lobanov-m / dcn-v2-pytorch

Deformable Convolution v2 PyTorch Layer for old networks like CenterNet v1

Deformable Convolution v2 PyTorch Layer for old networks like CenterNet v1

Old repositories (1 year-old+) like CenterNet are using PyTorch < 1.0.0, so DCN extension could not be built because of API change.

This repo contains useful replacement of DCN extension with torchvision function.

So you can install this module and replace such code:

from .DCNv2.dcn_v2 import DCN

to:

from dcn_v2_pytorch import DCNv2 as DCN

and models will work without diving in problems of extension compilation.

Installation

pip install git+https://github.com/lobanov-m/dcn-v2-pytorch.git

Repository requires pytorch >= 1.8.1 and torchvision >= 0.9.0 (in this version torchvision started to support modulated deformable convolutions).

dcn-v2-pytorch's Issues

Training time

I would like to ask, why does the use of DCV_V2 instead of the Conv layer in the network reduce the training speed several times?

Just wanna say thanks for this code!!!

Thanks!!!