Source code for darts.dataprocessing.transformers.base_data_transformer
"""
Data Transformer Base Class
---------------------------
"""
from abc import ABC, abstractmethod
from typing import Any, Generator, Iterable, List, Mapping, Optional, Sequence, Union
import numpy as np
import xarray as xr
from darts import TimeSeries
from darts.logging import get_logger, raise_log
from darts.utils import _build_tqdm_iterator, _parallel_apply
logger = get_logger(__name__)
[docs]class BaseDataTransformer(ABC):
def __init__(
self,
name: str = "BaseDataTransformer",
n_jobs: int = 1,
verbose: bool = False,
parallel_params: Union[bool, Sequence[str]] = False,
mask_components: bool = True,
):
"""Abstract class for data transformers.
All the deriving classes have to implement the static method :func:`ts_transform`; this implemented
method can then be applied to ``TimeSeries`` or ``Sequence[TimeSeries]`` inputs by calling the
:func:`transform()` method. Internally, :func:`transform()` parallelizes func:`ts_transform` over all
of the ``TimeSeries`` inputs passed to it. See the func:`ts_transform` method documentation
for further details on how to implement this method in a user-defined class.
Data transformers requiring to be fit first before calling :func:`transform` should derive
from :class:`.FittableDataTransformer` instead. Data transformers that are invertible should derive
from :class:`.InvertibleDataTransformer` instead. Transformers which are both fittable and invertible
should inherit from both :class:`.FittableDataTransformer` and :class:`.InvertibleDataTransformer`.
All Data Transformers can store *fixed parameters* that are automatically passed to func:`ts_transform`;
the fixed parameters of a data transformer object are taken to be all those attributes defined
in the `__init__` method of the child-most class *before* `super().__init__` is called. The fixed parameter
values can then be accessed within the func:`ts_transform` method through the `params` dictionary
argument. More specifically, `params['fixed']` stores a dictionary with all of the fixed parameter
values, where the keys are simply the attribute names of each fixed parameter (e.g. the `self._my_param`
fixed parameter attribute is accessed through `params['fixed']['_my_param']`).
Data Transformers which inherit from :class:`.FittableDataTransformer` can also store fitted parameters
alongside fixed parameters; please refer to the :class:`.FittableDataTransformer` documentation for
further details.
Note: the :func:`ts_transform` method is designed to be a static method instead of a instance method to allow an
efficient parallelisation also when the scaler instance is storing a non-negligible amount of data. Using
an instance method would imply copying the instance's data through multiple processes, which can easily
introduce a bottleneck and nullify parallelisation benefits.
Parameters
----------
name
The data transformer's name
n_jobs
The number of jobs to run in parallel. Parallel jobs are created only when a ``Sequence[TimeSeries]``
is passed as input to a method, parallelising operations regarding different TimeSeries.
Defaults to `1` (sequential). Setting the parameter to `-1` means using all the available processors.
Note: for a small amount of data, the parallelisation overhead could end up increasing the total
required amount of time.
verbose
Optionally, whether to print operations progress
parallel_params
Optionally, specifies which fixed parameters (i.e. the attributes initialized in the child-most
class's `__init__`) take on different values for different parallel jobs. Fixed parameters specified
by `parallel_params` are assumed to be a `Sequence` of values that should be used for that parameter
in each parallel job; the length of this `Sequence` should equal the number of parallel jobs. If
`parallel_params=True`, every fixed parameter will take on a different value for each
parallel job. If `parallel_params=False`, every fixed parameter will take on the same value for
each parallel job. If `parallel_params` is a `Sequence` of fixed attribute names, only those
attribute names specified will take on different values between different parallel jobs.
mask_components
Optionally, whether to automatically apply any provided `component_mask`s to the
`TimeSeries` inputs passed to `transform`, `fit`, `inverse_transform`, or `fit_transform`.
If `True`, any specified `component_mask` will be applied to each input timeseries
before passing them to the called method; the masked components will also be automatically
'unmasked' in the returned `TimeSeries`. If `False`, then `component_mask` (if provided) will
be passed as a keyword argument, but won't automatically be applied to the input timeseries.
See `apply_component_mask` for further details.
