Issue
On the official Tensorflow Lite website they show the support for GPU acceleration for Android and iOS but they don't say anything about support for Raspberry Pi.
Solution
If you want to use Rasperry Pi for training of neural nets within a mainstream framework, such as TensorFlow or PyTorch, you will likely not obtain GPU acceleration in a near term. The reason is that at least those two are mainstream GPU-centric, mostly NVidia, with AMD support being enabled little by little (here for TF and here for Torch).
Arguably, the most promising option for Pi is OpenCL support, however, by the time of writing and up to my knowledge, there is no OpenCL implementation for Raspberry Pi 4 yet.
Possibly with TensorFlow.js you can get something of the Raspberry Pi GPU, as far as it can use GPU by means of WebGL shaders. Therefore, a web-browser having WebGL support running on Pi may enable the GPU-accelerated training.
If you want to use Raspberry Pi for inference, it might be easier to benefit from the GPU acceleration.
- ncnn is a library using Vulkan API enabling GPU acceleration for Raspberry Pi 4.
- Beatmup library enables a GPU-accelerated inference on all Raspberry Pi models, not only 4. It also offers some conversion tools from Keras models, but is quite constrained in terms of implemented features.
Answered By - lnstadrum Answer Checked By - Timothy Miller (WPSolving Admin)