szha commented on issue #7445: Using cuDNN for CTC Loss
URL:
https://github.com/apache/incubator-mxnet/issues/7445#issuecomment-380191490
cudnn integration for ctc requires action from cudnn team in improving the
API.
This
szha commented on issue #7445: Using cuDNN for CTC Loss
URL:
https://github.com/apache/incubator-mxnet/issues/7445#issuecomment-354379548
@apache/mxnet-committers: This issue has been inactive for the past 90 days.
It has no label and needs triage.
For general "how-to" questions,
szha commented on issue #7445: Using cuDNN for CTC Loss
URL:
https://github.com/apache/incubator-mxnet/issues/7445#issuecomment-332921106
The input length note being supported means that there's no way for me to
enforce consistency between CPU and GPU implementation. cudnn version also
szha commented on issue #7445: Using cuDNN for CTC Loss
URL:
https://github.com/apache/incubator-mxnet/issues/7445#issuecomment-332652715
Unfortunately I have to turn cudnn CTC off because of the API design. I have
requested API changes to nv people and hopefully we could incorporate that
szha commented on issue #7445: Using cuDNN for CTC Loss
URL:
https://github.com/apache/incubator-mxnet/issues/7445#issuecomment-322811854
Current implementation still includes the WarpCTC implementation in the GPU
version and only enables cudnn version when all input requirements are met,
szha commented on issue #7445: Using cuDNN for CTC Loss
URL:
https://github.com/apache/incubator-mxnet/issues/7445#issuecomment-322313189
Good catch. Let me reflect this in the PR as well.
This is an automated message from
szha commented on issue #7445: Using cuDNN for CTC Loss
URL:
https://github.com/apache/incubator-mxnet/issues/7445#issuecomment-322308517
@sbodenstein I agree. There is only one catch. It seems that the current
WarpCTC supports variable-length inputs whereas cudnn7 only has the intention
szha commented on issue #7445: Using cuDNN for CTC Loss
URL:
https://github.com/apache/incubator-mxnet/issues/7445#issuecomment-322064100
Thanks for raising this, @sbodenstein. I'm working on using the cudnn7
implementation of CTC for GPU.