The Anderson-Darling test

probaforms/metrics/__init__.py

@@ -1,6 +1,7 @@
 from .fd import frechet_distance
 from .mmd import maximum_mean_discrepancy
-from .ks1d import kolmogorov_smirnov_1d, cramer_von_mises_1d, roc_auc_score_1d
+from .ks1d import kolmogorov_smirnov_1d, cramer_von_mises_1d
+from .ks1d import roc_auc_score_1d, anderson_darling_1d
 from .div1d import kullback_leibler_1d, kullback_leibler_1d_kde
 from .div1d import jensen_shannon_1d, jensen_shannon_1d_kde
 
@@ -14,5 +15,6 @@
     'kullback_leibler_1d_kde',
     'jensen_shannon_1d_kde',
     'cramer_von_mises_1d',
-    'roc_auc_score_1d'
+    'roc_auc_score_1d',
+    'anderson_darling_1d'
 ]

probaforms/metrics/ks1d.py

@@ -1,8 +1,12 @@
 import numpy as np
 from sklearn.utils import resample
-from scipy.stats import ks_2samp, cramervonmises_2samp
+from scipy.stats import ks_2samp, cramervonmises_2samp, anderson_ksamp
 from sklearn.metrics import roc_auc_score
 
+import warnings
+warnings.filterwarnings(action='ignore', category=UserWarning) # for _anderson1d
+
+
 
 def _ks1d(data1, data2):
     ks, _ = ks_2samp(data1, data2)
@@ -19,6 +23,10 @@ def _roc1d(x, y):
     auc = np.abs(auc - 0.5) + 0.5
     return auc
 
+def _anderson1d(x, y):
+    res = anderson_ksamp([x, y])
+    return res.statistic
+
 
 def _bootstrap_metric(metric_func, X_real, X_fake, n_iters=100, *args):
     '''
@@ -64,7 +72,7 @@ def _bootstrap_metric(metric_func, X_real, X_fake, n_iters=100, *args):
 
 def kolmogorov_smirnov_1d(X_real, X_fake, n_iters=100):
     '''
-    Calculates the Kolmogorov Smirnov statistics for real and fake samples.
+    Calculates the Kolmogorov Smirnov statistic for real and fake samples.
     The function calculates metric values for each input feature,
     and then averaged them.
 
@@ -80,7 +88,7 @@ def kolmogorov_smirnov_1d(X_real, X_fake, n_iters=100):
     Return:
     -------
     distance: float
-        The estimated KS statistics.
+        The estimated KS statistic.
     Std: fload
         The standard deviation of the distance.
     '''
@@ -106,7 +114,7 @@ def cramer_von_mises_1d(X_real, X_fake, n_iters=100):
     Return:
     -------
     distance: float
-        The estimated KS statistics.
+        The estimated CvM statistic.
     Std: fload
         The standard deviation of the distance.
     '''
@@ -132,9 +140,35 @@ def roc_auc_score_1d(X_real, X_fake, n_iters=100):
     Return:
     -------
     distance: float
-        The estimated KS statistics.
+        The estimated statistic.
     Std: fload
         The standard deviation of the distance.
     '''
 
     return _bootstrap_metric(_roc1d, X_real, X_fake, n_iters)
+
+
+def anderson_darling_1d(X_real, X_fake, n_iters=100):
+    '''
+    Calculates the Anderson-Darling statistic for real and fake samples.
+    The function calculates metric values for each input feature,
+    and then averaged them.
+
+    Parameters:
+    -----------
+    X_real: numpy.ndarray of shape [n_samples, n_features]
+        Real sample.
+    X_fake: numpy.ndarray of shape [n_samples, n_features]
+        Generated sample.
+    n_iters: int
+        The number of bootstrap iterations. Default = 100.
+
+    Return:
+    -------
+    distance: float
+        The estimated statistic.
+    Std: fload
+        The standard deviation of the distance.
+    '''
+
+    return _bootstrap_metric(_anderson1d, X_real, X_fake, n_iters)