Apply rotary positional embedding to input tensor T. check https://kexue.fm/archives/8265 for detailed formulas Args: t (Tensor): Inpu

Calculate Mean Absolute Error between predictions and ground truth. MAE measures the average magnitude of errors in predictions, without

Calculate Mean Squared Error between predictions and ground truth. MSE measures the average of the squares of the errors. It gives highe

auto regressive prediction with [single] head

Perform a forward pass through the output layer. Args: hidden_states (Tensor): Transformed hidden states of s

Perform patch embedding on the input time series. This method applies a linear transformation to the input tensor to convert

Change sign so the last dimension becomes [-odd, +even] Args: x (Tensor): Input tensor Returns: Tensor: Tensor rotated half

Perform batch evaluation on a single dataset. This method processes batches of time series data through the model, c

Combines expert outputs via communication and adds shared expert output. This method uses the time series(sample) dispatcher to combine the ou

Computes the output of the experts on the dispatched time series(sample). This method first post-processes the dispatched input to get permute

Generates matrix of frequencies based on positions in the sequence, used to create positional encodings

Calculate both MAE and MSE metrics for time series forecasting evaluation. This is a convenience function that computes both primary eva

Args: backcast (torch.Tensor): The previous layer's backcast time series (samples). [batch_size, seq_len]

Generates time series forecasts autoregressively. Args: time_series (torch.Tensor): Input time series data.

Normalization from Non-stationary Transformer

Compute and preprocess time series(sample) routing for dispatch. This method uses the router to determine which experts to send each time seri

Performs SwiGLU (Swish-Gated Linear Unit) activation function. Args: y (torch.Tensor): Input tensor to be split into two halves al

Run comprehensive evaluation across multiple benchmark datasets. This method orchestrates the complete evaluation pipeline:

Preprocess time series(sample) for dispatch. Applies RevIN to input time series(sample), and process the input mask (0: mask,

Get a single sample from the dataset. Args: idx (int): Index of the sample to retrieve Returns: dic

Initialize the evaluation class with configuration parameters.

Initialize the BenchmarkEvalDataset. Args: csv_path (str): Path to the CSV file containing time series data

Initialize FalconTST configuration.

Make the dataset iterable.

Return the total number of samples in the dataset.

Batch quantile forecast inference.

Batch quantile forecast inference.

Forward pass of RoPE embedding. Args: max_seq_len (int): Maximum size of sequence offset (int, optional): RoPE offset.

hidden_states [bs, patch_num, d_model]

Implements the multihead softmax attention. Arguments --------- q,k,v: The tensor containing the query, key, and value. [

x: [seq_len, batch_size, hidden_size] attention_mask: [batch_size, 1, seq_len, seq_len] rotary_pos_emb: [seq_len,

x: [seq_len, batch_size, hidden_size] attention_mask: [batch_size, 1, seq_len, seq_len] rotary_pos_emb: [seq_len,

Function to get the rotary sequence length. Args: transformer_input (Tensor): Input tensor to the transformer Returns:

Apply inverse transformation to denormalize predictions. This method reverses the StandardScaler normalization applied during

Quantile forecast inference.

Quantile forecast inference.