Remove helpers from tests, skip GPU tests. (#1105)

* fix 1098: remove unneccesary helper functions.

* fix 1093: add hook to skip GPU tests without device.

tests/conftest.py

@@ -16,3 +16,16 @@ def set_seed():
 @pytest.fixture(scope="session", autouse=True)
 def set_default_tensor_type():
     torch.set_default_tensor_type("torch.FloatTensor")
+
+
+# Pytest hook to skip GPU tests if no devices are available.
+def pytest_collection_modifyitems(config, items):
+    """Skip GPU tests if no devices are available."""
+    gpu_device_available = (
+        torch.cuda.is_available() or torch.backends.mps.is_available()
+    )
+    if not gpu_device_available:
+        skip_gpu = pytest.mark.skip(reason="No devices available")
+        for item in items:
+            if "gpu" in item.keywords:
+                item.add_marker(skip_gpu)

tests/ensemble_test.py

@@ -13,11 +13,6 @@
     linear_gaussian,
     true_posterior_linear_gaussian_mvn_prior,
 )
-from sbi.utils.user_input_checks import (
-    check_sbi_inputs,
-    process_prior,
-    process_simulator,
-)
 from tests.test_utils import check_c2st, get_dkl_gaussian_prior
 
 
@@ -34,13 +29,8 @@ def test_import_before_deprecation():
         prior_cov = eye(2)
         prior = MultivariateNormal(loc=prior_mean, covariance_matrix=prior_cov)
 
-        prior, _, prior_returns_numpy = process_prior(prior)
-        simulator = process_simulator(
-            lambda theta: linear_gaussian(theta, likelihood_shift, likelihood_cov),
-            prior,
-            prior_returns_numpy,
-        )
-        check_sbi_inputs(simulator, prior)
+        def simulator(theta):
+            return linear_gaussian(theta, likelihood_shift, likelihood_cov)
 
         theta = prior.sample((num_simulations,))
         x = simulator(theta)
@@ -87,13 +77,8 @@ def test_c2st_posterior_ensemble_on_linearGaussian(inference_method, num_trials)
     )
     target_samples = gt_posterior.sample((num_samples,))
 
-    prior, _, prior_returns_numpy = process_prior(prior)
-    simulator = process_simulator(
-        lambda theta: linear_gaussian(theta, likelihood_shift, likelihood_cov),
-        prior,
-        prior_returns_numpy,
-    )
-    check_sbi_inputs(simulator, prior)
+    def simulator(theta):
+        return linear_gaussian(theta, likelihood_shift, likelihood_cov)
 
     # train ensemble components
     posteriors = []

tests/inference_on_device_test.py

@@ -34,9 +34,6 @@
 from sbi.utils.torchutils import BoxUniform, gpu_available, process_device
 from sbi.utils.user_input_checks import (
     check_embedding_net_device,
-    check_sbi_inputs,
-    process_prior,
-    process_simulator,
     validate_theta_and_x,
 )
 
@@ -282,13 +279,11 @@ def test_train_with_different_data_and_training_device(
     training_device = process_device(training_device)
 
     num_dim = 2
-    prior_ = BoxUniform(
+    num_simulations = 32
+    prior = BoxUniform(
         -torch.ones(num_dim), torch.ones(num_dim), device=training_device
     )
-
-    prior, _, prior_returns_numpy = process_prior(prior_)
-    simulator = process_simulator(diagonal_linear_gaussian, prior, prior_returns_numpy)
-    check_sbi_inputs(simulator, prior)
+    simulator = diagonal_linear_gaussian
 
     inference = inference_method(
         prior,
@@ -305,8 +300,9 @@ def test_train_with_different_data_and_training_device(
         device=training_device,
     )
 
-    theta, x = simulate_for_sbi(simulator, prior, 32)
-    theta, x = theta.to(data_device), x.to(data_device)
+    theta = prior.sample((num_simulations,))
+    x = simulator(theta).to(data_device)
+    theta = theta.to(data_device)
     x_o = torch.zeros(x.shape[1])
     inference = inference.append_simulations(theta, x, data_device=data_device)
 
