def _make_prior(std, prior_W, weight_shape):
"""Check/make prior weight distributions."""
if prior_W is None:
prior_W = norm_prior(weight_shape, std=std)
assert _is_dim(prior_W, weight_shape), \
"Prior inconsistent dimension!"
return prior_W
with tc.test_session():
tf.global_variables_initializer().run()
P = Phi.eval(feed_dict={x_: x})
for i in range(P.shape[0]):
p = P[i]
assert p.shape == (N, 2 * D)
# Check behaving properly with k(x, x) ~ 1.0
assert np.allclose((p**2).sum(axis=1), np.ones(N))
# Make sure we get a valid KL
kl = KL.eval() if isinstance(KL, tf.Tensor) else KL
assert kl >= 0
@pytest.mark.parametrize('dists', [
{'prior_W': norm_prior((4, 4, 3, D), 1.), 'prior_b': norm_prior((D,), 1.)},
{'post_W': norm_prior((4, 4, 3, D), 1.), 'post_b': norm_prior((D,), 1.)},
{'prior_W': norm_prior((4, 4, 3, D), 1.),
'post_W': norm_prior((4, 4, 3, D), 1.)}
])
def test_conv2d_distribution(dists, make_image_data):
"""Test initialising dense variational layers with distributions."""
x, _, X = make_image_data
S = 3
x_, X_ = _make_placeholders(x, S)
N, height, width, channels = x.shape
Phi, KL = ab.Conv2DVariational(filters=D, kernel_size=(4, 4), **dists)(X_)
tc = tf.test.TestCase()