Add to MinPy’s operators¶
Since MinPy inherits nearly all operators from NumPy, you could expect to use a NumPy operator without additional instructions. The MinPy operator might actually run the NumPy implementation with the same name or corresponding MXNet implementation, depending on the policy.
Apart from the NumPy experience, all other automatic switching details should not worry users. But if you want to define a new operator, or implement some missing gradient for operators from NumPy, there are a few steps to follow.
1. Define the gradient of the NumPy version of the operator in
minpy/array_varians/numpy/numpy_core.py. There are already a bunch of operators
with patterns to follow. In essence you need to define a second order function
that takes all inputs to the original function, and the gradient passed down as well.
Your job is to have the function pass the gradient through the operator.
For example, in the
np.dot case, two gradient functions are
defined. One is for the first argument, same for the second. It is
argnum to the
def_grad method. The gradient itself
is a second order function which takes first the result of the
computation and orginal inputs, then the gradient. So it is
ans, a, b: lambda g: np.dot(g, b.T) in this case.
2. Follow the same steps for the MXNet version in
minpy/array_variants/mxnet/mxnet_core.py. Sometimes the MXNet
operator has a slightly different name, then you should
it under function
You could technically define only NumPy version or MXNet version. The policy is smart enought to fall back. But there will be performance penalty copying data back and forth.