Supporting N-dimensional deformable convolution
Original implementation:
DCNV2
DCNV4
Model example: please see here
- fp16. (cuda compute capability >= 7.0 && only gpu available, but when use fp16, it often occurs over/underflow problem. please use bfp16 instead.)
- bfp16. (cuda compute capability >= 8.0)
- torch.no_grad(). (enable gradient checkpointing)
- torch.autocast(). (enable pytorch AMP system)
- torch.autograd(). (just called function. do not need any additional implement.)
- 1d ~ 3d implement. (if you want to use over 3d, you just add dimension which you wanted TORCH_LIBRARY_IMPL in deform_conv.cpp, deform_conv_cpu.cpp and deform_conv_cuda.cu)
- channels last memory format.
- Deformable Convolution v1
- Deformable Convolution v2
- Deformable Convolution v3(InternImage)
- Deformable Convolution v4
- Pytorch
- CudaToolkit
- Python
- Ninja (Optional for fast build)
- OS: Windows10 with MSVC / Ubuntu 20.04.6 LTS with gcc
- GPU: NVIDA 3070TI 8G in Windows10 / NVIDIA A5000 24G in Ubuntu
- C++: std 17
- C: std 14
- Python: 3.10
- Pytorch: 2.1.0 / 2.4.0
- CudaToolkit: 11.8 / 12.4
(Optional) conda activate {your envirionment name}
python setup.py build installimport torch
import torch.nn as nn
from deform_conv import DeformConv2d
device = 'cuda' if torch.cuda.is_available() else 'cpu'
dtype = torch.bfloat16
inputs = torch.randn(2, 384, 64, 64).to(device)
deform_conv_layer = DeformConv2d(
in_channels=384,
out_channels=128,
kernel_size=3,
stride=1,
padding=1,
dilation=1,
groups=8,
deformable_groups_per_groups=2,
bias=True
).to(device)
with torch.autocast(device_type=device, dtype=dtype):
output = deform_conv_layer(inputs)
print(output)
"""
result is
1 2 3 4 ..... each values
... .... ........
device='cuda: 0', dtype=torch.bfloat16,
grad_fn=<CppNode<class DeformConvNdFunction<2>>>)
DeformConvNdFunction located at src/deform_conv.cpp
"""