Example
--------
>>> from darts.dataprocessing.transformers import BaseDataTransformer
>>> from darts.utils.timeseries_generation import linear_timeseries
>>>
>>> class SimpleTransform(BaseDataTransformer):
>>>
>>> def __init__(self, a):
>>> self._a = a
>>> super().__init__()
>>>
>>> @staticmethod
>>> def ts_transform(series, params, **kwargs):
>>> a = params['fixed']['_a']
>>> b = kwargs.pop('b')
>>> return a*series + b
>>>
>>> series = linear_timeseries(length=5)
>>> print(series)
<TimeSeries (DataArray) (time: 5, component: 1, sample: 1)>
array([[[0. ]],
[[0.25]],
[[0.5 ]],
[[0.75]],
[[1. ]]])
Coordinates:
* time (time) datetime64[ns] 2000-01-01 2000-01-02 ... 2000-01-05
* component (component) object 'linear'
Dimensions without coordinates: sample
Attributes:
static_covariates: None
hierarchy: None
>>> series = SimpleTransform(a=2).transform(series, b=3)
>>> print(series)
<TimeSeries (DataArray) (time: 5, component: 1, sample: 1)>
array([[[3. ]],
[[3.5]],
[[4. ]],
[[4.5]],
[[5. ]]])
Coordinates:
* time (time) datetime64[ns] 2000-01-01 2000-01-02 ... 2000-01-05
* component (component) object 'linear'
Dimensions without coordinates: sample
Attributes:
static_covariates: None
hierarchy: None
"""
# Assume `super().__init__` called at *very end* of
# child-most class's `__init__`, so `vars(self)` contains
# only attributes of this child-most class:
self._fixed_params = vars(self).copy()
# If `parallel_params = True`, but not a list of key values:
if parallel_params and not isinstance(parallel_params, Sequence):
parallel_params = self._fixed_params.keys()
# If `parallel_params = False`:
elif not parallel_params:
parallel_params = tuple()
self._parallel_params = parallel_params
self._mask_components = mask_components
self._name = name
self._verbose = verbose
self._n_jobs = n_jobs
[docs] def set_verbose(self, value: bool):
"""Set the verbosity status.
`True` for enabling the detailed report about scaler's operation progress,
`False` for no additional information.
Parameters
----------
value
New verbosity status
"""
if not isinstance(value, bool):
raise_log(ValueError("Verbosity status must be a boolean."), logger=logger)
self._verbose = value
[docs] def set_n_jobs(self, value: int):
"""Set the number of processors to be used by the transformer while processing multiple ``TimeSeries``.
Parameters
----------
value
New n_jobs value. Set to `-1` for using all the available cores.
"""
if not isinstance(value, int):
raise_log(ValueError("n_jobs must be an integer"), logger=logger)
self._n_jobs = value
[docs] @staticmethod
@abstractmethod
def ts_transform(series: TimeSeries, params: Mapping[str, Any]) -> TimeSeries:
"""The function that will be applied to each series when :func:`transform()` is called.
This method is not implemented in the base class and must be implemented in the deriving classes.
The function must take as first argument a ``TimeSeries`` object and, as a second argument, a
dictionary containing the fixed and/or fitted parameters of the transformation; this function
should then return a transformed ``TimeSeries`` object.
The `params` dictionary *can* contain up to two keys:
1. `params['fixed']` stores the fixed parameters of the transformation (i.e. attributed
defined in the `__init__` method of the child-most class *before* `super().__init__` is called);
`params['fixed']` is a dictionary itself, whose keys are the names of the fixed parameter
attributes. For example, if `_my_fixed_param` is defined as an attribute in the child-most
class, then this fixed parameter value can be accessed through `params['fixed']['_my_fixed_param']`.
2. If the transform inherits from the :class:`.FittableDataTransformer` class, then `params['fitted']`
will store the fitted parameters of the transformation; the fitted parameters are simply the output(s)
returned by the `ts_fit` function, whatever those output(s) may be. See :class:`.FittableDataTransformer`
for further details about fitted parameters.
Any positional/keyword argument supplied to the `transform` method are passed as positional/keyword arguments
to `ts_transform`; hence, `ts_transform` should also accept `*args` and/or `**kwargs` if positional/keyword
arguments are passed to `transform`. Note that if the `mask_components` attribute of `BaseDataTransformer`
is set to `False`, then the `component_mask` provided to `transform` will be passed as an additional keyword
argument to `ts_transform`.
The `BaseDataTransformer` class includes some helper methods which may prove useful when implementing a
`ts_transform` function:
1. The `apply_component_mask` and `unapply_component_mask` methods, which apply and 'unapply'
`component_mask`s to a `TimeSeries` respectively; these methods are automatically called in `transform`
if the `mask_component` attribute of `BaseDataTransformer` is set to `True`, but you may want to manually
call them if you set `mask_components` to `False` and wish to manually specify how `component_mask`s are
applied to a `TimeSeries`.
