Open binhtranmcs opened 2 months ago
Any updates on this?
@pkufool @csukuangfj do have any insight into this
sorry for delay, if you could.find.a way to.display your diff conveniently it may be easier for.us.to look...
On Thursday, May 2, 2024, binhtranmcs @.***> wrote:
@pkufool https://github.com/pkufool @csukuangfj https://github.com/csukuangfj do have any insight into this
— Reply to this email directly, view it on GitHub https://github.com/k2-fsa/k2/issues/1281#issuecomment-2089814622, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAZFLO3HIGRUI2PYIOBJ223ZAHUYHAVCNFSM6AAAAABGSPQDLSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDAOBZHAYTINRSGI . You are receiving this because you are subscribed to this thread.Message ID: @.***>
@binhtranmcs Have you figured out the reasons? I think it is because python breaks a large batch into smaller sub batches when doing intersect, see _intersect_device
in the python code.
@binhtranmcs Have you figured out the reasons? I think it is because python breaks a large batch into smaller sub batches when doing intersect, see
_intersect_device
in the python code.
tks @pkufool, I will have a look at this
I tried hlg decoding using both
hlg_decode.cu
andhlg_decode.py
, modified them a bit to get nbest from the lattice. The python api seems fine but i got OOM when using c++ api. As i know, python api is still c++ under the hood. So I wonder if you all face the same issue as mine, that the c++ api use a lot more memory than python api, or I installed k2 incorrectly in some ways. Please help me with this.Model: librispeech conformer ctc Code: python: hlg_decode.py.txt c++: hlg_decode.cu.txt Audio (change ext to .wav): testvram.txt
Thanks in advance!