Function _get_tsad_aligned_data

src/servers/tsfm/anomaly.py:368–432

(
    df_data, dataset_config, ad_config, tsmodel_prediction_dictionary
)

def _get_tsad_aligned_data(
    df_data, dataset_config, ad_config, tsmodel_prediction_dictionary
):
    from tsfm_public.toolkit.time_series_preprocessor import create_timestamps

    context_length = ad_config["context_length"]
    prediction_length = ad_config["prediction_length"]
    scaling = ad_config["scaling"]
    ix_target_features = list(
        np.arange(len(dataset_config["column_specifiers"]["target_columns"]))
    )

    df_data[dataset_config["column_specifiers"]["timestamp_column"]] = pd.to_datetime(
        df_data[dataset_config["column_specifiers"]["timestamp_column"]]
    )
    dataset_inference = _get_tsfm_dataloaders(
        df_data,
        {"prediction_length": prediction_length, "context_length": context_length},
        dataset_config,
        scaling=scaling,
    )
    X, y_gt, timestamp_id_value_dic = _tsfm_dataloader_to_array(
        dataset_inference, ix_target_features, x_context_window=context_length
    )

    source_timestamp = np.array(tsmodel_prediction_dictionary["timestamp"])[:, 0]
    target_timestamp = timestamp_id_value_dic["timestamp"]

    forecast_horizon = 1
    target_timestamp_updated = []
    for ts in target_timestamp:
        ts_updated = create_timestamps(
            last_timestamp=ts,
            time_sequence=target_timestamp,
            periods=forecast_horizon,
        )[0]
        target_timestamp_updated.append(ts_updated)
    target_timestamp = np.array(
        np.array(target_timestamp_updated, dtype="datetime64[ns]")
    )
    source_timestamp = np.array(np.array(source_timestamp, dtype="datetime64[ns]"))

    frequency_sampling_median = np.median(target_timestamp[1:] - target_timestamp[:-1])
    tolerance_frequency_sampling = 0.2

    time_diff = np.abs(target_timestamp[:, None] - source_timestamp)
    matching_pairs = np.where(
        time_diff <= frequency_sampling_median * tolerance_frequency_sampling
    )
    index_timestamp = matching_pairs[0]
    index_timestamp_source = matching_pairs[1]

    X_cp = X[index_timestamp]
    y_gt_cp = y_gt[index_timestamp]
    y_pred = np.array(tsmodel_prediction_dictionary["target_prediction"])[
        index_timestamp_source, 0, 0
    ]
    timestamps_source = np.array(source_timestamp)[index_timestamp_source]