Running model with torch 1.x

[SynStratos]

Hi,
could you suggest a solution to run the evaluation with pytorch 1? I would like to try it on Colab but installing torch 0.4.1 generates problem with Cuda so it is a bit problematic.

The main problem is related to the libs modules but I can't find a complete solution to it :(

[ziwei-zh]

Hi, I have not tested it on pytorch1, but you can refer to the following line in CE2P repo (https://github.com/liutinglt/CE2P):

to use Pytorch 1.0, just replace 'libs' with 'modules' in https://github.com/mapillary/inplace_abn, and rename it to 'libs'.

[SynStratos]

Yep, I tried to do so but doesn't solve the missing _wrap_function in the lib/_ext init :(
Do you have any idea what could be the correct substitute?

[ziwei-zh]

Could you show the errors?

[BLOO69]

Hi, I believe this is the same error.

Traceback (most recent call last):
File "evaluate.py", line 13, in
from networks.model import CorrPM_Model
File "/home/student/Documents/CorrPM/networks/model.py", line 7, in
from networks.non_local import NONLocalBlock2D
File "/home/student/Documents/CorrPM/networks/non_local.py", line 7, in
from libs import InPlaceABN, InPlaceABNSync
File "/home/student/Documents/CorrPM/libs/init.py", line 1, in
from .bn import ABN, InPlaceABN, InPlaceABNWrapper, InPlaceABNSync, InPlaceABNSyncWrapper
File "/home/student/Documents/CorrPM/libs/bn.py", line 15, in
from .functions import inplace_abn, inplace_abn_sync
File "/home/student/Documents/CorrPM/libs/functions.py", line 5, in
from . import _ext
File "/home/student/Documents/CorrPM/libs/_ext/init.py", line 2, in
from torch.utils.ffi import _wrap_function
File "/home/student/anaconda3/envs/CORR/lib/python3.6/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.

Dependecies

Pytorch == 1.7.0
CUDA == 10.2.89
CUDNN == 7.6.5_0

When i run nvcc: CUDA == 10.2.89
When i print torch.cuda.is_available() = True

Analysis

It seems that this error is specific for pytorch versions above 0.4.1. How would i be able to use this repo successfully?

If i try to install the dependencies based on environment.yaml:
Traceback (most recent call last):
File "evaluate.py", line 13, in
from networks.model import CorrPM_Model
File "/home/student/Documents/CorrPM/networks/model.py", line 7, in
from networks.non_local import NONLocalBlock2D
File "/home/student/Documents/CorrPM/networks/non_local.py", line 7, in
from libs import InPlaceABN, InPlaceABNSync
File "/home/student/Documents/CorrPM/libs/init.py", line 1, in
from .bn import ABN, InPlaceABN, InPlaceABNWrapper, InPlaceABNSync, InPlaceABNSyncWrapper
File "/home/student/Documents/CorrPM/libs/bn.py", line 15, in
from .functions import inplace_abn, inplace_abn_sync
File "/home/student/Documents/CorrPM/libs/functions.py", line 5, in
from . import _ext
File "/home/student/Documents/CorrPM/libs/_ext/init.py", line 3, in
from .__ext import lib as _lib, ffi as _ffi
ImportError: /home/student/Documents/CorrPM/libs/_ext/__ext.so: undefined symbol: __cudaRegisterFatBinaryEnd

This is most likely due to the CUDA = 9.0. I appreciate your help!

[xinzi2018]

i have the same questions

[Deepak2405]

Is this issue solved? I am having same error