A major bottleneck is the fact that we compute the info gain is python loop for each point separately. It' s conceivable that this can be vectorized, but probably this will require breaking up the kernel components of the GP and writing the vectorized computation by hand for each given kernel.
i.e. instead of
def info_gain(x_candidate,...)
X_all = np.concatenate([np.array([x_candidate]), meanX]).reshape(1 + n_samples, -1)
tocat = []
for gp in gps:
K_trans_all = gp.kernel_(X_all, gp.X_train_)
[info_gain(x,...) for x in scandetails.acqX]
do
info_gains(scandetails.acqX,...)
A major bottleneck is the fact that we compute the info gain is python loop for each point separately. It' s conceivable that this can be vectorized, but probably this will require breaking up the kernel components of the GP and writing the vectorized computation by hand for each given kernel.
i.e. instead of
do