this is a custom C++/Cuda implementation of Correlation module, used e.g. in FlowNetC
This tutorial was used as a basis for implementation, as well as NVIDIA's cuda code
- Build and Install C++ and CUDA extensions by executing
python setup.py install
, - Benchmark C++ vs. CUDA by running
python benchmark.py {cpu, cuda}
, - Run gradient checks on the code by running
python grad_check.py --backend {cpu, cuda}
.
This module is expected to compile for Pytorch 2.1.0
.
Before installation please check compatibility of your GPU and CUDA (Compute Capability) nvidia docs. e.g RTX 6000 is using CC=8.9 so we are setting the environment variable to
export TORCH_CUDA_ARCH_LIST="8.9+PTX"
be reminded this module requires python3-dev
to compile C++ code, e.g. on Ubuntu run:
apt install python3-dev
this module is available on pip
pip install spatial-correlation-sampler
For a cpu-only version, you can install from source with
python setup_cpu.py install
This module needs compatible gcc version and CUDA to be compiled. Namely, CUDA 9.1 and below will need gcc5, while CUDA 9.2 and 10.0 will need gcc7 See this issue for more information
API has a few difference with NVIDIA's module
- output is now a 5D tensor, which reflects the shifts horizontal and vertical.
input (B x C x H x W) -> output (B x PatchH x PatchW x oH x oW)
- Output sizes
oH
andoW
are no longer dependant of patch size, but only of kernel size and padding - Patch size
patch_size
is now the whole patch, and not only the radii. stride1
is nowstride
andstride2
isdilation_patch
, which behave like dilated convolutions- equivalent
max_displacement
is thendilation_patch * (patch_size - 1) / 2
. dilation
is a new parameter, it acts the same way as dilated convolution regarding the correlation kernel- to get the right parameters for FlowNetC, you would have
kernel_size=1
patch_size=21,
stride=1,
padding=0,
dilation=1
dilation_patch=2
import torch
from spatial_correlation_sampler import SpatialCorrelationSampler,
device = "cuda"
batch_size = 1
channel = 1
H = 10
W = 10
dtype = torch.float32
input1 = torch.randint(1, 4, (batch_size, channel, H, W), dtype=dtype, device=device, requires_grad=True)
input2 = torch.randint_like(input1, 1, 4).requires_grad_(True)
#You can either use the function or the module. Note that the module doesn't contain any parameter tensor.
#function
out = spatial_correlation_sample(input1,
input2,
kernel_size=3,
patch_size=1,
stride=2,
padding=0,
dilation=2,
dilation_patch=1)
#module
correlation_sampler = SpatialCorrelationSampler(
kernel_size=3,
patch_size=1,
stride=2,
padding=0,
dilation=2,
dilation_patch=1)
out = correlation_sampler(input1, input2)
- default parameters are from
benchmark.py
, FlowNetC parameters are same as use inFlowNetC
with a batch size of 4, described in this paper, implemented here and here. - Feel free to file an issue to add entries to this with your hardware !
- See here for a benchmark script working with NVIDIA's code, and Pytorch.
- Benchmark are launched with environment variable
CUDA_LAUNCH_BLOCKING
set to1
. - Only
float32
is benchmarked. - FlowNetC correlation parameters where launched with the following command:
CUDA_LAUNCH_BLOCKING=1 python benchmark.py --scale ms -k1 --patch 21 -s1 -p0 --patch_dilation 2 -b4 --height 48 --width 64 -c256 cuda -d float
CUDA_LAUNCH_BLOCKING=1 python NV_correlation_benchmark.py --scale ms -k1 --patch 21 -s1 -p0 --patch_dilation 2 -b4 --height 48 --width 64 -c256
implementation | Correlation parameters | device | pass | min time | avg time |
---|---|---|---|---|---|
ours | default | 980 GTX | forward | 5.745 ms | 5.851 ms |
ours | default | 980 GTX | backward | 77.694 ms | 77.957 ms |
NVIDIA | default | 980 GTX | forward | 13.779 ms | 13.853 ms |
NVIDIA | default | 980 GTX | backward | 73.383 ms | 73.708 ms |
ours | FlowNetC | 980 GTX | forward | 26.102 ms | 26.179 ms |
ours | FlowNetC | 980 GTX | backward | 208.091 ms | 208.510 ms |
NVIDIA | FlowNetC | 980 GTX | forward | 35.363 ms | 35.550 ms |
NVIDIA | FlowNetC | 980 GTX | backward | 283.748 ms | 284.346 ms |
- The overhead of our implementation regarding
kernel_size
> 1 during backward needs some investigation, feel free to dive in the code to improve it ! - The backward pass of NVIDIA is not entirely correct when stride1 > 1 and kernel_size > 1, because not everything is computed, see here.
- No other implementation is avalaible on CPU.
- It is obviously not recommended to run it on CPU if you have a GPU.
Correlation parameters | device | pass | min time | avg time |
---|---|---|---|---|
default | E5-2630 v3 @ 2.40GHz | forward | 159.616 ms | 188.727 ms |
default | E5-2630 v3 @ 2.40GHz | backward | 282.641 ms | 294.194 ms |
FlowNetC | E5-2630 v3 @ 2.40GHz | forward | 2.138 s | 2.144 s |
FlowNetC | E5-2630 v3 @ 2.40GHz | backward | 7.006 s | 7.075 s |