tteepe / dcnv2

Deformable Convolutional Networks v2 (DCNv2) with Pytorch 1.8+ & JIT Compilation

Archived: DCNv2

This repo has been archived! But there is a great maintained alternative by MMCV.

Use this guide to generate your install version, i.e.:

pip install mmcv-full==1.7.0 -f https://download.openmmlab.com/mmcv/dist/cu117/torch1.13/index.html

Then you can change your imports to replace your DCN layer:

from mmcv.ops.deform_conv import DeformConv2d as DCN

from mmcv.ops.modulated_deform_conv import ModulatedDeformConv2dPack as DCN

---

Old: DCNv2 with Pytorch 1.8+ & JIT Compilation

Requirements

• PyTorch 1.8+
• Ninja build

pip install torch torchvision torchaudio

sudo apt-get install ninja-build

Test

cd tests
python test_cuda.py  # run examples and gradient check on gpu
python test_cpu.py   # run examples and gradient check on cpu 

Note

Now the master branch is for pytorch 1.x, you can switch back to pytorch 0.4 with,

git checkout pytorch_0.4

Known Issues:

• Gradient check w.r.t offset (solved)
• Backward is not reentrant (minor)

This is an adaption of the official Deformable-ConvNets.

Update: all gradient check passes with double precision.

Another issue is that it raises RuntimeError: Backward is not reentrant. However, the error is very small (<1e-7 for float <1e-15 for double), so it may not be a serious problem (?)

Please post an issue or PR if you have any comments.