Merge pull request #69 from jet-net/norm_bug_fix

FPD and KPD normalization bug fix

jetnet/evaluation/gen_metrics.py

@@ -707,10 +707,10 @@ def fpd(
             stacklevel=2,
         )
 
-    real_features, gen_features = _check_get_ndarray(real_features, gen_features)
+    X, Y = _check_get_ndarray(real_features, gen_features)
 
     if normalise:
-        X, Y = _normalise_features(real_features, gen_features)
+        X, Y = _normalise_features(X, Y)
 
     # regular intervals in 1/N
     batches = (1 / np.linspace(1.0 / min_samples, 1.0 / max_samples, num_points)).astype("int32")
@@ -836,10 +836,10 @@ def kpd(
     Returns:
         Tuple[float, float]: median and error of KPD.
     """
-    real_features, gen_features = _check_get_ndarray(real_features, gen_features)
+    X, Y = _check_get_ndarray(real_features, gen_features)
 
     if normalise:
-        X, Y = _normalise_features(real_features, gen_features)
+        X, Y = _normalise_features(X, Y)
 
     if num_threads is None:
         vals_point = _kpd_batches(X, Y, num_batches, batch_size, seed)

tests/evaluation/test_gen_metrics.py

@@ -7,19 +7,37 @@
 
 test_zeros = np.zeros((50_000, 2))
 test_ones = np.ones((50_000, 2))
+test_twos = np.ones((50_000, 2)) * 2
 
 
 def test_fpd():
     val, err = evaluation.fpd(test_zeros, test_zeros)
     assert val == approx(0, abs=0.01)
     assert err < 1e-3
 
-    val, err = evaluation.fpd(test_zeros, test_ones)
-    assert val == approx(2, rel=0.01)
+    val, err = evaluation.fpd(test_twos, test_zeros)
+    assert val == approx(2, rel=0.01)  # 1^2 + 1^2
+    assert err < 1e-3
+
+    # test normalization
+    val, err = evaluation.fpd(test_zeros, test_zeros, normalise=False)  # should have no effect
+    assert val == approx(0, abs=0.01)
+    assert err < 1e-3
+
+    val, err = evaluation.fpd(test_twos, test_zeros, normalise=False)
+    assert val == approx(8, rel=0.01)  # 2^2 + 2^2
     assert err < 1e-3
 
 
 @pytest.mark.parametrize("num_threads", [None, 2])  # test numba parallelization
 def test_kpd(num_threads):
     assert evaluation.kpd(test_zeros, test_zeros, num_threads=num_threads) == approx([0, 0])
-    assert evaluation.kpd(test_zeros, test_ones, num_threads=num_threads) == approx([15, 0])
+    assert evaluation.kpd(test_twos, test_zeros, num_threads=num_threads) == approx([15, 0])
+
+    # test normalization
+    assert evaluation.kpd(
+        test_zeros, test_zeros, normalise=False, num_threads=num_threads
+    ) == approx([0, 0])
+    assert evaluation.kpd(
+        test_twos, test_zeros, normalise=False, num_threads=num_threads
+    ) == approx([624, 0])