📚 The doc issue

The doc of nn.CrossEntropyLoss() explains about target tensor in a complex way as shown below. *It's difficult to understand:

So from my understanding and experiments, these simple explanations below should be added to the doc above. *It's easy to understand:

• The target tensor whose size is different from input tensor is treated as class indices.
• The target tensor whose size is same as input tensor is the class probabilities which should be between [0, 1].

And from what the doc says below and my experiments, when target tensor is treated as class indices, softmax() is used both for input and target tensor internally:

The target that this criterion expects should contain either:

• Class indices in the range ...
  ...
  Note that this case is equivalent to applying LogSoftmax on an input, followed by NLLLoss.

But when target tensor is treated as class probabilities, softmax() is used only for input tensor internally, that's why the example of target tensor as class indices in the doc doesn't use softmax() externally while the example of target tensor as class probabilities in the doc uses softmax() externally as shown below:

So, the doc also should say something like as shown below. *You also use the words class indices mode and class probabilities mode :

• softmax() is used internally for input tensor, both when target tensor is treated as class indices and class probabilities so you don't need to use softmax() externally.
• softmax() is used internally for target tensor only when target tensor is treated as class indices so you should use softmax() externally for target tensor when target tensor is treated as class probabilities.

Suggest a potential alternative/fix

No response

cc @svekars @brycebortree @tstatler @albanD @mruberry @jbschlosser @walterddr @mikaylagawarecki