Add tests for evidence calculation sample batching.

Author: alicjapolanska

tests/test_evidence.py

@@ -16,6 +16,16 @@
 models_to_test_1 = [sphere_1000D, real_nvp_2D, spline_4D]
 models_to_test_2 = [sphere_2D, real_nvp_2D, spline_4D]
 
+models_to_test_2D = [
+    sphere_2D,
+    real_nvp_2D,
+    md.RealNVPModel(2, standardize=True),
+    md.RQSplineModel(2),
+    md.RQSplineModel(2, standardize=True),
+]
+
+chain_batching_options = [None, 2, 10]
+
 
 @pytest.mark.parametrize("model", models_to_test_1)
 def test_constructor(model):
@@ -59,6 +69,29 @@ def test_set_shift(model):
     assert rho.shift_set == True
 
 
+@pytest.mark.parametrize("model", models_to_test_2)
+def test_add_chains_sample_batching_error(model):
+
+    nchains = 10
+    n_samples = 20
+    ndim = model.ndim
+    num_slices = 300
+
+    X = np.zeros((nchains, n_samples, ndim))
+    Y = np.zeros((nchains, n_samples))
+
+    # Add samples to chains
+    chain = ch.Chains(ndim)
+    chain.add_chains_3d(X, Y)
+
+    model.fitted = True
+
+    # Calculate evidence
+    cal_ev = cbe.Evidence(nchains, model)
+    with pytest.raises(ValueError):
+        cal_ev.add_chains(chain, num_slices=num_slices)
+
+
 @pytest.mark.parametrize("model", models_to_test_1)
 def test_process_run_with_shift(model):
     nchains = 10
@@ -111,7 +144,9 @@ def test_process_run_with_shift(model):
     assert np.exp(rho.ln_evidence_inv_var_var) == pytest.approx(evidence_inv_var_var)
 
 
-def test_add_chains():
+@pytest.mark.parametrize("model", models_to_test_2D)
+@pytest.mark.parametrize("num_slices", chain_batching_options)
+def test_add_chains(model, num_slices):
     nchains = 200
     nsamples = 500
     ndim = 2
@@ -125,22 +160,25 @@ def test_add_chains():
     chain = ch.Chains(ndim)
     chain.add_chains_3d(X, Y)
 
-    # Fit the Hyper_sphere
-    domain = [np.array([1e-1, 1e1])]
-    sphere = mdl.HyperSphere(ndim, domain)
-    sphere.fit(chain.samples, chain.ln_posterior)
+    if hasattr(model, "flow"):
+        model.fit(chain.samples, epochs=5)
+    else:
+        model.fit(chain.samples, chain.ln_posterior)
 
     # Calculate evidence
-    cal_ev = cbe.Evidence(nchains, sphere, cbe.Shifting.MEAN_SHIFT)
-    cal_ev.add_chains(chain)
+    cal_ev = cbe.Evidence(nchains, model, cbe.Shifting.MEAN_SHIFT)
+    cal_ev.add_chains(chain, num_slices=num_slices)
 
     print("cal_ev.evidence_inv = {}".format(np.exp(cal_ev.ln_evidence_inv)))
 
-    assert np.exp(cal_ev.ln_evidence_inv) == pytest.approx(0.159438606)
-    assert np.exp(cal_ev.ln_evidence_inv_var) == pytest.approx(1.164628268e-07)
-    assert np.exp(cal_ev.ln_evidence_inv_var_var) ** 0.5 == pytest.approx(
-        1.142786462e-08
-    )
+    if hasattr(model, "flow"):
+        assert np.exp(cal_ev.ln_evidence_inv) == pytest.approx(0.159438606, rel=0.01)
+    else:
+        assert np.exp(cal_ev.ln_evidence_inv) == pytest.approx(0.159438606)
+        assert np.exp(cal_ev.ln_evidence_inv_var) == pytest.approx(1.164628268e-07)
+        assert np.exp(cal_ev.ln_evidence_inv_var_var) ** 0.5 == pytest.approx(
+            1.142786462e-08
+        )
 
     nsamples1 = 300
     chains1 = ch.Chains(ndim)
@@ -150,14 +188,19 @@ def test_add_chains():
     for i_chain in range(nchains):
         chains2.add_chain(X[i_chain, nsamples1:, :], Y[i_chain, nsamples1:])
 
-    ev = cbe.Evidence(nchains, sphere, cbe.Shifting.MEAN_SHIFT)
+    ev = cbe.Evidence(nchains, model, cbe.Shifting.MEAN_SHIFT)
     # Might have small numerical differences if don't use same mean_shift.
-    ev.add_chains(chains1)
-    ev.add_chains(chains2)
+    ev.add_chains(chains1, num_slices=num_slices)
+    ev.add_chains(chains2, num_slices=num_slices)
 
-    assert np.exp(ev.ln_evidence_inv) == pytest.approx(0.159438606)
-    assert np.exp(ev.ln_evidence_inv_var) == pytest.approx(1.164628268e-07)
-    assert np.exp(ev.ln_evidence_inv_var_var) ** 0.5 == pytest.approx(1.142786462e-08)
+    if hasattr(model, "flow"):
+        assert np.exp(ev.ln_evidence_inv) == pytest.approx(0.159438606, rel=0.01)
+    else:
+        assert np.exp(ev.ln_evidence_inv) == pytest.approx(0.159438606)
+        assert np.exp(ev.ln_evidence_inv_var) == pytest.approx(1.164628268e-07)
+        assert np.exp(ev.ln_evidence_inv_var_var) ** 0.5 == pytest.approx(
+            1.142786462e-08
+        )
 
     return