2. The `stack_samples` method, which stacks all the samples in a `TimeSeries` along
the component axis, so that the `TimeSeries` goes from shape `(n_timesteps, n_components, n_samples)` to
shape `(n_timesteps, n_components * n_samples)`. This stacking is useful if a pointwise transform is being
implemented (i.e. transforming the value at time `t` depends only on the value of the series at that
time `t`). Once transformed, the stacked `TimeSeries` can be 'unstacked' using the `unstack_samples` method.
Parameters
----------
series
series to be transformed.
params
Dictionary containing the parameters of the transformation function. Fixed parameters
(i.e. attributes defined in the child-most class of the transformation prior to
calling `super.__init__()`) are stored under the `'fixed'` key. If the transformation
inherits from the `FittableDataTransformer` class, then the fitted parameters of the
transformation (i.e. the values returned by `ts_fit`) are stored under the
`'fitted'` key.
args
Any poisitional arguments provided in addition to `series` when
kwargs
Any additional keyword arguments provided to `transform`. Note that if the `mask_component`
attribute of `BaseDataTransformer` is set to `False`, then `component_mask` will
be passed as a keyword argument.
Notes
-----
This method is designed to be a static method instead of instance method to allow an efficient
parallelisation also when the scaler instance is storing a non-negligible amount of data. Using instance
methods would imply copying the instance's data through multiple processes, which can easily introduce a
bottleneck and nullify parallelisation benefits.
"""
pass
[docs] def transform(
self,
series: Union[TimeSeries, Sequence[TimeSeries]],
*args,
component_mask: Optional[np.array] = None,
**kwargs,
) -> Union[TimeSeries, List[TimeSeries]]:
"""Transforms a (sequence of) of series by calling the user-implemeneted `ts_transform` method.
In case a ``Sequence[TimeSeries]`` is passed as input data, this function takes care of
parallelising the transformation of multiple series in the sequence at the same time. Additionally,
if the `mask_components` attribute was set to `True` when instantiating `BaseDataTransformer`,
then any provided `component_mask`s will be automatically applied to each input `TimeSeries`;
please refer to 'Notes' for further details on component masking.
Any additionally specified `*args` and `**kwargs` are automatically passed to `ts_transform`.
Parameters
----------
series
(sequence of) series to be transformed.
args
Additional positional arguments for each :func:`ts_transform()` method call
component_mask : Optional[np.ndarray] = None
Optionally, a 1-D boolean np.ndarray of length ``series.n_components`` that specifies which
components of the underlying `series` the transform should consider. If the `mask_components`
attribute was set to `True` when instantiating `BaseDataTransformer`, then the component mask
will be automatically applied to each `TimeSeries` input. Otherwise, `component_mask` will be
provided as an addition keyword argument to `ts_transform`. See 'Notes' for further details.
kwargs
Additional keyword arguments for each :func:`ts_transform()` method call
Returns
-------
Union[TimeSeries, List[TimeSeries]]
Transformed data.
Notes
-----
If the `mask_components` attribute was set to `True` when instantiating `BaseDataTransformer`,
then any provided `component_mask`s will be automatically applied to each `TimeSeries` input to
transform; `component_mask`s are simply boolean arrays of shape `(series.n_components,)` that
specify which components of each `series` should be transformed using `ts_transform` and which
components should not. If `component_mask[i]` is `True`, then the `i`th component of each
`series` will be transformed by `ts_transform`. Conversely, if `component_mask[i]` is `False`,
the `i`th component will be removed from each `series` before being passed to `ts_transform`;
after transforming this masked series, the untransformed `i`th component will be 'added back'
to the output. Note that automatic `component_mask`ing can only be performed if the `ts_transform`
does *not* change the number of timesteps in each series; if this were to happen, then the transformed
and untransformed components are unable to be concatenated back together along the component axis.
If `mask_components` was set to `False` when instantiating `BaseDataTransformer`, then any provided
`component_masks` will be passed as a keyword argument `ts_transform`; the user can then manually specify
how the `component_mask` should be applied to each series.
