Hi @tom,
Could I ask you one more question? According to warpctc
, the probs is the linear scale without LogSoftmax, and the CTCLoss makes the probs into log scale inside. And nn.CTCLoss
requires LogSoftmax output, not the log scale of the output. Here the difference is “softmax” existence, so I’d like to know whether the LogSoftmax output is the correct one I have to give to the loss function.
Thanks!