https://singularityfeed.com/optimizing-pytorch-model-inference-on-cpu/