gluonts.torch.model.tft.module module

class gluonts.torch.model.tft.module.TemporalFusionTransformerModel(context_length: int, prediction_length: int, d_feat_static_real: Optional[List[int]] = None, c_feat_static_cat: Optional[List[int]] = None, d_feat_dynamic_real: Optional[List[int]] = None, c_feat_dynamic_cat: Optional[List[int]] = None, d_past_feat_dynamic_real: Optional[List[int]] = None, c_past_feat_dynamic_cat: Optional[List[int]] = None, num_heads: int = 4, d_hidden: int = 32, d_var: int = 32, dropout_rate: float = 0.1, distr_output: Optional[gluonts.torch.distributions.output.Output] = None)

Bases: torch.nn.modules.module.Module

Temporal Fusion Transformer neural network.

Partially based on the implementation in github.com/kashif/pytorch-transformer-ts.

Inputs feat_static_real, feat_static_cat and feat_dynamic_real are mandatory. Inputs feat_dynamic_cat, past_feat_dynamic_real and past_feat_dynamic_cat are optional.

describe_inputs(batch_size=1) → gluonts.model.inputs.InputSpec

feat_dynamic_embed: Optional[FeatureEmbedder]

feat_dynamic_proj: Optional[FeatureProjector]

feat_static_embed: Optional[FeatureEmbedder]

feat_static_proj: Optional[FeatureProjector]

forward(past_target: torch.Tensor, past_observed_values: torch.Tensor, feat_static_real: Optional[torch.Tensor], feat_static_cat: Optional[torch.Tensor], feat_dynamic_real: Optional[torch.Tensor], feat_dynamic_cat: Optional[torch.Tensor] = None, past_feat_dynamic_real: Optional[torch.Tensor] = None, past_feat_dynamic_cat: Optional[torch.Tensor] = None) → Tuple[Tuple[torch.Tensor, ...], torch.Tensor, torch.Tensor]

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.

input_types() → Dict[str, torch.dtype]

loss(past_target: torch.Tensor, past_observed_values: torch.Tensor, future_target: torch.Tensor, future_observed_values: torch.Tensor, feat_static_real: torch.Tensor, feat_static_cat: torch.Tensor, feat_dynamic_real: torch.Tensor, feat_dynamic_cat: Optional[torch.Tensor] = None, past_feat_dynamic_real: Optional[torch.Tensor] = None, past_feat_dynamic_cat: Optional[torch.Tensor] = None) → torch.Tensor

past_feat_dynamic_embed: Optional[FeatureEmbedder]

past_feat_dynamic_proj: Optional[FeatureProjector]

training: bool