Why should we use log_softmax when training?

[Marcovaldong]

@HawkAaron Why should we use log_softmax when training at here. When training the CTC model, there is no activation function such as softmax and log_softmax at here

Is there something different?