@@ -453,10 +449,8 @@ def test_embedding_nets_integration_training_device(
 def test_nograd_after_inference_train(inference_method) -> None:
     """Test that no gradients are present after training."""
     num_dim = 2
-    prior_ = BoxUniform(-torch.ones(num_dim), torch.ones(num_dim))
-    prior, _, prior_returns_numpy = process_prior(prior_)
-    simulator = process_simulator(diagonal_linear_gaussian, prior, prior_returns_numpy)
-    check_sbi_inputs(simulator, prior)
+    prior = BoxUniform(-torch.ones(num_dim), torch.ones(num_dim))
+    simulator = diagonal_linear_gaussian
 
     inference = inference_method(
         prior,

tests/inference_with_NaN_simulator_test.py

@@ -23,7 +23,6 @@
 from sbi.utils.sbiutils import handle_invalid_x
 from sbi.utils.user_input_checks import (
     check_sbi_inputs,
-    process_prior,
     process_simulator,
 )
 
@@ -100,8 +99,7 @@ def linear_gaussian_nan(
         prior=prior,
     )
 
-    prior, _, prior_returns_numpy = process_prior(prior)
-    simulator = process_simulator(linear_gaussian_nan, prior, prior_returns_numpy)
+    simulator = process_simulator(linear_gaussian_nan, prior, False)
     check_sbi_inputs(simulator, prior)
     inference = method(prior=prior)
 
@@ -143,8 +141,7 @@ def linear_gaussian_nan(
         prior=prior,
     )
 
-    prior, _, prior_returns_numpy = process_prior(prior)
-    simulator = process_simulator(linear_gaussian_nan, prior, prior_returns_numpy)
+    simulator = process_simulator(linear_gaussian_nan, prior, False)
     check_sbi_inputs(simulator, prior)
     restriction_estimator = RestrictionEstimator(prior=prior)
     proposal = prior

tests/linearGaussian_mdn_test.py

@@ -20,11 +20,6 @@
     linear_gaussian,
     true_posterior_linear_gaussian_mvn_prior,
 )
-from sbi.utils.user_input_checks import (
-    check_sbi_inputs,
-    process_prior,
-    process_simulator,
-)
 from tests.test_utils import check_c2st
 
 
@@ -58,9 +53,6 @@ def mdn_inference_with_different_methods(method):
     def simulator(theta: Tensor) -> Tensor:
         return linear_gaussian(theta, likelihood_shift, likelihood_cov)
 
-    prior, _, prior_returns_numpy = process_prior(prior)
-    simulator = process_simulator(simulator, prior, prior_returns_numpy)
-    check_sbi_inputs(simulator, prior)
     inference = method(density_estimator="mdn")
 
     theta, x = simulate_for_sbi(simulator, prior, num_simulations)
@@ -108,9 +100,6 @@ def test_mdn_with_1D_uniform_prior():
     def simulator(theta: Tensor) -> Tensor:
         return linear_gaussian(theta, likelihood_shift, likelihood_cov)
 
-    prior, _, prior_returns_numpy = process_prior(prior)
-    simulator = process_simulator(simulator, prior, prior_returns_numpy)
-    check_sbi_inputs(simulator, prior)
     inference = SNPE(density_estimator="mdn")
 
     theta, x = simulate_for_sbi(simulator, prior, 100)

tests/linearGaussian_snle_test.py

@@ -27,9 +27,7 @@
 )
 from sbi.utils import BoxUniform
 from sbi.utils.user_input_checks import (
-    check_sbi_inputs,
     process_prior,
-    process_simulator,
 )
 
 from .test_utils import check_c2st, get_prob_outside_uniform_prior
@@ -58,9 +56,7 @@ def test_api_snle_multiple_trials_and_rounds_map(num_dim: int, prior_str: str):
     else:
         prior = BoxUniform(-2.0 * ones(num_dim), 2.0 * ones(num_dim))
 
