Oh I didn't realize you saw it as a simplification because all it does is remove the input size but then you need to create input placeholders. I thought it was meant to reimplement Keras on PyTorch.