[Numpy-discussion] "Best" dot product along an axis?

[Skipper Seabold]

Going through some of the recent threads on similar problems, I'm
trying to discern which is "best."

I have X is T x i, theta is T x i x j. I want a T by j array that
contains X[t].T.dot(theta[t]) along axis 0.

Say,

T,i,j = 166,7,3

X = np.random.random((T,i))
theta = np.random.random((T,i,j))

# Using sum

F1 = (X[:,:,None] * theta).sum(1)

# Using a loop

F2 = np.zeros((T,j))
for t in range(T):
    for j_idx in range(j):
        for i_idx in range(7):
            F2[t,j_idx] += X[t,i_idx] * theta[t,i_idx,j_idx]

# One way with dot, keeps an extra T index
F3 = np.squeeze(np.dot(X[:,None,:],theta)).reshape(-1,3)[::167]

# the above is way fast than
F3_2 = np.dot(X,theta).reshape(-1,3)[::167]

# zipped arrays
F4 = np.asarray([np.dot(x,theta_i) for x,theta_i in zip(X,theta)])

It seems that F3 is the fastest way to do this given T,i,j, but it
seems inefficient to compute things I don't want. Does it make sense
that F3 is faster than F3_2? I'm using generic Ubuntu ATLAS on this
machine. Is there a more efficient way to do this? I'm not coming up
with one.

Skipper