conscell

Changing operations order in __rtruediv__ and in __rsub__ prevents unnecessary __radd__ and __rmul__ calls.

Here is a vectorized implementation with PyTorch flavor built on top of NumPy / CuPy: https://github.com/conscell/ugrad

Fixes #59: Corrected the shape of `dbnmeani`, which should be [1, 64] instead of [64].

fixes #13 #45

1

Fixes #13 #45, where the `dhpreact` was not exactly matching the `hpreact.grad`.