import pytest
import tensorflow as tf
import gpflow.kernels as kernels
rng = np.random.RandomState(0)
class Datum:
num_data = 100
D = 100
X = rng.rand(num_data, D) * 100
kernel_list = [
kernels.Matern12(),
kernels.Matern32(),
kernels.Matern52(),
kernels.Exponential(),
kernels.Cosine()
]
@pytest.mark.parametrize('kernel', kernel_list)
def test_kernel_euclidean_distance(kernel):
'''
Tests output & gradients of kernels that are a function of the (scaled) euclidean distance
of the points. We test on a high dimensional space, which can generate very small distances
causing the scaled_square_dist to generate some negative values.
'''
K = kernel(Datum.X)
kernels.Constant,
# Stationary
kernels.SquaredExponential,
kernels.RationalQuadratic,
kernels.Exponential,
kernels.Matern12,
kernels.Matern32,
kernels.Matern52,
kernels.Cosine,
kernels.Linear,
kernels.Polynomial,
pytest.param(kernels.ArcCosine,
marks=pytest.mark.xfail), # broadcasting not implemented
kernels.Periodic,
lambda input_dim: kernels.White(input_dim) + kernels.Matern12(input_dim),
lambda input_dim: kernels.White(input_dim) * kernels.Matern12(input_dim),
]
@pytest.mark.parametrize("Kern", Kerns)
def test_broadcast_no_active_dims(Kern, session_tf):
S, N, M, D = 5, 4, 3, 2
X1 = tf.identity(np.random.randn(S, N, D))
X2 = tf.identity(np.random.randn(M, D))
kern = Kern(D) + gpflow.kernels.White(2)
compare_vs_map(X1, X2, kern, session_tf)
@pytest.mark.parametrize("Kern", [gpflow.kernels.SquaredExponential])
# Static kernels:
kernels.White,
kernels.Constant,
# Stationary kernels:
kernels.SquaredExponential,
kernels.RationalQuadratic,
kernels.Exponential,
kernels.Matern12,
kernels.Matern32,
kernels.Matern52,
# other kernels:
kernels.Cosine,
kernels.Linear,
kernels.Polynomial,
# sum and product kernels:
lambda: kernels.White() + kernels.Matern12(),
lambda: kernels.White() * kernels.Matern12(),
# Following kernels do not broadcast: see https://github.com/GPflow/GPflow/issues/1339
pytest.param(kernels.ArcCosine,
marks=pytest.mark.xfail), # broadcasting not implemented
pytest.param(kernels.Coregion,
marks=pytest.mark.xfail), # broadcasting not implemented
pytest.param(kernels.ChangePoints,
marks=pytest.mark.xfail), # broadcasting not implemented
pytest.param(kernels.Convolutional,
marks=pytest.mark.xfail), # broadcasting not implemented
]
def check_broadcasting(kernel):
S, N, M, D = 5, 4, 3, 2