Feel free to give it a go but I don't think you realize the mess underneath:- PlaidML was a startup that was bought by Intel a couple of years ago. It specialized on deep learning on OpenCL. Everyone was on Cuda and AMD was not really interested so Intel bought PlaidML.
- PlaidML had a Keras backend, which is a pure Python library with backends in Microsoft CNTK, Tensorflow, MxNet (Amazon/Apache) and Theano
- For Nim, I expect Keras is a no-go because that would mean shipping Python + dealing with CNTK/Tensorflow/MxNet which are huge in their own right.
- PlaidML has an alternative backend called Ngraph, from another startup Intel bought called Nervana which was focused on hardware.
- Ngraph is a compiler which can be used as a backend for Tensorflow, MxNet or declarative network in ONNX format (think the JSON of NN)
- Ngraph is built to be embedded in other applications, unlike PlaidML but besides ONNX we need a Nim frontend or still Tensorflow/MxNet.
- Plaidml requires Python + configure script + bazel + make for building, that is very complex.
i.e. I think you should start by getting a interface to Tensorflow or MxNet or build upon Nimtorch, write your own deep learning library with Ngraph as a backend or help on mine.