-    prior, _, prior_returns_numpy = process_prior(prior)
-    simulator = process_simulator(diagonal_linear_gaussian, prior, prior_returns_numpy)
-    check_sbi_inputs(simulator, prior)
+    simulator = diagonal_linear_gaussian
     inference = SNLE(prior=prior, density_estimator="mdn", show_progress_bars=False)
 
     proposals = [prior]
@@ -112,18 +108,14 @@ def test_c2st_snl_on_linear_gaussian_different_dims(model_str="maf"):
         num_discarded_dims=discard_dims,
         num_samples=num_samples,
     )
-    prior, _, prior_returns_numpy = process_prior(prior)
-    simulator = process_simulator(
-        lambda theta: linear_gaussian(
+
+    def simulator(theta):
+        return linear_gaussian(
             theta,
             likelihood_shift,
             likelihood_cov,
             num_discarded_dims=discard_dims,
-        ),
-        prior,
-        prior_returns_numpy,
-    )
-    check_sbi_inputs(simulator, prior)
+        )
 
     density_estimator = likelihood_nn(model=model_str, num_transforms=3)
     inference = SNLE(density_estimator=density_estimator, show_progress_bars=False)
@@ -177,13 +169,8 @@ def test_c2st_and_map_snl_on_linearGaussian_different(
     else:
         prior = BoxUniform(-2.0 * ones(num_dim), 2.0 * ones(num_dim))
 
-    prior, _, prior_returns_numpy = process_prior(prior)
-    simulator = process_simulator(
-        lambda theta: linear_gaussian(theta, likelihood_shift, likelihood_cov),
-        prior,
-        prior_returns_numpy,
-    )
-    check_sbi_inputs(simulator, prior)
+    def simulator(theta):
+        return linear_gaussian(theta, likelihood_shift, likelihood_cov)
 
     density_estimator = likelihood_nn(model_str, num_transforms=3)
     inference = SNLE(density_estimator=density_estimator, show_progress_bars=False)
@@ -306,13 +293,8 @@ def test_c2st_multi_round_snl_on_linearGaussian(num_trials: int):
     )
     target_samples = gt_posterior.sample((num_samples,))
 
-    prior, _, prior_returns_numpy = process_prior(prior)
-    simulator = process_simulator(
-        lambda theta: linear_gaussian(theta, likelihood_shift, likelihood_cov),
-        prior,
-        prior_returns_numpy,
-    )
-    check_sbi_inputs(simulator, prior)
+    def simulator(theta):
+        return linear_gaussian(theta, likelihood_shift, likelihood_cov)
 
     inference = SNLE(show_progress_bars=False)
 
@@ -375,13 +357,8 @@ def test_c2st_multi_round_snl_on_linearGaussian_vi(num_trials: int):
     )
     target_samples = gt_posterior.sample((num_samples,))
 
-    prior, _, prior_returns_numpy = process_prior(prior)
-    simulator = process_simulator(
-        lambda theta: linear_gaussian(theta, likelihood_shift, likelihood_cov),
-        prior,
-        prior_returns_numpy,
-    )
-    check_sbi_inputs(simulator, prior)
+    def simulator(theta):
+        return linear_gaussian(theta, likelihood_shift, likelihood_cov)
 
     inference = SNLE(show_progress_bars=False)
 
@@ -487,11 +464,7 @@ def test_api_snl_sampling_methods(
     # Thus, we would not like to run, e.g., VI with all init_strategies, but only once
     # (namely with `init_strategy=proposal`).
     if sample_with == "mcmc" or init_strategy == "proposal":
-        prior, _, prior_returns_numpy = process_prior(prior)
-        simulator = process_simulator(
-            diagonal_linear_gaussian, prior, prior_returns_numpy
-        )
-        check_sbi_inputs(simulator, prior)
+        simulator = diagonal_linear_gaussian
 
         inference = SNLE(show_progress_bars=False)