gluonts.torch.model.wavenet.module module#

class gluonts.torch.model.wavenet.module.CausalDilatedResidualLayer(num_residual_channels: int, num_skip_channels: int, dilation: int, kernel_size: int, return_dense_output: bool)[source]#

Bases: Module

forward(x: Tensor) → Tuple[Tensor, Tensor][source]#

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class gluonts.torch.model.wavenet.module.WaveNet(pred_length: int, bin_values: List[float], num_residual_channels: int, num_skip_channels: int, dilation_depth: int, num_stacks: int, num_feat_dynamic_real: int = 1, num_feat_static_real: int = 1, cardinality: List[int] = [1], embedding_dimension: int = 5, num_parallel_samples: int = 100, temperature: float = 1.0, use_log_scale_feature: bool = True)[source]#

Bases: Module

The WaveNet model.

Parameters:

  • pred_length – Prediction length.

  • bin_values – List of bin values.

  • num_residual_channels – Number of residual channels.

  • num_skip_channels – Number of skip channels.

  • dilation_depth – The depth of the dilated convolution.

  • num_stacks – The number of dilation stacks.

  • num_feat_dynamic_real – The number of dynamic real features, by default 1

  • optional – The number of dynamic real features, by default 1

  • num_feat_static_real – The number of static real features, by default 1

  • optional – The number of static real features, by default 1

  • cardinality – The cardinalities of static categorical features, by default [1]

  • optional – The cardinalities of static categorical features, by default [1]

  • embedding_dimension – The dimension of the embeddings for categorical features, by default 5

  • optional – The dimension of the embeddings for categorical features, by default 5

  • num_parallel_samples – The number of parallel samples to generate during inference. This parameter is only used in inference mode, by default 100

  • optional – The number of parallel samples to generate during inference. This parameter is only used in inference mode, by default 100

  • temperature – Temparature used for sampling from the output softmax distribution, by default 1.0

  • optional – Temparature used for sampling from the output softmax distribution, by default 1.0

base_net(inputs: Tensor, queues: Optional[List[Tensor]] = None) → Tuple[Tensor, List[Tensor]][source]#

Forward pass through the WaveNet.

Parameters:

  • inputs – A tensor of inputs Shape: (batch_size, num_residual_channels, sequence_length)

  • queues – Convolutional queues containing past computations. This speeds up predictions and must be provided during prediction mode. See [Paine et al., 2016] for details, by default None

  • optional – Convolutional queues containing past computations. This speeds up predictions and must be provided during prediction mode. See [Paine et al., 2016] for details, by default None

  • al. ([Paine et) – arXiv preprint arXiv:1611.09482 (2016).

  • algorithm." (2016] "Fast wavenet generation) – arXiv preprint arXiv:1611.09482 (2016).

Returns:

  • A tensor containing the unnormalized outputs of the network of

  • shape (batch_size, pred_length, num_bins) and a list containing the

  • convolutional queues for each layer. The queue corresponding to

  • layer `l` has shape ((batch_size, num_residual_channels, 2^l).)

forward(feat_static_cat: Tensor, feat_static_real: Tensor, past_target: Tensor, past_observed_values: Tensor, past_time_feat: Tensor, future_time_feat: Tensor, scale: Tensor, prediction_length: Optional[int] = None, num_parallel_samples: Optional[int] = None, temperature: Optional[float] = None) → Tensor[source]#

Generate predictions from the WaveNet model.

Parameters:

  • feat_static_cat – Static categorical features: (batch_size, num_cat_features)

  • feat_static_real – Static real-valued features: (batch_size, num_feat_static_real)

  • past_target – Past target: (batch_size, receptive_field)

  • past_observed_values – Observed value indicator for the past target: (batch_size, receptive_field)

  • past_time_feat – Past time features: (batch_size, num_time_features, receptive_field)

  • future_time_feat – Future time features: (batch_size, num_time_features, pred_length)

  • scale – Scale of the time series: (batch_size, 1)

  • prediction_length – Time length of the samples to generate. If not provided, use self.prediction_length.

  • num_parallel_samples – Number of samples to generate. If not provided, use self.num_parallel_samples.

  • temperature – Temperature to use in generating samples. If not provided, use self.temperature.

Return type:

Predictions with shape (batch_size, num_parallel_samples, pred_length)

get_full_features(feat_static_cat: Tensor, feat_static_real: Tensor, past_observed_values: Tensor, past_time_feat: Tensor, future_time_feat: Tensor, future_observed_values: Optional[Tensor], scale: Tensor) → Tensor[source]#

Prepares the inputs for the network by repeating static feature and concatenating it with time features and observed value indicator.

Parameters:

  • feat_static_cat – Static categorical features: (batch_size, num_cat_features)

  • feat_static_real – Static real-valued features: (batch_size, num_feat_static_real)

  • past_observed_values – Observed value indicator for the past target: (batch_size, receptive_field)

  • past_time_feat – Past time features: (batch_size, num_time_features, receptive_field)

  • future_time_feat – Future time features: (batch_size, num_time_features, pred_length)

  • future_observed_values – Observed value indicator for the future target: (batch_size, pred_length). This will be set to all ones, if not provided (e.g., during inference)

  • scale – scale of the time series: (batch_size, 1)

Returns:

  • A tensor containing all the features ready to be passed through the

  • network.

  • Shape ((batch_size, num_features, receptive_field + pred_length))

static get_receptive_field(dilation_depth: int, num_stacks: int) → int[source]#
loss(feat_static_cat: Tensor, feat_static_real: Tensor, past_target: Tensor, past_observed_values: Tensor, past_time_feat: Tensor, future_time_feat: Tensor, future_target: Tensor, future_observed_values: Tensor, scale: Tensor) → Tensor[source]#

Computes the training loss for the wavenet model.

Parameters:

  • feat_static_cat – Static categorical features: (batch_size, num_cat_features)

  • feat_static_real – Static real-valued features: (batch_size, num_feat_static_real)

  • past_target – Past target: (batch_size, receptive_field)

  • past_observed_values – Observed value indicator for the past target: (batch_size, receptive_field)

  • past_time_feat – Past time features: (batch_size, num_time_features, receptive_field)

  • future_time_feat – Future time features: (batch_size, num_time_features, pred_length)

  • future_target – Target on which the loss is computed: (batch_size, pred_length)

  • future_observed_values – Observed value indicator for the future target: (batch_size, pred_length). This will be set to all ones, if not provided (e.g., during inference)

  • scale – Scale of the time series: (batch_size, 1)

Return type:

Loss tensor with shape (batch_size, pred_length)

target_feature_embedding(target: Tensor, features: Tensor) → Tensor[source]#

Provides a joint embedding for the target and features.

Parameters:

  • target – Full target of shape (batch_size, sequence_length)

  • features – Full features of shape (batch_size, num_features, sequence_length)

Returns:

  • A tensor containing a joint embedding of target and features.

  • Shape ((batch_size, n_residue, sequence_length))