mindformers.core.CrossEntropyLoss

class mindformers.core.CrossEntropyLoss(parallel_config=default_dpmp_config, check_for_nan_in_loss_and_grad=False, calculate_per_token_loss=False, seq_split_num=1, **kwargs)

Calculate the cross entropy loss.

CrossEntropyLoss supports two different types of targets:

• Class indices (int), where the range of values is $[0,C)$ with $C$ being the number of classes. When reduction is set to 'none', the cross-entropy loss is computed as follows:

  $$\ell (x,y)=L=\{l_1,\dots ,l_N{\}}^{\mathrm{\top }},l_n=-w_{y_n}\log \frac{\exp (x_{n,y_n})}{\sum _{c=1}^C\exp (x_{n,c})}\cdot \mathbb{1}\{y_n\ne \text{ignore\_index}\}$$

  where $x$ denotes the predicted values, $t$ denotes the target values, $w$ denotes the weights, and $N$ is the batch size. The index $c$ ranges from [0, C-1], representing the class indices, where $C$ is the number of classes.

  If reduction is not set to 'none' (the default is 'mean'), the loss is computed as:

  $$\begin{array}{r}\ell (x,y)=\{\begin{array}{ll}\sum _{n=1}^N\frac{1}{\sum _{n=1}^N{w}_{y_n}\cdot \mathbb{1}\{y_n\ne \text{ignore\_index}\}}{l}_n,& \text{if reduction}=\text{'mean',}\\ \sum _{n=1}^N{l}_n,& \text{if reduction}=\text{'sum'.}\end{array}\end{array}$$

• Class probabilities (float), used when the target is a probability distribution over multiple class labels. When reduction is set to 'none', the cross-entropy loss is computed as follows:

  $$\ell (x,y)=L=\{l_1,\dots ,l_N{\}}^{\mathrm{\top }},l_n=-\sum _{c=1}^C{w}_c\log \frac{\exp (x_{n,c})}{\sum _{i=1}^C\exp (x_{n,i})}{y}_{n,c}$$

  where $x$ denotes the predicted values, $t$ denotes the target values, $w$ denotes the weights, and $N$ is the batch size. The index $c$ ranges from [0, C-1], representing the class indices, where $C$ is the number of classes.

  If reduction is not set to 'none' (the default is 'mean'), the loss is computed as:

  $$\begin{array}{r}\ell (x,y)=\{\begin{array}{ll}\frac{\sum _{n=1}^N{l}_n}{N},& \text{if reduction}=\text{'mean',}\\ \sum _{n=1}^N{l}_n,& \text{if reduction}=\text{'sum'.}\end{array}\end{array}$$

Parameters

• parallel_config (mindformers.modules.OpParallelConfig, optional) – The parallel configuration. Default: default_dpmp_config.

• check_for_nan_in_loss_and_grad (bool, optional) – Whether to print local loss. Default: False.

• calculate_per_token_loss (bool, optional) – Whether to use Megatron loss. Default: False.

• seq_split_num (int, optional) – Sequence split number in sequence pipeline parallel mode. Default: 1.

Inputs:

• logits (Tensor) - Tensor of shape (N, C). Data type must be float16 or float32. The output logits of the backbone.

• label (Tensor) - Tensor of shape (N, ). The ground truth label of the sample.

• input_mask (Tensor) - Tensor of shape (N, ). input_mask indicates whether there are padded inputs and for padded inputs it will not be counted into loss.

Returns

Tensor, the computed cross entropy loss value.

Examples

>>> import numpy as np
>>> from mindspore import dtype as mstype
>>> from mindspore import Tensor
>>> from mindformers.core import CrossEntropyLoss
>>> loss = CrossEntropyLoss()
>>> logits = Tensor(np.array([[3, 5, 6, 9, 12, 33, 42, 12, 32, 72]]), mstype.float32)
>>> labels_np = np.array([1]).astype(np.int32)
>>> input_mask = Tensor(np.ones(1).astype(np.float32))
>>> labels = Tensor(labels_np)
>>> output = loss(logits, labels, input_mask)
>>> output.shape
(1,)