This project demonstrates how to use OpenCV with CUDA modules and PyTorch/LibTorch in a TouchDesigner Custom Operator. Building this project is moderately complex and time-consuming, so read all the instructions to get a general overview before getting started.
PyTorchTOP uses Background Matting v2 as a foundation because their project is MIT-licensed (thank you!). Their research paper focuses on the speed and quality of the matting operation. For their quantitative results, they preprocess the images with OpenCV and don't add this computational cost to their numbers. It is ok to not preprocess the images but only if the frozen background image does not need to be transformed to match the live foreground images. This optional transformation is called homography in OpenCV. To see how the researchers do it, look at their HomographicAlignment class. PyTorchTOP has a toggle for using CUDA contrib modules from OpenCV to do the same homography efficiently on the GPU.
PyTorchTOP uses TorchScript models that can be downloaded from Background Matting V2. Download their "TorchScript" models and place them in this repo's models folder. More information on exporting models in this format is available here.
Install Python 3.7 or higher and make sure it's in your system path. This will help later when building OpenCV.
From https://pytorch.org/ download, 1.7.1 (stable), Windows, LibTorch, C++/Java, CUDA 11.0
From NVIDIA, install CUDA which will create C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.0. Check that this creates a system environment variable: CUDA_TOOLKIT_ROOT_DIR holding C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.0. Download the cuDNN version that corresponds to the CUDA version and place the files in the CUDA folder too.
Use git to clone OpenCV and OpenCV Contrib. It would be good to place them next to each other. Use CMake-GUI to select the source folder of OpenCV. In CMake-GUI, for the box Where is the source code, select the path to OpenCV. For Where to build the binaries, one suggestion is to create a build folder inside the OpenCV repo. You should permanently edit your system environment variables so that OpenCV_DIR holds this path. In CMake-GUI, press Configure. In the pop-up, select the latest version of Visual Studio, possibly Visual Studio 16 2019, and Finish. It will take less than a minute, and a table with lots of options will appear. Make the following modifications:
- Enable the checkbox
BUILD_opencv_world. - For
OPENCV_EXTRA_MODULES_PATH, enter the path tomodulesfolder in the opencv_contrib repository which you cloned earlier. - Enable the checkbox
WITH_CUDA.
Click Generate.
Check the entries in the table again. You should see a new entry for CUDA_TOOLKIT_ROOT_DIR, holding C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.0. If you don't, you may add the entry yourself, but this isn't guaranteed to work. You should probably check your CUDA installation, delete the CMake-GUI cache and restart the CMake-GUI steps.
If CUDA_TOOLKIT_ROOT_DIR is valid, then click Open Project. Build in Release mode. This will take between 1 to 2 hours. When it's complete, look for the "world" dll (opencv_world451.dll or equivalent) in the build's bin\Release folder and move it to %USERPROFILE%\Documents\Derivative\Plugins. If this folder doesn't already exist, create it.
To get started, we must first make sure several environment variables are correct. The process of installing CUDA and cuDNN should have already put them in your system path. If you have multiple versions of CUDA installed, you can temporarily modify your path to make sure the right one is found first. For example, since we're using CUDA 11.0, in a command window, do the following:
set CUDA_HOME=C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.0
set CUDA_PATH=C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.0
set PATH=%CUDA_HOME%;%CUDA_PATH%;%PATH%
This changes your system path but only in this command window. Furthermore, we need an environment variable OpenCV_DIR to be set to the path where OpenCV was built, such as C:/opencv/build. If you haven't set this variable, you can do it temporarily with the right path.
set OpenCV_DIR=C:/opencv/build
With the same window, inside the root of PyTorchTOP create a build folder:
mkdir build
cd build
cmake -DCMAKE_PREFIX_PATH=/path/to/libtorch ..
where /path/to/libtorch should be the full path to the unzipped LibTorch distribution. Expected output:
-- Selecting Windows SDK version 10.0.18362.0 to target Windows 10.0.
x64 architecture in use
-- Caffe2: CUDA detected: 11.0
-- Caffe2: CUDA nvcc is: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.0/bin/nvcc.exe
-- Caffe2: CUDA toolkit directory: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.0
-- Caffe2: Header version is: 11.0
-- Found cuDNN: v8.0.5 (include: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.0/include, library: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.0/lib/x64/cudnn.lib)
-- Autodetected CUDA architecture(s): 7.5
-- Added CUDA NVCC flags for: -gencode;arch=compute_75,code=sm_75
-- Configuring done
-- Generating done
-- Build files have been written to: /path/to/PyTorchTOP/build
If it works, you should end up with a Visual Studio solution build/PyTorchTOP.sln. Open it, select the Release build and press F5 to build the DLL and launch TouchDesigner. When you build, the newly built plugin and the necessary LibTorch DLLs will be copied to your C++ TouchDesigner plugin folder at %USERPROFILE%\Documents\Derivative\Plugins.
The steps to build a debug-mode Visual Studio solution are similar. You should download the debug version from PyTorch and use its path for CMake. Instead of build, make a folder build_debug.
mkdir build_debug
cd build_debug
set DEBUG=1
cmake -DCMAKE_PREFIX_PATH=/path/to/debug/libtorch ..
Now you can build build_debug\PyTorchTOP.sln in Debug mode. You can manually copy the .pdb files from the LibTorch folder to the Plugins folder in order to help with stack traces during debugging. You may also consider duplicating the steps for building OpenCV in debug mode.
When the project opens, check the Model path parameter on the plugin. If you select a 32-bit model, then the custom parameter for Bytes Per Model Input should be 4 because 4 bytes is 32 bits. If the model is 16-bit, then select 2. Then toggle the "Unload Plugin" parameter to load the plugin. The first input to the plugin is the live video and the second input is the static background image.
This example project has been tested with TouchDesigner 2020.28110 and libtorch with CUDA 11.0. At the time of writing this, TouchDesigner is using 10.1, so it's risky to use 11.0. Luckily it works. Choose your versions of the various components in this project at your own discretion. TouchDesigner's Release Notes often mention changes to their CUDA version.