Losses: keras_proba.losses

Negative log-likelihoods for the probabilistic regression heads.

Loss classes

GaussianNLL

GaussianNLL(name='gaussian_nll', reduction='sum_over_batch_size', dtype=None)

Bases: keras.src.losses.loss.Loss

Gaussian negative log-likelihood loss for GaussianHead outputs.

Example:

model = keras.Sequential([keras.Input((8,)), keras.layers.Dense(64, "relu"), GaussianHead(1)])
model.compile(optimizer="adam", loss=GaussianNLL())

__doc__ class-attribute

__doc__ = 'Gaussian negative log-likelihood loss for `GaussianHead` outputs.\n\n    Example:\n\n    ```python\n    model = keras.Sequential([keras.Input((8,)), keras.layers.Dense(64, "relu"), GaussianHead(1)])\n    model.compile(optimizer="adam", loss=GaussianNLL())\n    ```\n    '

str(object='') -> str str(bytes_or_buffer[, encoding[, errors]]) -> str

Create a new string object from the given object. If encoding or errors is specified, then the object must expose a data buffer that will be decoded using the given encoding and error handler. Otherwise, returns the result of object.str() (if defined) or repr(object). encoding defaults to sys.getdefaultencoding(). errors defaults to 'strict'.

__module__ class-attribute

__module__ = 'keras_proba.layers.gaussian_head'

str(object='') -> str str(bytes_or_buffer[, encoding[, errors]]) -> str

Create a new string object from the given object. If encoding or errors is specified, then the object must expose a data buffer that will be decoded using the given encoding and error handler. Otherwise, returns the result of object.str() (if defined) or repr(object). encoding defaults to sys.getdefaultencoding(). errors defaults to 'strict'.

MixtureGaussianNLL

MixtureGaussianNLL(units=None, name='mixture_gaussian_nll', reduction='sum_over_batch_size', dtype=None)

Bases: keras.src.losses.loss.Loss

Negative log-likelihood loss for MixtureGaussianHead outputs.

Example:

model = keras.Sequential([
    keras.Input((1,)),
    keras.layers.Dense(64, activation="relu"),
    MixtureGaussianHead(num_components=5),
])
model.compile(optimizer="adam", loss=MixtureGaussianNLL())

Parameters:

Name	Type	Description	Default
units		Number of regression targets. Inferred from the targets if None; set it when targets of shape (batch, units) could be confused with (batch,) targets for units=1.	None
name		Name of the loss.	'mixture_gaussian_nll'
reduction		Keras loss reduction.	'sum_over_batch_size'
dtype		Dtype of the loss computation.	None

__doc__ class-attribute

__doc__ = 'Negative log-likelihood loss for `MixtureGaussianHead` outputs.\n\n    Example:\n\n    ```python\n    model = keras.Sequential([\n        keras.Input((1,)),\n        keras.layers.Dense(64, activation="relu"),\n        MixtureGaussianHead(num_components=5),\n    ])\n    model.compile(optimizer="adam", loss=MixtureGaussianNLL())\n    ```\n\n    Args:\n        units: Number of regression targets. Inferred from the targets if\n            `None`; set it when targets of shape `(batch, units)` could be\n            confused with `(batch,)` targets for `units=1`.\n        name: Name of the loss.\n        reduction: Keras loss reduction.\n        dtype: Dtype of the loss computation.\n    '

str(object='') -> str str(bytes_or_buffer[, encoding[, errors]]) -> str

Create a new string object from the given object. If encoding or errors is specified, then the object must expose a data buffer that will be decoded using the given encoding and error handler. Otherwise, returns the result of object.str() (if defined) or repr(object). encoding defaults to sys.getdefaultencoding(). errors defaults to 'strict'.

__module__ class-attribute

__module__ = 'keras_proba.layers.mixture_gaussian_head'

str(object='') -> str str(bytes_or_buffer[, encoding[, errors]]) -> str

Create a new string object from the given object. If encoding or errors is specified, then the object must expose a data buffer that will be decoded using the given encoding and error handler. Otherwise, returns the result of object.str() (if defined) or repr(object). encoding defaults to sys.getdefaultencoding(). errors defaults to 'strict'.

Functions

gaussian_nll

gaussian_nll(y_true, y_pred)

Gaussian negative log-likelihood of y_true under a GaussianHead output.

Parameters:

Name	Type	Description	Default
y_true		Targets of shape (..., units), or (batch,) when units=1.	required
y_pred		GaussianHead output of shape (..., 2 * units).	required

Returns:

Type	Description
	Negative log-likelihood averaged over the last axis, shape (...).

mixture_gaussian_nll

mixture_gaussian_nll(y_true, y_pred, units=None)

Negative log-likelihood of y_true under a MixtureGaussianHead output.

Parameters:

Name	Type	Description	Default
y_true		Targets of shape (..., units), or (batch,) when units=1.	required
y_pred		MixtureGaussianHead output of shape (..., K * (1 + 2 * units)).	required
units		Number of regression targets. Inferred from y_true if None.	None

Returns:

Type	Description
	Negative log-likelihood of the joint units-dimensional target, shape
	(...). Unlike gaussian_nll, it is summed (not averaged) over units,
	as for tfd.Independent.