qumphy.models.timesnet module

File: qumphy/models/timesnet.py Project: 22HLT01 QUMPHY Contact: oskar.pfeffer@ptb.de Gitlab: https://gitlab.com/qumphy Description: Adapted from https://github.com/thuml/Time-Series-Library.

qumphy.models.timesnet.FFT_for_Period(x, k=2)[source]

Find dominant periods in a time-series batch using the FFT.

Parameters:

  • x (torch.Tensor) – Input tensor of shape (batch_size, sequence_length, channels).

  • k (int) – Number of dominant frequencies to select.

Returns:

Tuple containing the selected periods and their corresponding frequency weights.

Return type:

tuple

class qumphy.models.timesnet.TimesBlock(*args: Any, **kwargs: Any)[source]

Bases: Module

TimesNet block for multi-period temporal feature extraction.

forward(x)[source]
class qumphy.models.timesnet.TimesNet(*args: Any, **kwargs: Any)[source]

Bases: Module

TimesNet model for time series forecasting and classification.

Implements the TimesNet architecture as described in: https://openreview.net/pdf?id=ju_Uqw384Oq

Parameters:

  • seq_len (int) – Length of the input sequence.

  • label_len (int) – Length of the label/start token sequence (for forecasting).

  • pred_len (int, optional) – Length of the prediction/output sequence.

  • e_layers (int, optional) – Number of encoder layers. Default is 2.

  • d_model (int, optional) – Model hidden dimension. Default is 16.

  • d_ff (int, optional) – Dimension of the feed-forward network. Default is 32.

  • num_kernels (int, optional) – Number of kernels for the Inception-like blocks. Default is 6.

  • top_k (int, optional) – Top-k selection parameter for TimesBlock. Default is 5.

  • enc_in (int, optional) – Number of input channels/features. Default is 1.

  • c_out (int, optional) – Number of output channels/features or classes. Default is 1.

  • embed (str, optional) – Type of time feature embedding. Options are ‘timeF’, ‘fixed’, or ‘learned’. Default is ‘fixed’.

  • freq (str, optional) – Frequency string for time feature encoding (e.g., ‘h’ for hourly). Default is ‘s’.

  • dropout (float, optional) – Dropout rate. Default is 0.1.

  • num_class (int, optional) – Number of classes (for classification tasks).

  • task_name (str, optional) – Task type, either ‘classification’ or ‘forecasting’. Default is ‘classification’.

References

anomaly_detection(x_enc)[source]

Run anomaly detection with the TimesNet model.

Parameters:

x_enc (torch.Tensor) – Encoder input tensor of shape (batch_size, seq_len, enc_in).

Returns:

Reconstructed output tensor.

Return type:

torch.Tensor

classification(x_enc, x_mark_enc)[source]

Run classification with the TimesNet model.

Parameters:

  • x_enc (torch.Tensor) – Input tensor of shape (batch_size, enc_in, seq_len).

  • x_mark_enc (torch.Tensor) – Encoder time feature tensor. This parameter is kept for compatibility.

Returns:

Classification logits of shape (batch_size, num_class).

Return type:

torch.Tensor

forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)[source]

Run forecasting with the TimesNet model.

Parameters:

  • x_enc (torch.Tensor) – Encoder input tensor of shape (batch_size, seq_len, enc_in).

  • x_mark_enc (torch.Tensor) – Encoder time feature tensor.

  • x_dec (torch.Tensor) – Decoder input tensor. This parameter is kept for compatibility.

  • x_mark_dec (torch.Tensor) – Decoder time feature tensor. This parameter is kept for compatibility.

Returns:

Forecast output tensor of shape (batch_size, seq_len + pred_len, c_out).

Return type:

torch.Tensor

forward(x_enc, x_mark_enc=None, x_dec=None, x_mark_dec=None, mask=None)[source]
imputation(x_enc, x_mark_enc, x_dec, x_mark_dec, mask)[source]

Run imputation with the TimesNet model.

Parameters:

  • x_enc (torch.Tensor) – Encoder input tensor of shape (batch_size, seq_len, enc_in).

  • x_mark_enc (torch.Tensor) – Encoder time feature tensor.

  • x_dec (torch.Tensor) – Decoder input tensor. This parameter is kept for compatibility.

  • x_mark_dec (torch.Tensor) – Decoder time feature tensor. This parameter is kept for compatibility.

  • mask (torch.Tensor) – Mask tensor indicating observed and missing values.

Returns:

Imputed output tensor.

Return type:

torch.Tensor