mindspore.ops.NLLLoss

class mindspore.ops.NLLLoss(reduction='mean', ignore_index=- 100)

Gets the negative log likelihood loss between logits and labels.

The nll loss with \(reduction = none\) can be described as:

\[\ell(x, t)=L=\left\{l_{1}, \ldots, l_{N}\right\}^{\top}, \quad l_{n}=-w_{t_{n}} x_{n, t_{n}}, \quad w_{c}=\text { weight }[c] \cdot 1\]

where \(x\) is the logits, \(t\) is the labels, \(w\) is the weight, \(N\) is the batch size, \(c\) belonging to [0, C-1] is class index, where \(C\) is the number of classes.

If \(reduction \neq none\) (default 'mean' ), then

\[\begin{split}\ell(x, t)=\left\{\begin{array}{ll} \sum_{n=1}^{N} \frac{1}{\sum_{n=1}^{N} w_{t n}} l_{n}, & \text { if reduction }=\text { 'mean'; } \\ \sum_{n=1}^{N} l_{n}, & \text { if reduction }=\text { 'sum' } \end{array}\right.\end{split}\]

Warning

This is an experimental API that is subject to change or deletion.

Parameters

• reduction (str, optional) –

  Apply specific reduction method to the output: 'none' , 'mean' , 'sum' . Default: 'mean' .

  • 'none': no reduction will be applied.

  • 'mean': compute and return the weighted mean of elements in the output.

  • 'sum': the output elements will be summed.

• ignore_index (int, optional) – Specifies a target value that is ignored and does not contribute to the input gradient. Default: -100 .

Inputs:

• logits (Tensor) - Input logits, with shape \((N, C)\). Data type only supports float32 or float16 or bfloat16(only supported by Atlas A2 training series products).

• labels (Tensor) - Ground truth labels, with shape \((N,)\), where each value belong to \([0, C-1]\). Data type only supports uint8 or int32 or int64.

• weight (Tensor) - The rescaling weight to each class, with shape \((C,)\) and data type only supports float32 or float16 or bfloat16(only supported by Atlas A2 training series products). It should have the same data type as logits .

Returns

Tuple of 2 tensors composed with loss and total_weight.

• loss (Tensor) - When reduction is 'none' and logits is a 2D tensor, the loss shape is \((N,)\). Otherwise, the loss is a scalar. The data type is the same with that of logits.

• total_weight (Tensor) - The total_weight is a scalar. The data type is the same with that of weight.

Supported Platforms:

Ascend GPU CPU

Examples

>>> import numpy as np
>>> from mindspore import Tensor, ops
>>> logits = Tensor(np.array([[0.5488135, 0.71518934],
...                           [0.60276335, 0.5448832],
...                           [0.4236548, 0.6458941]]).astype(np.float32))
>>> labels = Tensor(np.array([0, 0, 0]).astype(np.int32))
>>> weight = Tensor(np.array([0.3834415, 0.79172504]).astype(np.float32))
>>> nll_loss = ops.NLLLoss(reduction="mean")
>>> loss, weight = nll_loss(logits, labels, weight)
>>> print(loss)
-0.52507716
>>> print(weight)
1.1503246