Module qbiome.forecaster

Classes

class Forecaster (qnet_orchestrator)

Forecast the data week by week by sequantially generating qnet predictions for the next timestamp and using the filled timestamp to update qnet predictions

Initialization

Args

qnet_orchestrator : qbiome.QnetOrchestrator

an instance with a trained qnet model

Expand source code

class Forecaster:
    """Forecast the data week by week by sequantially generating qnet predictions for the next timestamp and using the filled timestamp to update qnet predictions
    """

    def __init__(self, qnet_orchestrator):
        """Initialization

        Args:
            qnet_orchestrator (qbiome.QnetOrchestrator): an instance with a trained qnet model
        """
        self.qnet_orchestrator = qnet_orchestrator
        self.quantizer = qnet_orchestrator.quantizer

    def forecast_data(self,
                      data,
                      start_week,
                      end_week=None,
                      n_samples=100,
                      subject_id=None):
        """Forecast the data matrix from `start_week` to `end_week`

        Output format:

        |   subject_id | variable         |   week |    value |
        |-------------:|:-----------------|-------:|---------:|
        |            1 | Actinobacteriota |     27 | 0.36665  |
        |            1 | Bacteroidota     |     27 | 0.507248 |
        |            1 | Campilobacterota |     27 | 0.002032 |

        Args:
            data (numpy.ndarray): 2D array of label strings, produced by `self.get_qnet_inputs`
            start_week (int): start predicting from this week
            end_week (int, optional): end predicting after this week
            n_samples (int, optional): the number of times to sample from qnet predictions for one masked entry. Defaults to 100.
            subject_id (list[str], optional): subject id for each row of data. If None, we assume data has as many rows as quantized originally. This must be set for new patient data.

        Returns:
            pandas.DataFrame: see format above
        """
        if end_week is None:
            end_week = self.qnet_orchestrator.get_max_timestamp()
        forecasted_matrix = np.empty(data.shape)
        for idx, seq in enumerate(data):
            forecasted_seq = self.qnet_orchestrator.predict_sequentially_by_week(
                seq, start_week, end_week, n_samples=n_samples
            )
            forecasted_matrix[idx] = forecasted_seq

        if subject_id is None:
            df = self.quantizer.add_meta_to_matrix(forecasted_matrix)
        else:
            df = self.quantizer.add_meta_to_matrix(forecasted_matrix, add_subject_id=False)
            if len(subject_id) != df.index.size:
                raise('Subject_id list must match in length to number of rows in input data matrix in forecaster')
            df['subject_id']=subject_id
        # convert to plottable format
        plot_df = self.quantizer.melt_into_plot_format(df)
        return plot_df

Methods

def forecast_data(self, data, start_week, end_week=None, n_samples=100, subject_id=None)

Forecast the data matrix from start_week to end_week

Output format:

subject_id	variable	week	value
1	Actinobacteriota	27	0.36665
1	Bacteroidota	27	0.507248
1	Campilobacterota	27	0.002032

Args

data : numpy.ndarray

2D array of label strings, produced by self.get_qnet_inputs

start_week : int

start predicting from this week

end_week : int, optional

end predicting after this week

n_samples : int, optional

the number of times to sample from qnet predictions for one masked entry. Defaults to 100.

subject_id : list[str], optional

subject id for each row of data. If None, we assume data has as many rows as quantized originally. This must be set for new patient data.

Returns

pandas.DataFrame

see format above

Expand source code

def forecast_data(self,
                  data,
                  start_week,
                  end_week=None,
                  n_samples=100,
                  subject_id=None):
    """Forecast the data matrix from `start_week` to `end_week`

    Output format:

    |   subject_id | variable         |   week |    value |
    |-------------:|:-----------------|-------:|---------:|
    |            1 | Actinobacteriota |     27 | 0.36665  |
    |            1 | Bacteroidota     |     27 | 0.507248 |
    |            1 | Campilobacterota |     27 | 0.002032 |

    Args:
        data (numpy.ndarray): 2D array of label strings, produced by `self.get_qnet_inputs`
        start_week (int): start predicting from this week
        end_week (int, optional): end predicting after this week
        n_samples (int, optional): the number of times to sample from qnet predictions for one masked entry. Defaults to 100.
        subject_id (list[str], optional): subject id for each row of data. If None, we assume data has as many rows as quantized originally. This must be set for new patient data.

    Returns:
        pandas.DataFrame: see format above
    """
    if end_week is None:
        end_week = self.qnet_orchestrator.get_max_timestamp()
    forecasted_matrix = np.empty(data.shape)
    for idx, seq in enumerate(data):
        forecasted_seq = self.qnet_orchestrator.predict_sequentially_by_week(
            seq, start_week, end_week, n_samples=n_samples
        )
        forecasted_matrix[idx] = forecasted_seq

    if subject_id is None:
        df = self.quantizer.add_meta_to_matrix(forecasted_matrix)
    else:
        df = self.quantizer.add_meta_to_matrix(forecasted_matrix, add_subject_id=False)
        if len(subject_id) != df.index.size:
            raise('Subject_id list must match in length to number of rows in input data matrix in forecaster')
        df['subject_id']=subject_id
    # convert to plottable format
    plot_df = self.quantizer.melt_into_plot_format(df)
    return plot_df