treex.losses.MeanAbsoluteError
Computes the mean absolute errors between target and predictions.
loss = mean(abs(target - preds))
Usage:
target = jnp.array([[0.0, 1.0], [0.0, 0.0]])
preds = jnp.array([[1.0, 1.0], [1.0, 0.0]])
# Using 'auto'/'sum_over_batch_size' reduction type.
mae = tx.losses.MeanAbsoluteError()
assert mae(target, preds) == 0.5
# Calling with 'sample_weight'.
assert mae(target, preds, sample_weight=jnp.array([0.7, 0.3])) == 0.25
# Using 'sum' reduction type.
mae = tx.losses.MeanAbsoluteError(reduction=tx.losses.Reduction.SUM)
assert mae(target, preds) == 1.0
# Using 'none' reduction type.
mae = tx.losses.MeanAbsoluteError(reduction=tx.losses.Reduction.NONE)
assert list(mae(target, preds)) == [0.5, 0.5]
model = elegy.Model(
module_fn,
loss=tx.losses.MeanAbsoluteError(),
metrics=elegy.metrics.Mean(),
)
Source code in treex/losses/mean_absolute_error.py
class MeanAbsoluteError(Loss):
"""
Computes the mean absolute errors between target and predictions.
`loss = mean(abs(target - preds))`
Usage:
```python
target = jnp.array([[0.0, 1.0], [0.0, 0.0]])
preds = jnp.array([[1.0, 1.0], [1.0, 0.0]])
# Using 'auto'/'sum_over_batch_size' reduction type.
mae = tx.losses.MeanAbsoluteError()
assert mae(target, preds) == 0.5
# Calling with 'sample_weight'.
assert mae(target, preds, sample_weight=jnp.array([0.7, 0.3])) == 0.25
# Using 'sum' reduction type.
mae = tx.losses.MeanAbsoluteError(reduction=tx.losses.Reduction.SUM)
assert mae(target, preds) == 1.0
# Using 'none' reduction type.
mae = tx.losses.MeanAbsoluteError(reduction=tx.losses.Reduction.NONE)
assert list(mae(target, preds)) == [0.5, 0.5]
```
Usage with the Elegy API:
```python
model = elegy.Model(
module_fn,
loss=tx.losses.MeanAbsoluteError(),
metrics=elegy.metrics.Mean(),
)
```
"""
def __init__(
self,
reduction: tp.Optional[Reduction] = None,
weight: tp.Optional[float] = None,
on: tp.Optional[types.IndexLike] = None,
**kwargs
):
"""
Initializes `Mean` class.
Arguments:
reduction: (Optional) Type of `tx.losses.Reduction` to apply to
loss. Default value is `SUM_OVER_BATCH_SIZE`. For almost all cases
this defaults to `SUM_OVER_BATCH_SIZE`.
weight: Optional weight contribution for the total loss. Defaults to `1`.
on: A string or integer, or iterable of string or integers, that
indicate how to index/filter the `target` and `preds`
arguments before passing them to `call`. For example if `on = "a"` then
`target = target["a"]`. If `on` is an iterable
the structures will be indexed iteratively, for example if `on = ["a", 0, "b"]`
then `target = target["a"][0]["b"]`, same for `preds`. For more information
check out [Keras-like behavior](https://poets-ai.github.io/elegy/guides/modules-losses-metrics/#keras-like-behavior).
"""
return super().__init__(reduction=reduction, weight=weight, on=on, **kwargs)
def call(
self,
target: jnp.ndarray,
preds: jnp.ndarray,
sample_weight: tp.Optional[
jnp.ndarray
] = None, # not used, __call__ handles it, left for documentation purposes.
) -> jnp.ndarray:
"""
Invokes the `MeanAbsoluteError` instance.
Arguments:
target: Ground truth values. shape = `[batch_size, d0, .. dN]`, except
sparse loss functions such as sparse categorical crossentropy where
shape = `[batch_size, d0, .. dN-1]`
preds: The predicted values. shape = `[batch_size, d0, .. dN]`
sample_weight: Optional `sample_weight` acts as a
coefficient for the loss. If a scalar is provided, then the loss is
simply scaled by the given value. If `sample_weight` is a tensor of size
`[batch_size]`, then the total loss for each sample of the batch is
rescaled by the corresponding element in the `sample_weight` vector. If
the shape of `sample_weight` is `[batch_size, d0, .. dN-1]` (or can be
broadcasted to this shape), then each loss element of `preds` is scaled
by the corresponding value of `sample_weight`. (Note on`dN-1`: all loss
functions reduce by 1 dimension, usually axis=-1.)
