A more practical solution is to train your network on other beefy CPU/GPU/TPUs and convert+run that on the Nano using TensorRT. TensorRT supports import of models in Tensorflow (UFF) and ONNX formats.