"""
Functions to assess loss.
"""
import numpy as np
from typing import Callable, Union, List, Optional
import pandas as pd
import tensorflow as tf
from keras.backend import epsilon
from keras.losses import LossFunctionWrapper
from keras.utils import losses_utils
from tensorflow.python.util import dispatch
from ..utils import quantiles_handler
from .. import AutopycoinBaseClass
Yannotation = Union[tf.Tensor, pd.DataFrame, np.array, list]
@dispatch.add_dispatch_support
def smape(y_true: Yannotation, y_pred: Yannotation):
"""
Calculate the symmetric mean absolute percentage error between `y_true`and `y_pred`.
Parameters
----------
y_true : array, `dataframe`, list or `tensor`
Ground truth values. shape = `[batch_size, d0, .. dN]`.
y_pred : array, `dataframe`, list or `tensor`
The predicted values. shape = `[batch_size, d0, .. dN]`.
Returns
-------
error : `tensor`
Symetric mean absolute percentage error values. shape = `[batch_size, d0, .. dN-1]`.
Examples
--------
>>> from autopycoin.losses import smape
>>> import tensorflow as tf
>>> y_true = [[0., 1.], [0., 0.]]
>>> y_pred = [[1., 1.], [1., 0.]]
>>> smape(y_true, y_pred).numpy()
array([99.999985, 99.999985], dtype=float32)
"""
if not isinstance(y_pred, tf.RaggedTensor):
y_pred = tf.convert_to_tensor(y_pred)
y_true = tf.cast(y_true, dtype=y_pred.dtype)
error = tf.abs(y_true - y_pred) / (
tf.maximum(tf.abs(y_true), epsilon()) + tf.abs(y_pred)
)
return 200.0 * tf.reduce_mean(error, axis=-1)
def quantile_loss(
y_true: Yannotation, y_pred: Yannotation, quantiles: List[List[float]] = [[0.5]]
) -> tf.Tensor:
"""
Calculate the quantile loss function, summed across all quantile outputs.
Parameters
----------
y_true : array, `dataframe`, list or `tensor of shape (batch_size, d0, .. dN)`
Ground truth values.
y_pred : array, `dataframe`, list or `tensor of shape (batch_size, d0, .. dN)`
The predicted values.
quantiles : list[float]
The set of quantiles on which is calculated the quantile loss. The list is 1 dimension.
Returns
-------
error : `tensor of shape (batch_size, d0, .. dN)`
The error.
Examples
--------
>>> from autopycoin.losses import quantile_loss
>>> import tensorflow as tf
>>> y_true = [[[0.], [1.]], [[0.], [0.]]]
>>> y_pred = [[[1.], [1.]], [[1.], [0.]]]
>>> quantile_loss(y_true, y_pred, quantiles=[0.5]).numpy()
array([0.25, 0.25], dtype=float32)
"""
if not isinstance(y_pred, tf.RaggedTensor):
y_pred = tf.convert_to_tensor(y_pred)
y_true = tf.cast(y_true, dtype=y_pred.dtype)
quantiles = tf.convert_to_tensor(quantiles)
diff = tf.math.subtract(y_pred, y_true)
q_loss = tf.math.add(
quantiles * tf.clip_by_value(diff, 0.0, np.inf),
(1 - quantiles) * tf.clip_by_value(tf.math.negative(diff), 0.0, np.inf),
)
error = tf.reduce_mean(q_loss, axis=-2)
return tf.reduce_sum(error, axis=-1)
[docs]class SymetricMeanAbsolutePercentageError(LossFunctionWrapper, AutopycoinBaseClass):
"""
Calculate the symetric mean absolute percentage error between `y_true` and `y_pred`.
To avoid infinite error, we add epsilon value to zeros denominator.
This behavior can be modified by setting mask to True.
then, infinite instances are not taken into account in calculation.
Parameters
----------
reduction : `tf.keras.losses.Reduction, Optional`
Type of `tf.keras.losses.Reduction` to apply to
loss. Default value is `AUTO`. `AUTO` indicates that the reduction
option will be determined by the usage context. For almost all
cases this defaults to `SUM_OVER_BATCH_SIZE`. When used with
`tf.distribute.Strategy`, outside of built-in training lotf such
as `tf.keras` `compile` and `fit`, using `AUTO` or `SUM_OVER_BATCH_SIZE`
will raise an error. Please see this custom training [tutorial](
https://www.tensorflow.org/tutorials/distribute/custom_training)
for more details.
name : str, `Optional`
name for the op. Default to "smape".
Examples
--------
>>> import tensorflow as tf
>>> from autopycoin.losses import SymetricMeanAbsolutePercentageError
>>> y_true = [[0., 1.], [0., 0.]]
>>> y_pred = [[1., 1.], [1., 0.]]
>>> # Using 'auto'/'sum_over_batch_size' reduction type.
