_segment, past_next_changepoint, meta_name_tensor_one_hot=None)

This method overrides the method from the PiecewiseLinear class.

compute_trend(t, k_t, m_t, meta_name_tensor_one_hot=None)

This method overrides the method from the PiecewiseLinear class.

class neuralprophet.components.trend.piecewise_linear.PiecewiseLinearTrend(config, id_list, quantiles, num_trends_modelled, n_forecasts, device)

add_regularization()

add regularization to loss

compute_k_t(current_segment, past_next_changepoint, meta_name_tensor_one_hot)

For segmentwise, k_t is the model parameter representing the trend slope(actually, trend slope-k_0) in the current_segment at time t (for each sample of the batch).

For not segmentwise, k_t is the model parameter representing the difference between trend slope in the current_segment at time t and the trend slope in the previous segment (for each sample of the batch).

Parameters

• current_segment (torch.Tensor, int) – segment corresponding to time t (batch_size, n_forecasts, segments (+ 1))

• past_next_changepoint (torch.Tensor, bool) – whether the a changepoint >= time t (batch_size, n_forecasts, segments (+ 1))

• meta_name_tensor_one_hot (torch.Tensor, float) –

  Metadata about the all the samples of the model input batch.

  Contains the following:

  • df_name (list, str), time series name ID corresponding to each sample of the input batch.

Returns

k_t, ( batch_size, n_forecasts, quantiles_size)

Return type

torch.Tensor

compute_m_t(current_segment, past_next_changepoint, meta_name_tensor_one_hot)

m_t represents the value at the origin(t=0) that we would need to have so that if we use (k_t + k_0) as slope, we reach the same value at time = chagepoint_start_of_segment_i as we would reach by following the segmented slope (having in each segment the slope trend_deltas(i) + k_0)

Parameters

• current_segment (torch.Tensor, int) – segment corresponding to time t (batch_size, n_forecasts, segments (+ 1))

• past_next_changepoint (torch.Tensor, bool) – whether the a changepoint >= time t (batch_size, n_forecasts, segments (+ 1))

• meta_name_tensor_one_hot (torch.Tensor, float) –

  Metadata about the all the samples of the model input batch.

  Contains the following:

  • df_name (list, str), time series name ID corresponding to each sample of the input batch.

Returns

m_t, ( batch_size, n_forecasts, quantiles_size)

Return type

torch.Tensor

compute_trend(t, k_t, m_t, meta_name_tensor_one_hot=None)

This method computes the trend component of the model.

Parameters

• t (torch.Tensor, float) – time

• k_t (torch.Tensor, int) – see compute_k_t

• m_t (torch.Tensor, bool) – see compute_m_t

• meta_name_tensor_one_hot (torch.Tensor, float) –

  Metadata about the all the samples of the model input batch.

  Contains the following:

  • df_name (list, str), time series name ID corresponding to each sample of the input batch.

Returns

trend_component, ( batch_size, n_forecasts, quantiles_size)

Return type

torch.Tensor

forward(t, meta)

Computes trend based on model configuration.

Parameters

• t (torch.Tensor float) – normalized time, dim: (batch, n_forecasts)

• meta (dict) –

  Metadata about the all the samples of the model input batch. Contains the following:

  • df_name (list, str), time series ID corresponding to each sample of the input batch.

Returns

Trend component, same dimensions as input t

Return type

torch.Tensor

property get_trend_deltas

trend deltas for regularization.

update if trend is modelled differently

class neuralprophet.components.trend.static.StaticTrend(config, id_list, quantiles, num_trends_modelled, n_forecasts, device)

add_regularization()

add regularization to loss

forward(t, meta)

Computes trend based on model configuration.

Parameters

• t (torch.Tensor float) – normalized time, dim: (batch, n_forecasts)

• meta (dict) –

  Metadata about the all the samples of the model input batch. Contains the following:

  • df_name (list, str), time series ID corresponding to each sample of the input batch.

Returns

Trend component, same dimensions as input t

Return type

torch.Tensor

property get_trend_deltas

trend deltas for regularization.

update if trend is modelled differently