d1ac7b831ad36cd0e4bdd7980819f83208345148,gpflow/expectations.py,,_expectation,#Any#Any#Any#Any#Any#,512

Before Change

@dispatch(Gaussian, kernels.Sum, InducingPoints, kernels.Sum, InducingPoints)
@quadrature_fallback
def _expectation(p, kern1, feat1, kern2, feat2):
    if feat1 != feat2:
        raise NotImplementedError("Different features are not supported")

    feat = feat1
    crossexps = []

    for k1, k2 in it.product(kern1.kern_list, kern2.kern_list):
        if k1.on_separate_dims(k2):
            eKxz1 = _expectation(p, k1, feat, None, None)
            eKxz2 = _expectation(p, k2, feat, None, None)
            result = eKxz1[:, :, None] * eKxz2[:, None, :]
        else:
            result = _expectation(p, k1, feat, k2, feat)

        crossexps.append(result)

    return functools.reduce(tf.add, crossexps)


// Product

After Change

    
    crossexps = []

    if kern1 == kern2 and feat1 == feat2:  // avoid duplicate computation by using transposes
        for i, k1 in enumerate(kern1.kern_list):
            crossexps.append(expectation(p, (k1, feat1), (k1, feat1)))

            for k2 in kern1.kern_list[:i]:
                eKK = expectation(p, (k1, feat1), (k2, feat2))
                eKK += tf.matrix_transpose(eKK)
                crossexps.append(eKK)
    else:
        for k1, k2 in it.product(kern1.kern_list, kern2.kern_list):
            crossexps.append(expectation(p, (k1, feat1), (k2, feat2)))

    return functools.reduce(tf.add, crossexps)


// =================== Cross Kernel expectations (eK1zxK2xz) ===================

Instances