I've made an example style transfer project with PyTorch https://github.com/DBraun/PyTorchTOP-cpumem This project copies from the GPU to the CPU before loading into PyTorch. In a private repo, I have the same style transfer example working but with a full CUDA pipeline (no CPU-GPU or GPU-CPU transfer). The benefits of the CUDA technique are larger for larger images. Furthermore, there's no usage of Spout In/Out, so I can guarantee no frames are dropped when running in non-real-time mode for an HD render.
I've also made an example project using TensorFlow with CUDA in a TOP (no CPU-GPU or GPU-CPU transfer).
Finally, I've coded a network entirely with the TensorFlow C++ API (no Python involved) and run it in a C++ TOP. This is pictured here https://www.instagram.com/p/CAYhW6mHnT6/ Going forward, it would be interesting to train a TensorFlow network interactively inside TouchDesigner. The TensorFlow C++ API makes this possible. There are also many pretrained and freely licensed TensorFlow models such as PoseNet, hand pose net, image segmentation models, and GANs which I'd like to run in TouchDesigner in the future.
To summarize, I have good knowledge of PyTorch, TensorFlow, and both of their Python/C++ APIs. I can create complex deep learning networks and want to keep learning. I would love to have a paid project in which I can further integrate deep learning frameworks with TouchDesigner and create amazing results. Please contact me with any questions!