def testGaussianProcessPrior(self):
batch_size = 3
input_dim = 4
output_dim = 5
features = tf.to_float(np.random.rand(batch_size, input_dim))
labels = tf.to_float(np.random.rand(batch_size, output_dim))
model = tf.keras.Sequential([
tf.keras.layers.Dense(2, activation=None),
bayes.GaussianProcess(output_dim),
])
outputs = model(features)
log_prob = outputs.distribution.log_prob(labels)
self.evaluate(tf.global_variables_initializer())
log_prob_val, outputs_val = self.evaluate([log_prob, outputs])
self.assertEqual(log_prob_val.shape, ())
self.assertLessEqual(log_prob_val, 0.)
self.assertEqual(outputs_val.shape, (batch_size, output_dim))
def testGaussianProcessPosterior(self):
train_batch_size = 3
test_batch_size = 2
input_dim = 4
output_dim = 5
features = tf.to_float(np.random.rand(train_batch_size, input_dim))
labels = tf.to_float(np.random.rand(train_batch_size, output_dim))
layer = bayes.GaussianProcess(output_dim,
conditional_inputs=features,
conditional_outputs=labels)
test_features = tf.to_float(np.random.rand(test_batch_size, input_dim))
test_labels = tf.to_float(np.random.rand(test_batch_size, output_dim))
test_outputs = layer(test_features)
test_nats = -test_outputs.distribution.log_prob(test_labels)
self.evaluate(tf.global_variables_initializer())
test_nats_val, outputs_val = self.evaluate([test_nats, test_outputs])
self.assertEqual(test_nats_val.shape, ())
self.assertGreaterEqual(test_nats_val, 0.)
self.assertEqual(outputs_val.shape, (test_batch_size, output_dim))