"""
desc = f"Transform ({self._name})"
# Take note of original input for unmasking purposes:
if isinstance(series, TimeSeries):
input_series = [series]
data = [series]
transformer_selector = [0]
else:
input_series = series
data = series
transformer_selector = range(len(series))
if self._mask_components:
data = [
self.apply_component_mask(ts, component_mask, return_ts=True)
for ts in data
]
else:
kwargs["component_mask"] = component_mask
input_iterator = _build_tqdm_iterator(
zip(data, self._get_params(transformer_selector=transformer_selector)),
verbose=self._verbose,
desc=desc,
total=len(data),
)
transformed_data = _parallel_apply(
input_iterator, self.__class__.ts_transform, self._n_jobs, args, kwargs
)
if self._mask_components:
unmasked = []
for ts, transformed_ts in zip(input_series, transformed_data):
unmasked.append(
self.unapply_component_mask(ts, transformed_ts, component_mask)
)
transformed_data = unmasked
return (
transformed_data[0] if isinstance(series, TimeSeries) else transformed_data
)
def _get_params(
self, transformer_selector: Iterable
) -> Generator[Mapping[str, Any], None, None]:
"""
Creates generator of dictionaries containing fixed parameter values
(i.e. attributes defined in the child-most class). Those fixed parameters
specified by `parallel_params` are given different values over each of the
parallel jobs. Called by `transform` and `inverse_transform`,
if `Transformer` does *not* inherit from `FittableTransformer`.
"""
self._check_fixed_params(transformer_selector)
def params_generator(transformer_selector, fixed_params, parallel_params):
fixed_params_copy = fixed_params.copy()
for i in transformer_selector:
for key in parallel_params:
fixed_params_copy[key] = fixed_params[key][i]
if fixed_params_copy:
params = {"fixed": fixed_params_copy}
else:
params = None
yield params
return None
return params_generator(
transformer_selector, self._fixed_params, self._parallel_params
)
def _check_fixed_params(self, transformer_selector: Iterable) -> None:
"""
Raises `ValueError` if `self._parallel_params` specifies a `key` in
`self._fixed_params` that should be distributed, but
`len(self._fixed_params[key])` does not equal to the number of time series
(the maximum value + 1 from `transformer_selector`).
"""
for key in self._parallel_params:
n_timeseries_ = max(transformer_selector) + 1
if n_timeseries_ > len(self._fixed_params[key]):
raise_log(
ValueError(
f"{n_timeseries_} TimeSeries were provided "
f"but only {len(self._fixed_params[key])} {key} values "
f"were specified upon initialising {self.name}."
),
logger=logger,
)
elif n_timeseries_ < len(self._fixed_params[key]):
logger.warning(
f"Only {n_timeseries_} TimeSeries were provided "
f"which is lower than the number of {key} values "
f"(n={len(self._fixed_params[key])}) that were specified "
f"upon initialising {self.name}."
)
return None
[docs] @staticmethod
def apply_component_mask(
series: TimeSeries, component_mask: Optional[np.ndarray] = None, return_ts=False
) -> np.ndarray:
"""
Extracts components specified by `component_mask` from `series`
Parameters
----------
series
input TimeSeries to be fed into transformer.
component_mask
Optionally, np.ndarray boolean mask of shape (n_components, 1) specifying which components to
extract from `series`. The `i`th component of `series` is kept only if `component_mask[i] = True`.
If not specified, no masking is performed.
return_ts
Optionally, specifies that a `TimeSeries` should be returned, rather than an `np.ndarray`.
Returns
-------
masked
`TimeSeries` (if `return_ts = True`) or `np.ndarray` (if `return_ts = False`) with only those components
specified by `component_mask` remaining.
"""
if component_mask is None:
masked = series.copy() if return_ts else series.all_values()
else:
if not (
isinstance(component_mask, np.ndarray) and component_mask.dtype == bool
):
raise_log(
ValueError(
f"`component_mask` must be a boolean `np.ndarray`, not a {type(component_mask)}."
),
logger=logger,
)
if not series.width == len(component_mask):
raise_log(
ValueError(
"mismatch between number of components in `series` and length of `component_mask`"
),
logger=logger,
)
masked = series.all_values(copy=False)[:, component_mask, :]
if return_ts:
# Remove masked components from coords:
coords = dict(series._xa.coords)
coords["component"] = coords["component"][component_mask]
new_xa = xr.DataArray(
masked, dims=series._xa.dims, coords=coords, attrs=series._xa.attrs
)
masked = TimeSeries(new_xa)
return masked
[docs] @staticmethod
def unapply_component_mask(
series: TimeSeries,
vals: Union[np.ndarray, TimeSeries],
component_mask: Optional[np.ndarray] = None,
) -> Union[np.ndarray, TimeSeries]:
"""
Adds back components previously removed by `component_mask` in `apply_component_mask` method.
