tensorflow 中 softmax_cross_entropy_with_logits 与 sparse_softmax_cross_entropy_with_logits 的区别

http://stackoverflow.com/questions/37312421/tensorflow-whats-the-difference-between-sparse-softmax-cross-entropy-with-logi

Having two different functions is a convenience, as they produce the same result.

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.