MLP layer.

BasicTS Forecasting Config, including general configuration, dataset and scaler configuration, model configuration, \ metrics configuration,

RevIN Paper: Reversible Instance Normalization for Accurate Time-Series Forecasting Against Distribution Shift Official Code: https://git

BasicTS Multi-Head Attention module. Features: - Can be used as self-/cross-attention. - MHA/MQA/GQA with various Key-Va

FeatureEmbedding layer is used to embed the time series from the feature dimension to hidden dimension, \ i.e., [batch_size, seq_len, num_fea

BasicTS Transformer encoder.

Gated Linear Unit

This is the MLP-based Residual Block

Config class for iTransformer model.

BasicTS Transformer block.

Time series decomposition layer by moving average.

PatchEmbedding layer is used to embed the time series from the patch dimension to hidden dimension, \ i.e., [batch_size, seq_len, num_feature

Paper: A Time Series is Worth 64 Words: Long-term Forecasting with Transformers Link: https://arxiv.org/abs/2211.14730 Official Code: htt

Config class for DLinear model.

Multilayer perceptron to encode/decode high dimension representation of sequential data

SequenceEmbedding layer is used to embed the time series from the temporal dimension to hidden dimension, \ i.e., [batch_size, seq_len, num_f

Auto correlation layer from Autoformer.

BasicTS Imputation Config, including general configuration, dataset and scaler configuration, model configuration, \ metrics configuration, t

De-stationary Attention Layer.

IEBlock in LightTS.

BasicTS KV cache for generative Transformers. This cache is used to store the Key and Value states of the attention layers. It is a simple ve

Paper: Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting Official Code: https://github.com/thuml/Nonstationa

Head layer for PatchTST.

PositionEmbedding layer is used to embed the position of the time series to hidden dimension, \ i.e., [batch_size, seq_len, hidden_size] -> [

Seq2Seq decoder for encoder-decoder architecture.

Paper: iTransformer: Inverted Transformers Are Effective for Time Series Forecasting Official Code: https://github.com/thuml/iTransformer

BasicTS Classification Config, including general configuration, dataset and scaler configuration, model configuration, \ metrics configuratio

Forecasting Task Flow

A base runner that uses epoch as the fundamental training unit. This is a general runner. Other Features: - support torch.compil

Centered LayerNorm with zero mean, special designed for the seasonal part of time series. Original implementation is from Autoformer.

Crossformer Encoder Layer

Distributional Router Encoder

frequency-domain MLPs Attributes: real (nn.Parameter): the real part of weights imag (nn.Parameter): the imaginary part of w

Inception block v1 in TimesNet.

Leddam Transformer block.

Moving average block to highlight the trend of time series. Args: kernel_size (int): kernel size of moving average. stri

Config class for PatchTST model.

Probabilistic Sparse Attention layer in Informer. Modified to follow BasicTS style with clear type hints and structure.

MLP to learn the De-stationary factors Paper link: https://openreview.net/pdf?id=ucNDIDRNjjv

M_KAN layer.

Inference engine for EasyTorch.

Adding auxiliary loss callback.

Autoregressive Attention module for Leddam.

Auto-regressive decoder for decoder-only architecture.

Autoformer decoder with the progressive decomposition architecture

Autoformer decoder block with the progressive decomposition architecture

Autoformer encoder block with the progressive decomposition architecture

Handler for BasicTS callbacks.

Classification Task Flow

Imputation Task Flow

Channel Projection Layer

Kolmogorov-Arnold Network layer using Chebyshev polynomials instead of splines coefficients.

Convolutional layer for the Informer model.

Config class for Crossformer model.

The decoder of Crossformer, making the final prediction by adding up predictions at each scale

Crossformer Decoder

Crossformer Encoder

Time series decomposition layer by discrete Fourier transform.

Paper: Are Transformers Effective for Time Series Forecasting? Link: https://arxiv.org/abs/2205.13504 Official Code: https://

Config class for DUET model.

BasicTS Transformer decoder layer for decoder-only architecture. It is a variant of the encoder layer, with improvements for autoregressive d

Early stopping callback. Args: patience (int, optional): Number of epochs with no improvement after which training will be stopped.

Config class for FITS model.

Factorized Channel Mixing Layer

Factorized Temporal Mixing Layer

Flatten head for TimeXer.

Fourier Filter: to time-variant and time-invariant term

Config class for FreTS model.

Frequency decomposition layer that separates high and low frequency components.

Frequency mixing layer that combines low and high frequency components.

Clip gradient norm. Args: max_norm (float): max norm of the gradients norm_type (float): type of the used p-norm. Can be ``'

Hippo projection layer.

Config class for Informer model.

Informer encoder with additional convolutional layers.

A demonstration of finding one step transition of linear system by DMD iteratively

Find koopman transition of linear system by DMD with multistep K approximation

Config class for Koopa model.

Callback for initializing Koopa mask during training. Changes made: - Robust handling when training loader is empty. - Ensure

Learnable Decomposition module for Leddam.

Config class for Leddam model.

Config class for LightTS model.

Call a Sparsely gated mixture of experts layer with 1-layer Feed-Forward networks as experts. Args: input_size: integer - size of the input

Config class for MTSMixer model.

Mahalanobis mask module. Args: input_len (int): Input sequence length.

Meter container

Mixer Layer of MTSMixer

Bottom-up mixing season pattern

Top-down mixing trend pattern

Config class for NLinear model.

Decoder layer for non-stationary transformer. Extra arguments `tau` and `delta` are passed for de-stationary attention.

Encoder layer for non-stationary transformer. Extra arguments `tau` and `delta` are passed for de-stationary attention.

Past decomposable mixing layer.

Patch Merging Layer. The adjacent `win_size` segments in each dimension will be merged into one patch \ to get representation of a coarse

MLP layer with residual connection.

STar Aggregate-Redistribute Module.

Config class for STID model.

Config class for SegRNN model.

Selective learning callback. Paper: Selective Learning for Deep Time Series Forecasting Venue: NeurIPS 2025 Task: Long-term Time Seri

BasicTS Transformer decoder layer for encoder-decoder architecture.