sparse_softmax_cross_entropy_with_logits 与 softmax_cross_entropy_with_logits 的区别

The difference is simple:

• For sparse_softmax_cross_entropy_with_logits, labels must have the shape [batch_size] and the dtype int32 or int64. Each label is an int in range [0, num_classes-1].
• For softmax_cross_entropy_with_logits, labels must have the shape [batch_size, num_classes] and dtype float32 or float64.

Labels used in softmax_cross_entropy_with_logits are the one hot version of labels used in sparse_softmax_cross_entropy_with_logits.

Another tiny difference is that with sparse_softmax_cross_entropy_with_logits, you can give -1 as a label to have loss 0 on this label.