Source code for lightwood.data.timeseries_transform
from typing import Dict, Optional
from functools import partial
import multiprocessing as mp
import numpy as np
import pandas as pd
from lightwood.helpers.parallelism import get_nr_procs
from type_infer.dtype import dtype
from lightwood.api.types import TimeseriesSettings, PredictionArguments
from lightwood.helpers.log import log
[docs]def transform_timeseries(
data: pd.DataFrame, dtype_dict: Dict[str, str],
timeseries_settings: TimeseriesSettings, target: str, mode: str,
pred_args: Optional[PredictionArguments] = None
) -> pd.DataFrame:
"""
Block that transforms the dataframe of a time series task to a convenient format for use in posterior phases like model training.
The main transformations performed by this block are:
- Type casting (e.g. to numerical for `order_by` column).
- Windowing functions for historical context based on `TimeseriesSettings.window` parameter.
- Explicitly add target columns according to the `TimeseriesSettings.horizon` parameter.
- Flag all rows that are "predictable" based on all `TimeseriesSettings`.
- Plus, handle all logic for the streaming use case (where forecasts are only emitted for the last observed data point).
:param data: Dataframe with data to transform.
:param dtype_dict: Dictionary with the types of each column.
:param timeseries_settings: A `TimeseriesSettings` object.
:param target: The name of the target column to forecast.
:param mode: Either "train" or "predict", depending on what phase is calling this procedure.
:param pred_args: Optional prediction arguments to control the transformation process.
:return: A dataframe with all the transformations applied.
""" # noqa
pred_args = PredictionArguments() if pred_args is None else pred_args
tss = timeseries_settings
gb_arr = tss.group_by if tss.group_by is not None else []
oby = tss.order_by
window = tss.window
oby_col = tss.order_by
if tss.use_previous_target and target not in data.columns:
raise Exception(f"Cannot transform. Missing historical values for target column {target} (`use_previous_target` is set to True).") # noqa
for hcol in tss.historical_columns:
if hcol not in data.columns or data[hcol].isna().any():
raise Exception(f"Cannot transform. Missing values in historical column {hcol}.")
# initial stable sort and per-partition deduplication TODO: slowish, add a top-level param to disable if needed
data = data.sort_values(by=oby_col, kind='mergesort')
data = data.drop_duplicates(subset=[oby_col, *gb_arr], keep='first')
# pass seconds to timestamps according to each group's inferred freq, and force this freq on index
grouped = data.groupby(by=tss.group_by) if tss.group_by else data.groupby(lambda x: True)
reindexed = []
# TODO: introduce MP here
for name, group in grouped:
name = name if tss.group_by and len(tss.group_by) > 1 else (name, ) # guaranteed tuple type
if group.shape[0] > 0:
if group[tss.order_by].value_counts().max() > 1 and group.shape[0] > 1:
raise Exception(f"Partition is not valid, faulty group {name}. Please make sure you group by a set of columns that ensures unique measurements for each grouping through time.") # noqa
index = pd.to_datetime(group[oby_col], unit='s', utc=True)
group.index = pd.date_range(start=index.iloc[0], end=index.iloc[-1], periods=len(group))
resampled = group
group['__mdb_inferred_freq'] = None
if len(group) > 2:
freq = pd.infer_freq(group.index)
if freq is not None:
group['__mdb_inferred_freq'] = freq # sets constant column because pd.concat forgets freq (see: https://github.com/pandas-dev/pandas/issues/3232) # noqa
resampled = group.resample(freq).first()
reindexed.append(resampled)
original_df = pd.concat(reindexed).sort_values(by='__mdb_original_index')
if '__mdb_forecast_offset' in original_df.columns:
""" This special column can be either None or an integer. If this column is passed, then the TS transformation will react to the values within:
* If all rows = `None`, proceed as usual. This ends up generating one HORIZON-length forecast for each row in the DF.
* If all rows have the same value `N <= 0`, then cutoff the dataframe latest `-N` rows after TS shaping and prime the DF (with `__make_predictions` column) so that a forecast is generated only for the last row (thus more efficient). This enables `WHERE T = LATEST - K` with `0 <= K < WINDOW` syntax upstream in MindsDB.
* If all rows have the same value `N = 1`, then activate streaming inference mode where a single forecast will be emitted for the timestamp inferred by the `_ts_infer_next_row` method. This enables the (already supported) `WHERE T > LATEST` syntax.