>>> smape = SymetricMeanAbsolutePercentageError()
>>> smape(y_true, y_pred).numpy()
99.999985
>>> # Calling with 'sample_weight'.
>>> smape(y_true, y_pred, sample_weight=[0.7, 0.3]).numpy()
49.999992
>>> # Using 'sum' reduction type.
>>> smape = SymetricMeanAbsolutePercentageError(
... reduction=tf.keras.losses.Reduction.SUM)
>>> smape(y_true, y_pred).numpy()
199.99997
>>> # Using 'none' reduction type.
>>> smape = SymetricMeanAbsolutePercentageError(
... reduction=tf.keras.losses.Reduction.NONE)
>>> smape(y_true, y_pred).numpy()
array([100., 100.], dtype=float32)
"""
def __init__(
self,
reduction: Optional[str] = losses_utils.ReductionV2.AUTO,
name: Optional[str] = "smape",
):
super().__init__(smape, name=name, reduction=reduction)
# TODO: ragged tensor
[docs]class QuantileLossError(LossFunctionWrapper, AutopycoinBaseClass):
"""
Calculate the quantile loss error between `y_true` and `y_pred`
across all examples.
Parameters
----------
quantiles : array, `dataframe`, list or `tensor of shape (batch_size, d0, .. dN)`
The set of quantiles on which is calculated the
quantile loss.
reduction : `tf.keras.losses.Reduction, Optional`
Type of `tf.keras.losses.Reduction`to apply to
loss. Default value is `AUTO`. `AUTO` indicates that the reduction
option will be determined by the usage context. For almost all
cases this defaults to `SUM_OVER_BATCH_SIZE`. When used with
`tf.distribute.Strategy`, outside of built-in training lotf such
as `tf.keras` `compile` and `fit`, using `AUTO` or `SUM_OVER_BATCH_SIZE`
will raise an error. Please see this custom training [tutorial](
https://www.tensorflow.org/tutorials/distribute/custom_training)
for more details.
name : str, `Optional`
name for the op. Default to 'quantile_loss'.
Returns
-------
error : `tensor of shape (batch_size, d0, .. dN)`
The error.
Attributes
----------
quantiles : list[float]
Examples
--------
>>> import tensorflow as tf
>>> from autopycoin.losses import QuantileLossError
>>> y_true = [[[0.], [1.]], [[0.], [0.]]]
>>> y_pred = [[[1.], [1.]], [[1.], [0.]]]
>>> # Using 'auto'/'sum_over_batch_size' reduction type.
>>> ql = QuantileLossError(quantiles=[0.5])
>>> ql(y_true, y_pred).numpy()
0.5
>>> # Calling with 'sample_weight'.
>>> ql(y_true, y_pred, sample_weight=[0.7, 0.3]).numpy()
0.25
>>> # Using 'AUTO' reduction type.
>>> ql = QuantileLossError(quantiles=[0.5],
... reduction=tf.keras.losses.Reduction.AUTO)
>>> ql(y_true, y_pred).numpy()
0.25
>>> # Using 'none' reduction type.
>>> ql = QuantileLossError(quantiles=[0.5],
... reduction=tf.keras.losses.Reduction.NONE)
>>> ql(y_true, y_pred).numpy()
array([0.25, 0.25], dtype=float32)
>>> # Using multiple quantiles.
>>> y_pred = [[[1.,1.,1.], [1.,1.,1.]], [[1.,1.,1.], [0.,0.,0.]]]
>>> ql = QuantileLossError(quantiles=[0.1, 0.5, 0.9])
>>> ql(y_true, y_pred).numpy()
1.5
"""
def __init__(
self,
quantiles: Union[
int, float, List[Union[int, float, List[Union[int, float]]]]
] = 0.5,
reduction: Optional[str] = losses_utils.ReductionV2.SUM,
name: Optional[str] = "q_loss",
**kwargs: dict
):
self.quantiles = quantiles_handler(quantiles)
super().__init__(
quantile_loss,
quantiles=self.quantiles,
name=name,
reduction=reduction,
**kwargs
)
# TODO: write doc, test and use LossFunctionWrapper, autpycoinBaseClass.
class LagError(tf.keras.losses.Loss):
"""1 lag error MSE. Often used in time series analysis.
Parameters
----------
lag : int
"""
def __init__(self, fn_loss: Callable, lag: int, **kwargs):
super().__init__(**kwargs)
self.lag = lag
self.fn_loss = fn_loss
def call(self, y_true, y_pred):
y_pred = tf.convert_to_tensor(y_pred)
y_true = tf.cast(y_true, y_pred.dtype)
y_pred = y_pred[:, self.lag :] - y_pred[:, : -self.lag]
y_true = y_true[:, self.lag :] - y_true[:, : -self.lag]
return self.fn_loss(y_true, y_pred)