Closed jinserk closed 5 years ago
oh this is awesome! I'll pull this in for sure, i'll give it a look over ASAP. But yes will be deprecating this for the built in loss function into pytorch, installing warp-ctc has been a history of pain for people :D
It looks like the CI only checks cpu version pytorch, so I realize that the previous PR only works in gpu version. I've modified it to work on both of cpu and gpu versions. :D
Hmm. the CI still failed. I guess the CI's pytorch is still the current stable 0.4.1 or so, not supporting the cpp extensions of pytorch.
I'll keep this PR open till it becomes the stable GA release, that way I don't need to change the travis file to point to the nightly build!
@jinserk: This works with 1.0.0, have you tested?
Thanks Jinserk! I made a few changes/rebased down to one commit and got the tests working
It's my big pleasure! Merry Christmas and Happy New Year to you!
On Fri, Dec 14, 2018 at 8:16 AM Sean Naren notifications@github.com wrote:
Thanks Jinserk! I made a few changes/rebased down to one commit and got the tests working
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/SeanNaren/warp-ctc/pull/89#issuecomment-447338080, or mute the thread https://github.com/notifications/unsubscribe-auth/AAyPtmKjasneDyz2e-GY7IlQc3KTTg42ks5u47LNgaJpZM4Xsigw .
Hi @SeanNaren, I know that PyTorch 1.0.0 includes brand new CTCLoss, however, my old codes were not being trained well with it. I've discussed about the issue with @t-vi here, but it's still unclear that the bpc metric he mentioned could make the difference. So I modified the old ffi extension interface of this project to the new CppExtension and CUDAExtention of PyTorch 1.0.0 to keep it with my legacy codes. I also know that you are going to make this project deprecated. It's totally up to your decision to merge this PR. Thank you!