OmniTrax - CUDA installation guide
Installing CUDA (Compute Unified Device Architecture) is widely regarded as one of the least fun but sadly most essential endeavours when diving into machine learning. Here, we will keep things as straight forward as possible and only install what is necessary to run OmniTrax with GPU support (which will result in an inference speed increase of at least one order of magnitude).
To match the dependencies of darknet and DeepLabCut-Live we are going to install CUDA 11.2 and cuDNN 8.1.0.
Download CUDA 11.2 and make sure you select the version matching your system.
Once you ensured you have picked the correct version for your system, download the Base Installer
Next, download cuDNN 8.1.0. Pay close attention to the selected version. In this case we need v8.1.0 for CUDA 11.0, 11.1, 11.2, specifically the cuDNN Library for Windows (x86).
If this is your first time using CUDA, you will need to first create an NVIDIA account. (Please don't ask me why.)
Once you have created an account, sign in and continue the download.
The files are relatively large (even more so, when unpacked), so make sure to select a drive with sufficient space. You can delete them, once the installation is completed.
Open the downloaded CUDA installer and chose a location to temporarily store the unpacked files (~5 GB). These temporary files will be removed once the installation is completed.
Unless you really know what you are doing, choosing the express installation option should be sufficient.
If your installation does not work as expected, you may need to install Visual Studio for your system. However, in most cases this is not strictly necessary.
Wait for the installation to complete.
When the installation has completed, doublecheck that CUDA 11.2 has been added to your Environment variables. Type Edit the system environment variables into the windows search bar and open the first suggested application.
Click on Environment Variables... and check whether CUDA_PATH as well as CUDA_PATH_V11_2 are listed.
All that is left to do now, is move the files contained in the downloaded cuDNN###.zip into the appropriate CUDA directory. Unpack the cuDNN###.zip and check its contents:
Simply drag these contents into your CUDA/v11.2 directory, which, by default, is located in C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.2
And that's (hopefully) it.
Restart your system before continuing the OmniTrax installation. If all has gone according to plan, you should now be able to select your GPU(s) as (a) compute device(s) within OmniTrax.
