Improve IDW

[patel-zeel]

The following lines in IDW might slow down IDW since it is a for loop over all examples. Can we use a clever math/code trick to get rid of them?

polire/polire/idw/idw.py

Lines 85 to 91 in 4b9360f

	# if point is from train data, ground truth must not change
	for i in range(X.shape[0]):
	mask = np.equal(X[i], self.X).all(axis=1)
	if mask.any():
	result[i] = (self.y * mask).sum()
	return result

A relevant trick is used in causal attention. Consider the following sentence:
I play cricket

A self-attention matrix will be 3x3 for this. Let's populate some values:

.	I	play	cricket
I	a11	0	0
play	a21	a22	0
cricket	a31	a32	a33

To make sure, softmax works correctly, we replace zeros with -inf.

.	I	play	cricket
I	a11	-inf	-inf
play	a21	a22	-inf
cricket	a31	a32	a33

Softmax output:

.	I	play	cricket
I	1	0	0
play	exp(a21)/row_sum	exp(a22)/row_sum	0
cricket	exp(a31)/row_sum	exp(a32)/row_sum	exp(a33)/row_sum