Basic Transformer

class flood_forecast.transformer_xl.transformer_basic.SimpleTransformer(number_time_series: int, seq_length: int = 48, output_seq_len: int = None, d_model: int = 128, n_heads: int = 8, dropout=0.1, forward_dim=2048, sigmoid=False)

__init__(number_time_series: int, seq_length: int = 48, output_seq_len: int = None, d_model: int = 128, n_heads: int = 8, dropout=0.1, forward_dim=2048, sigmoid=False)

A full transformer model.

Parameters:

• number_time_series (int) – The total number of time series present (e.g. n_feature_time_series + n_targets)

• seq_length (int, optional) – The length of your input sequence, defaults to 48

• output_seq_len (int, optional) – The length of your output sequence, defaults to None

• d_model (int, optional) – The dimensions of your model, defaults to 128

• n_heads (int, optional) – The number of heads in each encoder/decoder block, defaults to 8

• dropout (float, optional) – The fraction of dropout you wish to apply during training, defaults to 0.1 (currently not functional)

• forward_dim (int, optional) – Currently not functional, defaults to 2048

• sigmoid (bool, optional) – Whether to apply a sigmoid activation to the final layer (useful for binary classification), defaults to False

forward(x: Tensor, t: Tensor, tgt_mask=None, src_mask=None)

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.

basic_feature(x: Tensor)

encode_sequence(x, src_mask=None)

decode_seq(mem, t, tgt_mask=None, view_number=None) → Tensor

class flood_forecast.transformer_xl.transformer_basic.CustomTransformerDecoder(seq_length: int, output_seq_length: int, n_time_series: int, d_model=128, output_dim=1, n_layers_encoder=6, forward_dim=2048, dropout=0.1, use_mask=False, meta_data=None, final_act=None, squashed_embedding=False, n_heads=8)

__init__(seq_length: int, output_seq_length: int, n_time_series: int, d_model=128, output_dim=1, n_layers_encoder=6, forward_dim=2048, dropout=0.1, use_mask=False, meta_data=None, final_act=None, squashed_embedding=False, n_heads=8)

Uses a number of encoder layers with simple linear decoder layer.

Parameters:

• seq_length (int) – The number of historical time-steps fed into the model in each forward pass.

• output_seq_length (int) – The number of forecasted time-steps outputted by the model.

• n_time_series (int) – The total number of time series present (targets + features)

• d_model (int, optional) – The embedding dim of the mode, defaults to 128

• output_dim (int, optional) – The output dimension (should correspond to n_targets), defaults to 1

• n_layers_encoder (int, optional) – The number of encoder layers, defaults to 6

• forward_dim (int, optional) – The forward embedding dim, defaults to 2048

• dropout (float, optional) – How much dropout to use, defaults to 0.1

• use_mask (bool, optional) – Whether to use subsquent sequence mask during training, defaults to False

• meta_data (str, optional) – Whether to use static meta-data, defaults to None

• final_act (str, optional) – Whether to use a final activation function, defaults to None

• squashed_embedding (bool, optional) – Whether to create a one 1-D time embedding, defaults to False

• n_heads (int, optional) – [description], defaults to 8

make_embedding(x: Tensor)

forward(x: Tensor, meta_data=None) → Tensor

Performs forward pass on tensor of (batch_size, sequence_length, n_time_series) Return tensor of dim (batch_size, output_seq_length)

class flood_forecast.transformer_xl.transformer_basic.SimplePositionalEncoding(d_model, dropout=0.1, max_len=5000)

__init__(d_model, dropout=0.1, max_len=5000)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor) → Tensor

Creates a basic positional encoding

flood_forecast.transformer_xl.transformer_basic.greedy_decode(model, src: Tensor, max_len: int, real_target: Tensor, unsqueeze_dim=1, output_len=1, device='cpu', multi_targets=1, probabilistic=False, scaler=None)

Mechanism to sequentially decode the model :src The Historical time series values :real_target The real values (they should be masked), however if you want can include known real values. :returns torch.Tensor