Returns:
Weighted loss float `Tensor`. If `reduction` is `NONE`, this has
shape `[batch_size, d0, .. dN-1]`; otherwise, it is scalar. (Note `dN-1`
because all loss functions reduce by 1 dimension, usually axis=-1.)
Raises:
ValueError: If the shape of `sample_weight` is invalid.
"""
return mean_absolute_error(target, preds)
__init__(self, reduction=None, weight=None, on=None, **kwargs)
special
Initializes Mean class.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
reduction |
Optional[treex.losses.loss.Reduction] |
(Optional) Type of |
None |
weight |
Optional[float] |
Optional weight contribution for the total loss. Defaults to |
None |
on |
Union[str, int, Sequence[Union[str, int]]] |
A string or integer, or iterable of string or integers, that
indicate how to index/filter the |
None |
Source code in treex/losses/mean_absolute_error.py
def __init__(
self,
reduction: tp.Optional[Reduction] = None,
weight: tp.Optional[float] = None,
on: tp.Optional[types.IndexLike] = None,
**kwargs
):
"""
Initializes `Mean` class.
Arguments:
reduction: (Optional) Type of `tx.losses.Reduction` to apply to
loss. Default value is `SUM_OVER_BATCH_SIZE`. For almost all cases
this defaults to `SUM_OVER_BATCH_SIZE`.
weight: Optional weight contribution for the total loss. Defaults to `1`.
on: A string or integer, or iterable of string or integers, that
indicate how to index/filter the `target` and `preds`
arguments before passing them to `call`. For example if `on = "a"` then
`target = target["a"]`. If `on` is an iterable
the structures will be indexed iteratively, for example if `on = ["a", 0, "b"]`
then `target = target["a"][0]["b"]`, same for `preds`. For more information
check out [Keras-like behavior](https://poets-ai.github.io/elegy/guides/modules-losses-metrics/#keras-like-behavior).
"""
return super().__init__(reduction=reduction, weight=weight, on=on, **kwargs)
call(self, target, preds, sample_weight=None)
Invokes the MeanAbsoluteError instance.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
target |
ndarray |
Ground truth values. shape = |
required |
preds |
ndarray |
The predicted values. shape = |
required |
sample_weight |
Optional[jax._src.numpy.lax_numpy.ndarray] |
Optional |
None |
Returns:
| Type | Description |
|---|---|
ndarray |
Weighted loss float |
Exceptions:
| Type | Description |
|---|---|
ValueError |
If the shape of |
Source code in treex/losses/mean_absolute_error.py
def call(
self,
target: jnp.ndarray,
preds: jnp.ndarray,
sample_weight: tp.Optional[
jnp.ndarray
] = None, # not used, __call__ handles it, left for documentation purposes.
) -> jnp.ndarray:
"""
Invokes the `MeanAbsoluteError` instance.
Arguments:
target: Ground truth values. shape = `[batch_size, d0, .. dN]`, except
sparse loss functions such as sparse categorical crossentropy where
shape = `[batch_size, d0, .. dN-1]`
preds: The predicted values. shape = `[batch_size, d0, .. dN]`
sample_weight: Optional `sample_weight` acts as a
coefficient for the loss. If a scalar is provided, then the loss is
simply scaled by the given value. If `sample_weight` is a tensor of size
`[batch_size]`, then the total loss for each sample of the batch is
rescaled by the corresponding element in the `sample_weight` vector. If
the shape of `sample_weight` is `[batch_size, d0, .. dN-1]` (or can be
broadcasted to this shape), then each loss element of `preds` is scaled
by the corresponding value of `sample_weight`. (Note on`dN-1`: all loss
functions reduce by 1 dimension, usually axis=-1.)
Returns:
Weighted loss float `Tensor`. If `reduction` is `NONE`, this has
shape `[batch_size, d0, .. dN-1]`; otherwise, it is scalar. (Note `dN-1`
because all loss functions reduce by 1 dimension, usually axis=-1.)
Raises:
ValueError: If the shape of `sample_weight` is invalid.
"""
return mean_absolute_error(target, preds)