Closed ysk24ok closed 3 years ago
Thanks @ysk24ok but out of curiosity have you tried using the native CTCLoss function? After further experiments seems it works well enough for me!
Yes, we actually use both warp CTC and native PyTorch CTC. The reason is 1) backward compatibility and 2) we used to observe strange behavior more for native PyTorch CTC when we used longer sequences (probably It might be gone, now.). Now, we're trying to move to espnet2 and we want to make sure especially for the backward compatibility. If espnet2 can guarantee performance comparable to espnet1 for various applications, we may completely switch to native PyTorch CTC in the near future.
Awesome, one less thing to hassle with in installation :) thanks guys!
Hello. Why not add a test for
forward
andbackward
ofCTCLoss
? I'm glad you review this PR.