Centernet: DCNv2 can not built on pytorch1.0

Created on 18 Apr 2019 · 9Comments · Source: xingyizhou/CenterNet

    from ._ext import dcn_v2 as _backend
ModuleNotFoundError: No module named 'models.networks.DCNv2._ext'

the DCNv2 can not be built on pytorch 1.0.

Traceback (most recent call last):
  File "build.py", line 3, in <module>
    from torch.utils.ffi import create_extension
  File "/usr/local/lib/python3.6/dist-packages/torch/utils/ffi/__init__.py", line 1, in <module>
    raise ImportError("torch.utils.ffi is deprecated. Please use cpp extensions instead.")
ImportError: torch.utils.ffi is deprecated. Please use cpp extensions instead.

Any fix on this?

Source

jinfagang

👍1

Most helpful comment

Hi,
Thanks for your interests again. You can check the Pytorch1.0 DCNv2 from the original repo https://github.com/CharlesShang/DCNv2 . Remember also to disable cudnn BN for pytorch 1.0. Our preliminary result on pytorch 1.0 is about 0.4 AP lower than pytorch 0.4.1 (for ctdet_coco_dla_1x). Not sure if this is due to randomness or internal difference between the two versions. Other experiments are not fully tested in pytorch 1.0. You can still test with our pretrained model on pytorch 1.0.

xingyizhou on 18 Apr 2019

👍5

All 9 comments

xingyizhou on 18 Apr 2019

👍5

@xingyizhou I already update DCNv2 to pytorch 1.0 now. It works fine, thanks for your reply.

BTW, how to disable cudnn BN in pytorch?

jinfagang on 18 Apr 2019

You can
vim ~/anaconda3/lib/python3.6/site-packages/torch/nn/functional.py
or vim ~/anaconda3/envs/ENV_NAME/lib/python3.6/site-packages/torch/nn/functional.py (if you use
conda environment)
And change torch.backend.cudnn.enabled to False in line 1623.

xingyizhou on 18 Apr 2019

👍4

@xingyizhou I already update DCNv2 to pytorch 1.0 now. It works fine, thanks for your reply.

BTW, how to disable cudnn BN in pytorch?

@jinfagang How would you update DCNv2 to pytorch 1.0

azhaoaigaga on 25 Jun 2019

@xingyizhou I already update DCNv2 to pytorch 1.0 now. It works fine, thanks for your reply.
BTW, how to disable cudnn BN in pytorch?

@jinfagang How would you update DCNv2 to pytorch 1.0

Download the new code from the original repo and replace them

VRCMF on 28 Jun 2019

👍1

@xingyizhou I already update DCNv2 to pytorch 1.0 now. It works fine, thanks for your reply.
BTW, how to disable cudnn BN in pytorch?

@jinfagang How would you update DCNv2 to pytorch 1.0

Download the new code from the original repo and replace them

TKS IT WORKS WELL

clemente0420 on 3 Aug 2019

I just git clone https://github.com/CharlesShang/DCNv2.git to my repo
and change my pytorch==1.0 for cuda 10
solve the problem
ImportError: torch.utils.ffi is deprecated. Please use cpp extensions instead.

clemente0420 on 3 Aug 2019

@clemente620 Please refer the question #7

VRCMF on 5 Aug 2019

Hi,
Thanks for your interests again. You can check the Pytorch1.0 DCNv2 from the original repo https://github.com/CharlesShang/DCNv2 . Remember also to disable cudnn BN for pytorch 1.0. Our preliminary result on pytorch 1.0 is about 0.4 AP lower than pytorch 0.4.1 (for ctdet_coco_dla_1x). Not sure if this is due to randomness or internal difference between the two versions. Other experiments are not fully tested in pytorch 1.0. You can still test with our pretrained model on pytorch 1.0.

Can this method solve his first question？from ._ext import dcn_v2 as _backend ModuleNotFoundError: No module named 'models.networks.DCNv2._ext'