torch.utils.ffi is deprecated

Question

torch.utils.ffi is deprecated

Opened this issue 5 years ago · 4 comments

Hi there,

thank you for the good structured repo here.
Unfortunately I am getting this error (relatively new to python):
Compiling crop_and_resize kernels by nvcc...
./make.sh: line 12: nvcc: command not found
Traceback (most recent call last):
File "build.py", line 3, in
from torch.utils.ffi import create_extension
File "/Users//opt/anaconda3/lib/python3.7/site-packages/torch/utils/ffi/init.py", line 1, in
raise ImportError("torch.utils.ffi is deprecated. Please use cpp extensions instead.")
ImportError: torch.utils.ffi is deprecated. Please use cpp extensions instead.
Compiling nms kernels by nvcc...
./make.sh: line 20: nvcc: command not found
Traceback (most recent call last):
File "build.py", line 3, in
from torch.utils.ffi import create_extension
File "/Users//opt/anaconda3/lib/python3.7/site-packages/torch/utils/ffi/init.py", line 1, in
raise ImportError("torch.utils.ffi is deprecated. Please use cpp extensions instead.")

What can I da about it?

Downgrading to torch 1.0 does not help (as intended in another issue). Also I dont want to downgrade to 0.4 since the version is too old.
Also I dont know how to change it to torch.utils.cpp_extension.

Regards

Answer 1 · 2019-12-10T18:05:20.000Z

You need nvcc installed as this compiles stuff for cuda. it is part of cudatoolkit. The other issue I am not sure. However the code was working with pytorch 1.0. You might try finding old source code for torch.utils.ffi and putting it in your python path. I remember having the error and doing something like that to fix it. I think you only need it to compile the c code. Alternatively find a cpp version of nms and cropresize. At the time nobody had done that though.

…

On Mon, 9 Dec 2019 at 10:22, 1seck ***@***.***> wrote: Hi there, thank you for the good structured repo here. Unfortunately I am getting this error (relatively new to python): Compiling crop_and_resize kernels by nvcc... ./make.sh: line 12: nvcc: command not found Traceback (most recent call last): File "build.py", line 3, in from torch.utils.ffi import create_extension File "/Users/ */opt/anaconda3/lib/python3.7/site-packages/torch/utils/ffi/init.py", line 1, in raise ImportError("torch.utils.ffi is deprecated. Please use cpp extensions instead.") ImportError: torch.utils.ffi is deprecated. Please use cpp extensions instead. Compiling nms kernels by nvcc... ./make.sh: line 20: nvcc: command not found Traceback (most recent call last): File "build.py", line 3, in from torch.utils.ffi import create_extension File "/Users/*/opt/anaconda3/lib/python3.7/site-packages/torch/utils/ffi/*init*.py", line 1, in raise ImportError("torch.utils.ffi is deprecated. Please use cpp extensions instead.") What can I da about it? Downgrading to torch 1.0 does not help (as intended in another issue). Also I dont want to downgrade to 0.4 since the version is too old. Also I dont know how to change it to torch.utils.cpp_extension. Regards — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1?email_source=notifications&email_token=AAJE32NCGD5UI2HZOFJFTD3QXYL5XA5CNFSM4JYGXRH2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4H7A256A>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAJE32PJ2LPHU3OBXJOU2ELQXYL5XANCNFSM4JYGXRHQ> .

Answer 2 · 2019-12-10T18:30:56.000Z

Hi Simon, thank you for the answer, I will try this although I am looking into facebooks implementation detectron2 right now.

…

Am 10.12.2019 um 19:05 schrieb simon ***@***.***>: You need nvcc installed as this compiles stuff for cuda. it is part of cudatoolkit. The other issue I am not sure. However the code was working with pytorch 1.0. You might try finding old source code for torch.utils.ffi and putting it in your python path. I remember having the error and doing something like that to fix it. I think you only need it to compile the c code. Alternatively find a cpp version of nms and cropresize. At the time nobody had done that though. On Mon, 9 Dec 2019 at 10:22, 1seck ***@***.***> wrote: > Hi there, > > thank you for the good structured repo here. > Unfortunately I am getting this error (relatively new to python): > Compiling crop_and_resize kernels by nvcc... > ./make.sh: line 12: nvcc: command not found > Traceback (most recent call last): > File "build.py", line 3, in > from torch.utils.ffi import create_extension > File "/Users/ > > > > > > > > */opt/anaconda3/lib/python3.7/site-packages/torch/utils/ffi/init.py", line > 1, in raise ImportError("torch.utils.ffi is deprecated. Please use cpp > extensions instead.") ImportError: torch.utils.ffi is deprecated. Please > use cpp extensions instead. Compiling nms kernels by nvcc... ./make.sh: > line 20: nvcc: command not found Traceback (most recent call last): File > "build.py", line 3, in from torch.utils.ffi import create_extension File > "/Users/*/opt/anaconda3/lib/python3.7/site-packages/torch/utils/ffi/*init*.py", > line 1, in > raise ImportError("torch.utils.ffi is deprecated. Please use cpp > extensions instead.") > > What can I da about it? > > Downgrading to torch 1.0 does not help (as intended in another issue). > Also I dont want to downgrade to 0.4 since the version is too old. > Also I dont know how to change it to torch.utils.cpp_extension. > > Regards > > — > You are receiving this because you are subscribed to this thread. > Reply to this email directly, view it on GitHub > <#1?email_source=notifications&email_token=AAJE32NCGD5UI2HZOFJFTD3QXYL5XA5CNFSM4JYGXRH2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4H7A256A>, > or unsubscribe > <https://github.com/notifications/unsubscribe-auth/AAJE32PJ2LPHU3OBXJOU2ELQXYL5XANCNFSM4JYGXRHQ> > . > — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1?email_source=notifications&email_token=AB6B2R2GJE5QQGRN6FNETMDQX7K6DA5CNFSM4JYGXRH2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEGQGASI#issuecomment-564158537>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB6B2R3MSN77KJYW4YP3R2TQX7K6DANCNFSM4JYGXRHQ>.

Answer 3 · 2019-12-10T18:38:44.000Z

I actually like fastai more than fb's stuff but am undecided since they are really good supported. But fastai has the 1-cycle super convergence training etc. and your implementation is just more straight forward than fbs with lots of variables and yaml files...

Answer 4 · 2019-12-23T19:27:34.000Z

its also part of torchvision now
(https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py)