Parameters
----------
series
input TimeSeries that was fed into transformer.
vals:
`np.ndarray` or `TimeSeries` to 'unmask'
component_mask
Optionally, np.ndarray boolean mask of shape (n_components, 1) specifying which components were extracted
from `series`. If given, insert `vals` back into the columns of the original array. If not specified,
nothing is 'unmasked'.
Returns
-------
unmasked
`TimeSeries` (if `vals` is a `TimeSeries`) or `np.ndarray` (if `vals` is an `np.ndarray`) with those
components previously removed by `component_mask` now 'added back'.
"""
if component_mask is None:
unmasked = vals
else:
if not (
isinstance(component_mask, np.ndarray) and component_mask.dtype == bool
):
raise_log(
ValueError(
"If `component_mask` is given, must be a boolean np.ndarray`"
),
logger=logger,
)
if not series.width == len(component_mask):
raise_log(
ValueError(
"mismatch between number of components in `series` and length of `component_mask`"
),
logger=logger,
)
unmasked = series.all_values()
if isinstance(vals, TimeSeries):
unmasked[:, component_mask, :] = vals.all_values()
# Remove timepoints not present in transformed data:
unmasked = series.slice_intersect(vals).with_values(unmasked)
else:
unmasked[:, component_mask, :] = vals
return unmasked
[docs] @staticmethod
def stack_samples(vals: Union[np.ndarray, TimeSeries]) -> np.ndarray:
"""
Creates an array of shape `(n_timesteps * n_samples, n_components)` from
either a `TimeSeries` or the `array_values` of a `TimeSeries`.
Each column of the returned array corresponds to a component (dimension)
of the series and is formed by concatenating all of the samples associated
with that component together. More specifically, the `i`th column is formed
by concatenating
`[component_i_sample_1, component_i_sample_2, ..., component_i_sample_n]`.
Stacking is useful when implementing a transformation that applies the exact same
change to every timestep in the timeseries. In such cases, the samples of each
component can be stacked together into a single column, and the transformation can then be
applied to each column, thereby 'vectorising' the transformation over all samples of that component;
the `unstack_samples` method can then be used to reshape the output. For transformations that depend
on the `time_index` or the temporal ordering of the observations, stacking should *not* be employed.
Parameters
----------
vals
`Timeseries` or `np.ndarray` of shape `(n_timesteps, n_components, n_samples)` to be 'stacked'.
Returns
-------
stacked
`np.ndarray` of shape `(n_timesteps * n_samples, n_components)`, where the `i`th column is formed
by concatenating all of the samples of the `i`th component in `vals`.
"""
if isinstance(vals, TimeSeries):
vals = vals.all_values()
shape = vals.shape
new_shape = (shape[0] * shape[2], shape[1])
stacked = np.swapaxes(vals, 1, 2).reshape(new_shape)
return stacked
[docs] @staticmethod
def unstack_samples(
vals: np.ndarray,
n_timesteps: Optional[int] = None,
n_samples: Optional[int] = None,
series: Optional[TimeSeries] = None,
) -> np.ndarray:
"""
Reshapes the 2D array returned by `stack_samples` back into an array of shape
`(n_timesteps, n_components, n_samples)`; this 'undoes' the reshaping of
`stack_samples`. Either `n_components`, `n_samples`, or `series` must be specified.
Parameters
----------
vals
`np.ndarray` of shape `(n_timesteps * n_samples, n_components)` to be 'unstacked'.
n_timesteps
Optionally, the number of timesteps in the array originally passed to `stack_samples`.
Does *not* need to be provided if `series` is specified.
n_samples
Optionally, the number of samples in the array originally passed to `stack_samples`.
Does *not* need to be provided if `series` is specified.
series
Optionally, the `TimeSeries` object used to create `vals`; `n_samples` is inferred
from this.
Returns
-------
unstacked
`np.ndarray` of shape `(n_timesteps, n_components, n_samples)`.
"""
if series is not None:
n_samples = series.n_samples
else:
if all(x is None for x in [n_timesteps, n_samples]):
raise_log(
ValueError(
"Must specify either `n_timesteps`, `n_samples`, or `series`."
),
logger=logger,
)
n_components = vals.shape[-1]
if n_timesteps is not None:
reshaped_vals = vals.reshape(n_timesteps, -1, n_components)
else:
reshaped_vals = vals.reshape(-1, n_samples, n_components)
unstacked = np.swapaxes(reshaped_vals, 1, 2)
return unstacked
@property
def name(self):
"""Name of the data transformer."""
return self._name
def __str__(self):
return self._name
def __repr__(self):
return self.__str__()