Launches the training process. This method initializes the runner specified in the configuration, sets up logging, and start

Build layer. Args: layer (Callable | Tuple[Callable, Any]): Layer norm function or tuple of layer norm function and kwargs. Returns:

Feed-forward network sublayer. Args: hidden_states (torch.Tensor): Input hidden states. **kwargs: Additional

Args: decoder_layers (nn.ModuleList): The decoder layers. layer_norm (Optional[nn.Module], optional):

The complete evaluation process. Args: ckpt_path (str, optional): Path to the model checkpoint file. Defaults to None.

Get the weight of the forward return

Train model. Args: epoch (int): current epoch iter_index (int): current iter index data (Dict): data

Pack params to the config and add the extra keys to `_keys_to_pop`. Args: params (dict): params obj_params (dict): pa

Get value. Args: name (str): meter name. Returns: avg (float)

Create a mask for the input data. Args: data (torch.Tensor): Input data. mask_ratio (float): Ratio of the data to be masked.

Apply the scaling transformation to the input data. Args: input_data (torch.Tensor): Input data to be transformed.

Updates the cache with the new `key_states` and `value_states` for the layer `layer_idx`. Parameters: key_states (`torch

Update average meter. Args: name (str): meter name. value (Union[float, Tuple[float]]): value. n: (int):

patch_len: int, patch len for patch_embedding stride: int, stride for patch_embedding

Build dataloader from BasicTSConfig structure of `data_cfg` is { 'BATCH_SIZE': (int, optional) batch size of data loader

Mask abnormal values in the sequence. Args: seq (np.ndarray): input sequence Returns: np.ndarray: s

Reshape input to (batch, heads, seq_len, head_dim).

The complete test process. It should not be directly called. Args: train_epoch (int, optional): Current epoch during tra

Calculate the transition matrix as proposed in DCRNN and Graph WaveNet. The transition matrix is defined as P = D^{-1} A, where D is the deg

Initialize logger. Args: logger (logging.Logger, optional): specified logger. logger_name (str, optional): specified

Apply the inverse scaling transformation to the input data. Args: input_data (torch.Tensor): Input data to be transforme

Run the task flow

Self-attention sublayer. Args: hidden_states (torch.Tensor): Input hidden states. attention_mask (Optional[t

Initialize the Auto-Attention module. Args: d_model (int): The input and output dimension for queries, keys, and values.

Create a SparseDispatcher.

Build a dataset with the given config and mode. Args: cfg (BasicTSConfig): config mode (BasicTSMode): mode R

evaluate model.

Get checkpoint path. The format is "{ckpt_save_dir}/{model_name}_{epoch}" Args: epoch (int): current epoch. Ret

Initialize the test data loader and related settings. Args: cfg (BasicTSConfig): Configuration dictionary.

Initialize validation

Load data from local file or remotely. Returns: np.ndarray: loaded data

Load model state dict. if param `ckpt_path` is None, load the last checkpoint in `self.ckpt_save_dir`, else load checkpoint from `ckpt

Zscore normalize the inputs and targets. Args: inputs (np.ndarray): input sequences targets (Optional[np.nda

Common processing pipeline for frequency levels. Args: level_list: List of frequency levels (from high to low)

Rotates half the hidden dims of the input.

Save the best model while training. Examples: >>> def on_validating_end(self, train_epoch: Optional[int]): >>> if

Save checkpoint every epoch. checkpoint format is { "epoch": current epoch ([1, num_epochs]), "model_state_dict": sta

Calculate the symmetric normalized Laplacian. The symmetric normalized Laplacian matrix is given by: L^{Sym} = I - D^{-1/2} A D^{-1/2},

Forward pass of the cross-attention layer. Args: hidden_states (torch.Tensor): The input hidden states.

The squared coefficient of variation of a sample. Useful as a loss to encourage a positive distribution to be more uniform. Epsilons a

Fit the scaler to the training data. Args: data (torch.Tensor): Training data used to fit the scaler.

获取指定目录下指定后缀的文件字典 :param file_dir: 文件目录 :param postfix: 文件后缀 :param specify_dir_depth: 指定目录深度，0表示不指定 :param key_remain_length

Load and preprocess raw data, selecting the specified channel(s).

Load data from a pickle file. Args: pickle_file (str): Path to the pickle file. Returns: object: Loaded object from the

Run the task flow

Generate a causal attention mask (including padding mask) for the decoder. Args: input_shape: (batch_size, seq_len) inpu

Validate model. Args: cfg (Dict, optional): config train_epoch (int, optional): current epoch if in training process.

Initialize the MLP layer. Args: input_size (int): Input feature dimension. intermediate_size (int): Hidden d

Build model. Initialize model by calling ```self.define_model```, Moves model to the GPU. If DDP is initialized, initialize

Check required callbacks.

Compute the residual entropy for time series. This is an implementation for residual that follows a normal distribution. Arg