RNN¶

class pytorch_forecasting.models.nn.rnn.RNN(mode: str, input_size: int, hidden_size: int, num_layers: int = 1, bias: bool = True, batch_first: bool = False, dropout: float = 0.0, bidirectional: bool = False, proj_size: int = 0, device=None, dtype=None)[source]¶

Bases: abc.ABC, torch.nn.modules.rnn.RNNBase

Base class flexible RNNs.

Forward function can handle sequences of length 0.

Methods

`forward`(x[, hx, lengths, enforce_sorted])	Forward function of rnn that allows zero-length sequences.
`handle_no_encoding`(hidden_state, ...)	Mask the hidden_state where there is no encoding.
`init_hidden_state`(x)	Initialise a hidden_state.
`repeat_interleave`(hidden_state, n_samples)	Duplicate the hidden_state n_samples times.

forward(x: Union[torch.nn.utils.rnn.PackedSequence, torch.Tensor], hx: Optional[Union[Tuple[torch.Tensor, torch.Tensor], torch.Tensor]] = None, lengths: Optional[torch.LongTensor] = None, enforce_sorted: bool = True) → Tuple[Union[torch.nn.utils.rnn.PackedSequence, torch.Tensor], Union[Tuple[torch.Tensor, torch.Tensor], torch.Tensor]][source]¶

Forward function of rnn that allows zero-length sequences.

Functions as normal for RNN. Only changes output if lengths are defined.

Parameters

x (Union[rnn.PackedSequence, torch.Tensor]) – input to RNN. either packed sequence or tensor of padded sequences
hx (HiddenState, optional) – hidden state. Defaults to None.
lengths (torch.LongTensor, optional) – lengths of sequences. If not None, used to determine correct returned hidden state. Can contain zeros. Defaults to None.
enforce_sorted (bool, optional) – if lengths are passed, determines if RNN expects them to be sorted. Defaults to True.

Returns

output and hidden state.: Output is packed sequence if input has been a packed sequence.

Return type

Tuple[Union[rnn.PackedSequence, torch.Tensor], HiddenState]

abstract handle_no_encoding(hidden_state: Union[Tuple[torch.Tensor, torch.Tensor], torch.Tensor], no_encoding: torch.BoolTensor, initial_hidden_state: Union[Tuple[torch.Tensor, torch.Tensor], torch.Tensor]) → Union[Tuple[torch.Tensor, torch.Tensor], torch.Tensor][source]¶

Mask the hidden_state where there is no encoding.

Parameters

hidden_state (HiddenState) – hidden state where some entries need replacement
no_encoding (torch.BoolTensor) – positions that need replacement
initial_hidden_state (HiddenState) – hidden state to use for replacement

Returns

hidden state with propagated initial hidden state where appropriate

Return type

HiddenState

abstract init_hidden_state(x: torch.Tensor) → Union[Tuple[torch.Tensor, torch.Tensor], torch.Tensor][source]¶

Initialise a hidden_state.

Parameters: x (torch.Tensor) – network input
Returns: default (zero-like) hidden state
Return type: HiddenState

abstract repeat_interleave(hidden_state: Union[Tuple[torch.Tensor, torch.Tensor], torch.Tensor], n_samples: int) → Union[Tuple[torch.Tensor, torch.Tensor], torch.Tensor][source]¶

Duplicate the hidden_state n_samples times.

Parameters

hidden_state (HiddenState) – hidden state to repeat
n_samples (int) – number of repetitions

Returns

repeated hidden state

Return type

HiddenState

rnn