""" # noqa
index = original_df[~original_df['__mdb_forecast_offset'].isin([None])] # trigger if col is constant & != None
offset_available = index.shape[0] == len(original_df) and \
original_df['__mdb_forecast_offset'].unique().tolist() != [None]
if offset_available:
offset = min(int(original_df['__mdb_forecast_offset'].unique()[0]), 1)
else:
offset = 0
cutoff_mode = offset_available and offset == 1
else:
offset_available = False
offset = 0
cutoff_mode = False
if '__mdb_forecast_offset' in original_df.columns or cutoff_mode:
original_df['__make_predictions'] = True
original_df['original_index'] = np.arange(len(original_df))
else:
original_df['__make_predictions'] = False
original_df['original_index'] = None
secondary_type_dict = {}
if dtype_dict[oby] in (dtype.date, dtype.integer, dtype.float):
secondary_type_dict[oby] = dtype_dict[oby]
original_df[f'__mdb_original_{oby}'] = original_df[oby]
original_df = _ts_to_obj(original_df, [oby] + tss.historical_columns)
group_lengths = []
if len(gb_arr) > 0:
df_arr = []
for _, df in original_df.groupby(gb_arr):
df_arr.append(df.sort_values(by=oby))
group_lengths.append(len(df))
else:
df_arr = [original_df.sort_values(by=oby)]
group_lengths.append(len(original_df))
n_groups = len(df_arr)
for i, subdf in enumerate(df_arr):
if '__mdb_forecast_offset' in subdf.columns and mode == 'predict':
if cutoff_mode:
df_arr[i] = _ts_infer_next_row(subdf, oby)
make_preds = [False for _ in range(max(0, len(df_arr[i]) - 1))] + [True]
elif offset_available:
# truncate to forecast up until some len(df) + offset (which is <= 0)
new_index = df_arr[i].index[:len(df_arr[i].index) + offset]
make_preds = [False for _ in range(max(0, len(new_index) - 1))] + [True]
df_arr[i] = df_arr[i].loc[new_index]
else:
if pred_args.force_ts_infer:
df_arr[i] = _ts_infer_next_row(subdf, oby) # force-infer out-of-sample forecast in default mode
make_preds = [True for _ in range(len(df_arr[i]))]
df_arr[i]['__make_predictions'] = make_preds
if len(df_arr) > 1 and len(original_df) > 5000:
# @TODO: restore possibility to override this with args
biggest_sub_df = df_arr[np.argmax(group_lengths)]
nr_procs = min(get_nr_procs(biggest_sub_df), len(df_arr))
log.info(f'Using {nr_procs} processes to reshape.')
with mp.Pool(processes=nr_procs) as pool:
df_arr = pool.map(
partial(_ts_add_previous_rows, order_cols=[oby] + tss.historical_columns, window=window),
df_arr
)
df_arr = pool.map(
partial(_ts_add_future_target, target=target, horizon=tss.horizon,
data_dtype=tss.target_type, mode=mode),
df_arr
)
if tss.use_previous_target:
df_arr = pool.map(
partial(_ts_add_previous_target, target=target, window=tss.window),
df_arr
)
else:
for i in range(n_groups):
df_arr[i] = _ts_add_previous_rows(df_arr[i],
order_cols=[oby] + tss.historical_columns, window=window)
df_arr[i] = _ts_add_future_target(df_arr[i], target=target, horizon=tss.horizon,
data_dtype=tss.target_type, mode=mode)
if tss.use_previous_target:
df_arr[i] = _ts_add_previous_target(df_arr[i], target=target, window=tss.window)
combined_df = pd.concat(df_arr)
if '__mdb_forecast_offset' in combined_df.columns:
combined_df = pd.DataFrame(combined_df[combined_df['__make_predictions']]) # filters by True only
del combined_df['__make_predictions']
if not cutoff_mode and any([i < tss.window for i in group_lengths]):
if tss.allow_incomplete_history:
log.warning("Forecasting with incomplete historical context, predictions might be subpar")
else:
raise Exception(f'Not enough historical context to make a timeseries prediction (`allow_incomplete_history` is set to False). Please provide a number of rows greater or equal to the window size - currently (number_rows, window_size) = ({min(group_lengths)}, {tss.window}). If you can\'t get enough rows, consider lowering your window size. If you want to force timeseries predictions lacking historical context please set the `allow_incomplete_history` timeseries setting to `True`, but this might lead to subpar predictions depending on the mixer.') # noqa
if n_groups > 1:
df_gb_list = list(combined_df.groupby(tss.group_by))
df_gb_map = {}
for gb, df in df_gb_list:
df_gb_map['_' + '_'.join(str(gb))] = df
del combined_df['original_index']
return combined_df
def _ts_infer_next_row(df: pd.DataFrame, ob: str) -> pd.DataFrame:
"""
Adds an inferred next row for streaming mode purposes.
