Sebastian/faster tests

tests/integration/models/gpflux/test_predictions.py

@@ -31,10 +31,10 @@ def test_dgp_model_close_to_actuals(
     Ensure that DGP model fits well and predictions are close to actual output values.
     """
 
-    dataset_size = 100
-    num_inducing = 100
-    batch_size = 100
-    epochs = 2000
+    dataset_size = 50
+    num_inducing = 50
+    batch_size = 50
+    epochs = 500
 
     example_data = hartmann_6_dataset_function(dataset_size)
 
@@ -51,16 +51,16 @@ def test_dgp_model_close_to_actuals(
     model.optimize(example_data)
     predicted_means, _ = model.predict(example_data.query_points)
 
-    np.testing.assert_allclose(predicted_means, example_data.observations, atol=0.1, rtol=0.2)
+    np.testing.assert_allclose(predicted_means, example_data.observations, atol=0.2, rtol=0.2)
 
 
 @random_seed
 def test_dgp_model_close_to_simple_implementation(
     hartmann_6_dataset_function: Callable[[int], Dataset], depth: int, keras_float: None
 ) -> None:
-    dataset_size = 200
-    num_inducing = 100
-    batch_size = 100
+    dataset_size = 50
+    num_inducing = 50
+    batch_size = 50
     epochs = 500
     learning_rate = 0.01
 

tests/integration/test_bayesian_optimization.py

@@ -194,7 +194,7 @@ def build_model(data: Dataset) -> GaussianProcessRegression:
 @pytest.mark.parametrize(
     "num_steps, acquisition_rule",
     [
-        (5, DiscreteThompsonSampling(1000, 50)),
+        (1, DiscreteThompsonSampling(1000, 50)),
     ],
 )
 def test_two_layer_dgp_optimizer_finds_minima_of_michalewicz_function(
@@ -207,7 +207,7 @@ def test_two_layer_dgp_optimizer_finds_minima_of_michalewicz_function(
     search_space = Box(MICHALEWICZ_2_MINIMIZER[0] - 0.5, MICHALEWICZ_2_MINIMIZER[0] + 0.5)
 
     def build_model(data: Dataset) -> DeepGaussianProcess:
-        epochs = int(2e3)
+        epochs = int(4e2)
         batch_size = 100
 
         dgp = two_layer_dgp_model(data.query_points)
@@ -228,7 +228,7 @@ def scheduler(epoch: int, lr: float) -> float:
 
         return DeepGaussianProcess(model=dgp, optimizer=optimizer, fit_args=fit_args)
 
-    initial_query_points = search_space.sample(50)
+    initial_query_points = search_space.sample_sobol(20)
     observer = mk_observer(michalewicz, OBJECTIVE)
     initial_data = observer(initial_query_points)
     model = build_model(initial_data[OBJECTIVE])

tests/unit/models/gpflow/test_models.py

@@ -446,7 +446,7 @@ def test_gaussian_process_regression_optimize(
 
 @random_seed
 def test_variational_gaussian_process_predict() -> None:
-    x_observed = tf.constant(np.arange(100).reshape((-1, 1)), dtype=gpflow.default_float())
+    x_observed = tf.constant(np.arange(3).reshape((-1, 1)), dtype=gpflow.default_float())
     y_observed = _3x_plus_gaussian_noise(x_observed)
     model = VariationalGaussianProcess(vgp_model(x_observed, y_observed))
     internal_model = model.model
@@ -455,7 +455,7 @@ def test_variational_gaussian_process_predict() -> None:
         internal_model.training_loss_closure(),
         internal_model.trainable_variables,
     )
-    x_predict = tf.constant([[50.5]], gpflow.default_float())
+    x_predict = tf.constant([[1.5]], gpflow.default_float())
     mean, variance = model.predict(x_predict)
     mean_y, variance_y = model.predict_y(x_predict)
 

tests/util/models/gpflux/models.py

@@ -98,7 +98,7 @@ def trieste_deep_gaussian_process(
 ) -> Tuple[DeepGaussianProcess, Dict[str, Any]]:
     dgp = build_vanilla_deep_gp(query_points, num_layers=depth, num_inducing=num_inducing)
     if fix_noise:
-        dgp.likelihood_layer.likelihood.variance.assign(1e-4)
+        dgp.likelihood_layer.likelihood.variance.assign(1e-5)
         set_trainable(dgp.likelihood_layer, False)
     optimizer = tf.optimizers.Adam(learning_rate)