search

SeanNaren / warp-ctc

Pytorch Bindings for warp-ctc
Apache License 2.0
756 stars 271 forks source link

c10/cuda/CUDAGuard.h not found ？？compilation terminated #109

Closed OKhyc closed 5 years ago

OKhyc commented 5 years ago

我运行到 python setup.py install 的爆出来的 错误，怎么解决呢

SeanNaren commented 5 years ago

A recent commit i merged doesn't work for 1.0, will revert and create a new PR till 1.1 is stable

SeanNaren commented 5 years ago

Pull from master

Adhders commented 5 years ago

A recent commit i merged doesn't work for 1.0, will revert and create a new PR till 1.1 is stable

same error for warp-ctc-1.1_fix

alchemi5t commented 5 years ago

I've pulled from master and have "ImportError: torch.utils.ffi is deprecated. Please use cpp extensions instead." now.

SeanNaren commented 5 years ago

@alchemi5t are you running off branch pytorch_bindings?

alchemi5t commented 5 years ago

@SeanNaren No. Anyway, I needed it for running your deepspeech implementation on pytorch 1.0.1 . I just found out that I can replace warp-ctc with the torch.nn.CTCLoss. Hope that works.

SeanNaren commented 5 years ago

it should work if you use the branch pytorch_bindings. Just FYI as well, the torch.nn.CTCLoss is known to have a few kinks to work out, so if you start seeing weird behaviour and losses, it might be the loss function!

alchemi5t commented 5 years ago

@SeanNaren what do you mean by branch pytorch_bindings. I understand there's a folder named that which had the installation step, i was unable install because of "c10/cuda/CUDAGuard.h not found" error. so i skipped that and decided to use the system pytorch loss.

rusiaaman commented 5 years ago

@alchemi5t are you running off branch pytorch_bindings?

I got the same error on master branch: ImportError: torch.utils.ffi is deprecated. Please use cpp extensions instead.

On pytorch_bindings branch I got the original error: src/binding.cpp:10:33: fatal error: c10/cuda/CUDAGuard.h: No such file or directory

I am on pytorch 1.0.0 and CUDA 10. It works for pytorch 0.4.0 but it doesn't support CUDA 10. And Tesla T4 doesn't support below CUDA 10 (I think) so I can't downgrade. Any help? Do you have any update on torch.nn.CTCLoss, and if it works good enough for most purpose in deepspeech.pytorch?

EDIT: It worked on torch==1.1.0 but not on torch==1.0.0.

shuaijiang commented 4 years ago

You can try pip install warpctc-pytorch10-cuda90==0.1.3 instead of python setup.py install