:param df: dataframe from which next row is inferred.
:param ob: `order_by` column.
:return: Modified `df` with the inferred row appended to it.
"""
original_index = df.index.copy()
start = original_index.min()
new_index = pd.date_range(start=start, periods=len(original_index) + 1, freq=df['__mdb_inferred_freq'].iloc[0])
last_row = df.iloc[[-1]].copy()
last_row['__mdb_ts_inferred'] = True
if df.shape[0] > 1:
butlast_row = df.iloc[[-2]]
delta = (last_row[ob].values - butlast_row[ob].values).flatten()[0]
else:
delta = 1
last_row[ob] += delta
new_df = pd.concat([df, last_row], ignore_index=True)
new_df.index = pd.DatetimeIndex(new_index)
return new_df
def _ts_to_obj(df: pd.DataFrame, historical_columns: list) -> pd.DataFrame:
"""
Casts all historical columns in a dataframe to `object` type.
:param df: Input dataframe
:param historical_columns: Historical columns to type cast
:return: Dataframe with `object`-typed historical columns
"""
for hist_col in historical_columns:
df.loc[:, hist_col] = df[hist_col].astype(object)
return df
def _ts_add_previous_rows(df: pd.DataFrame, order_cols: list, window: int) -> pd.DataFrame:
"""
Adds previous rows (as determined by `TimeseriesSettings.window`) into the cells of the `order_by` column.
:param df: Input dataframe.
:param order_cols: `order_by` column and other columns flagged as `historical`.
:param window: value of `TimeseriesSettings.window` parameter.
:return: Dataframe with all `order_cols` modified so that their values are now arrays of historical context.
""" # noqa
for order_col in order_cols:
new_vals = np.zeros((len(df), window))
for i in range(window, 0, -1):
new_vals[:, i - 1] = df[order_col].shift(window - i).values
new_vals = np.nan_to_num(new_vals, nan=0.0)
for i in range(len(df)):
df.at[df.index[i], order_col] = new_vals[i, :]
return df
def _ts_add_previous_target(df: pd.DataFrame, target: str, window: int) -> pd.DataFrame:
"""
Adds previous rows (as determined by `TimeseriesSettings.window`) into the cells of the target column.
:param df: Input dataframe.
:param target: target column name.
:param window: value of `TimeseriesSettings.window` parameter.
:return: Dataframe with new `__mdb_ts_previous_{target}` column that contains historical target context.
""" # noqa
if target not in df:
return df
previous_target_values = list(df[target])
del previous_target_values[-1]
previous_target_values = [None] + previous_target_values
previous_target_values_arr = []
for i in range(len(previous_target_values)):
prev_vals = previous_target_values[max(i - window, 0):i + 1]
arr = [None] * (window - len(prev_vals))
arr.extend(prev_vals)
previous_target_values_arr.append(arr)
df[f'__mdb_ts_previous_{target}'] = previous_target_values_arr
return df
def _ts_add_future_target(df, target, horizon, data_dtype, mode):
"""
Adds as many columns to the input dataframe as the forecasting horizon asks for (as determined by `TimeseriesSettings.horizon`).
:param df: Input dataframe.
:param target: target column name.
:param horizon: value of `TimeseriesSettings.horizon` parameter.
:param data_dtype: dictionary with types of all input columns
:param mode: either "train" or "predict". `Train` will drop rows with incomplet target info. `Predict` has no effect, for now.
:return: Dataframe with new `{target}_timestep_{i}'` columns that contains target labels at timestep `i` of a total `TimeseriesSettings.horizon`.
""" # noqa
if target not in df:
return df
if data_dtype in (dtype.integer, dtype.float, dtype.num_array, dtype.num_tsarray):
df[target] = df[target].astype(float)
for timestep_index in range(1, horizon):
next_target_value_arr = list(df[target])
for del_index in range(0, min(timestep_index, len(next_target_value_arr))):
del next_target_value_arr[0]
next_target_value_arr.append(None)
col_name = f'{target}_timestep_{timestep_index}'
df[col_name] = next_target_value_arr
df[col_name] = df[col_name].fillna(value=np.nan)
return df