Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time


Update December 27, 2020: a GPU version of PyTorchTOP has been released here. This branch demonstrates Style Transfer.

Installation on Windows (no OSX support)

Download LibTorch

From download, 1.4 (stable), Windows, LibTorch, C++/Java, CUDA 10.1

TouchDesigner Hack :/

I've tested TouchDesigner 2020.22080. From the place where you downloaded LibTorch, go to libtorch\lib. Then take libiomp5md.dll and overwrite the libiomp5md.dll for your TouchDesigner: C:\Program Files\Derivative\TouchDesigner099\bin.

Download PyTorchTOP.dll from the Releases page of this repo. Place it in this repo's Plugins folder. Copy the DLL files from libtorch\lib into Plugins too. Congrats! You're done and can open PyTorchTOP.toe! The remaining steps are for building PyTorchTOP.dll.


From NVIDIA, install CUDA 10.1, which will create C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1. Download cuDNN 7.6.5 for 10.1 and place the files into this folder too.


I've tested CMake 3.15.1. Inside the root of PyTorchTOP-cpumem:

mkdir build_release
cd build_release
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: 10.1
-- Caffe2: CUDA nvcc is: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v10.1/bin/nvcc.exe
-- Caffe2: CUDA toolkit directory: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v10.1
-- Caffe2: Header version is: 10.1
-- Found cuDNN: v7.6.5  (include: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v10.1/include, library: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v10.1/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-cpumem/build_release

If it works, you should end up with a Visual Studio Solution build_release\PyTorchTOP.sln. Open it and select the Release build. Press F5 to build the DLL and launch TouchDesigner.

The steps to build a debug-mode Visual Studio solution are similar. Instead of build_release, make a folder build_debug.

mkdir build_debug
cd build_debug
set DEBUG=1
cmake -DCMAKE_PREFIX_PATH=/path/to/libtorch-debug-version ..

Now you can build build_debug\PyTorchTOP.sln in Debug mode. You should copy the .pdb files from the libtorch-debug-version folder to this repo's Plugins folder in order to help with stack traces during debugging.

Neural Style Transfer

This project uses models that have been exported from Fast Neural Style. These are the steps to creating your own models. Install pytorch for python. Open neural_style\ Look for this line

output = style_model(content_image).cpu()

Before it, write these lines:

traced_script_module = torch.jit.trace(style_model, content_image)"")

To save a new, run:

python neural_style/ eval --content-image test640x360.jpeg --model saved_models/udnie.pth --output-image myoutput.png --cuda 1

Notice that you've provided a content image of a certain resolution, selected a model path, and enabled cuda. Because test640x360.jpeg is a 640x360 image, the newly created will work with 640x360 images in TouchDesigner. Export a model for each size resolution you need and rename as necessary. In TouchDesigner, select the model with the custom parameter Modelfilepath.

Extra notes

Use the channel mix TOP to swap your red channel and blue channel before sending to PyTorchTOP.

The Future

  • Fix the clumsiness of the RGBA-BGRA swapping in TouchDesigner.
  • Better handling of input resolution and output resolution. Can it be stored in the pt file?


PyTorch / libtorch in TouchDesigner based on the CPUMemoryTOP example.








No packages published