Others

Data Manager

nazuna.data_manager.TimeSeriesDataManager

Manager for time-series data that handles loading and splitting.

Note on seq_len and pred_len

These parameters affect data splitting. When you specify a data range (e.g., 0.0-0.5 for training), the actual row indices depend on seq_len and pred_len. The actual sequence lengths used during evaluation or training can be shorter (specified in the criterion and model). E.g., for fair comparison across experiments with varying input lengths, set seq_len to the maximum value you plan to use.

__init__(path, colname_timestamp, seq_len, pred_len, white_list=None, step_start=0, step_width=1)

Parameters:

• path (str) –

  Path to the CSV file.

• colname_timestamp (str) –

  Column name for timestamps.

• seq_len (int) –

  Input sequence length (affects data splitting).

• pred_len (int) –

  Prediction sequence length (affects data splitting).

• white_list (str | list[str] | None, default: None ) –

  Column names to use. If None, all columns are used.

Batch Sampler

nazuna.batch_samplers.BatchSampler

A sampler that splits samples into batches in sequential order.

__init__(n_sample, batch_size)

Parameters:

• n_sample (int) –

  Total number of samples.

• batch_size (int) –

  Batch size.

nazuna.batch_samplers.BatchSamplerShuffle

Bases: BatchSamplerRandom

A sampler that shuffles samples and splits them into batches.

• At the start of each iteration, sample indices are shuffled, and batches are created based on the shuffled order.

__init__(n_sample, batch_size, seed=0)

Parameters:

• n_sample (int) –

  Total number of samples.

• batch_size (int) –

  Batch size.

• seed (int, default: 0 ) –

  Random seed.

Criteria

nazuna.criteria.MSE

Bases: BaseError

Mean Squared Error with channel and sequence weighting.

_setup(n_channel, pred_len, decay_rate=None, tolerance=0)

Parameters:

• n_channel (int) –

  Number of channels in the prediction.

• pred_len (int) –

  Length of the prediction sequence.

• decay_rate (float | None, default: None ) –

  Exponential decay rate for sequence weighting. If None or 1, uniform weights are used.

• tolerance (float, default: 0 ) –

  Errors below this threshold are treated as zero. Default is 0 (no tolerance).

Note

The weight for step i is decay_rate^i (before normalization). For example, if decay_rate=0.9, step 0 has weight 1.0, step 1 has 0.9, step 2 has 0.81, and so on.

nazuna.criteria.MAE

Bases: MSE

nazuna.criteria.ImprovementRate

Bases: BaseImprovement

Computes improvement rate over a baseline prediction.

Improvement rate = 1 - (pred_error / baseline_error) - Positive values indicate improvement over baseline - Zero indicates equal performance - Negative values indicate worse performance than baseline

_setup(n_channel, pred_len, decay_rate=None, tolerance=0, epsilon=1e-06, error_type='mae')

Parameters:

• n_channel (int) –

  Number of channels in the prediction.

• pred_len (int) –

  Length of the prediction sequence.

• decay_rate (float | None, default: None ) –

  Exponential decay rate for sequence weighting. If None or 1, uniform weights are used.

• tolerance (float, default: 0 ) –

  Errors below this threshold are treated as zero. Default is 0 (no tolerance).

• epsilon (float, default: 1e-06 ) –

  Small value to avoid division by zero.

• error_type (str, default: 'mae' ) –

  Type of error metric ("mse" or "mae").

forward(baseline, pred, true)

Parameters:

• baseline –

  Baseline model prediction (batch_size, pred_len, n_channel)

• pred –

  Model prediction to evaluate (batch_size, pred_len, n_channel)

• true –

  Ground truth (batch_size, pred_len, n_channel)

Returns:

• TimeSeriesError –

  Aggregated improvement rates