[DGL-LifeSci] Handle Same Labels for ROC AUC (#1581)

* Handle corner case for ROC AUC

* Update

* Update doct

apps/life_sci/python/dgllife/utils/eval.py

@@ -173,7 +173,9 @@ class Meter(object):
             task_w = mask[:, task]
             task_y_true = y_true[:, task][task_w != 0]
             task_y_pred = y_pred[:, task][task_w != 0]
-            scores.append(score_func(task_y_true, task_y_pred))
+            task_score = score_func(task_y_true, task_y_pred)
+            if task_score is not None:
+                scores.append(task_score)
         return self._reduce_scores(scores, reduction)
 
     def pearson_r2(self, reduction='none'):
@@ -236,6 +238,9 @@ class Meter(object):
     def roc_auc_score(self, reduction='none'):
         """Compute roc-auc score for binary classification.
 
+        ROC-AUC scores are not well-defined in cases where labels for a task have one single
+        class only. In this case we will simply ignore this task and print a warning message.
+
         Parameters
         ----------
         reduction : 'none' or 'mean' or 'sum'
@@ -253,7 +258,12 @@ class Meter(object):
         assert (self.mean is None) and (self.std is None), \
             'Label normalization should not be performed for binary classification.'
         def score(y_true, y_pred):
-            return roc_auc_score(y_true.long().numpy(), torch.sigmoid(y_pred).numpy())
+            if len(y_true.unique()) == 1:
+                print('Warning: Only one class {} present in y_true for a task. '
+                      'ROC AUC score is not defined in that case.'.format(y_true[0]))
+                return None
+            else:
+                return roc_auc_score(y_true.long().numpy(), torch.sigmoid(y_pred).numpy())
         return self.multilabel_score(score, reduction)
 
     def compute_metric(self, metric_name, reduction='none'):

apps/life_sci/tests/utils/test_eval.py

@@ -17,8 +17,8 @@ def test_Meter():
 
     # pearson r2
     meter = Meter(label_mean, label_std)
-    meter.update(label, pred)
-    true_scores = [0.7499999999999999, 0.7499999999999999]
+    meter.update(pred, label)
+    true_scores = [0.7500000774286983, 0.7500000516191412]
     assert meter.pearson_r2() == true_scores
     assert meter.pearson_r2('mean') == np.mean(true_scores)
     assert meter.pearson_r2('sum') == np.sum(true_scores)
@@ -27,7 +27,7 @@ def test_Meter():
     assert meter.compute_metric('r2', 'sum') == np.sum(true_scores)
 
     meter = Meter(label_mean, label_std)
-    meter.update(label, pred, mask)
+    meter.update(pred, label, mask)
     true_scores = [1.0, 1.0]
     assert meter.pearson_r2() == true_scores
     assert meter.pearson_r2('mean') == np.mean(true_scores)
@@ -38,7 +38,7 @@ def test_Meter():
 
     # mae
     meter = Meter()
-    meter.update(label, pred)
+    meter.update(pred, label)
     true_scores = [0.1666666716337204, 0.1666666716337204]
     assert meter.mae() == true_scores
     assert meter.mae('mean') == np.mean(true_scores)
@@ -48,7 +48,7 @@ def test_Meter():
     assert meter.compute_metric('mae', 'sum') == np.sum(true_scores)
 
     meter = Meter()
-    meter.update(label, pred, mask)
+    meter.update(pred, label, mask)
     true_scores = [0.25, 0.0]
     assert meter.mae() == true_scores
     assert meter.mae('mean') == np.mean(true_scores)
@@ -57,23 +57,23 @@ def test_Meter():
     assert meter.compute_metric('mae', 'mean') == np.mean(true_scores)
     assert meter.compute_metric('mae', 'sum') == np.sum(true_scores)
 
-    # rmse
+    # rmsef
     meter = Meter(label_mean, label_std)
-    meter.update(label, pred)
-    true_scores = [0.22125875529784111, 0.5937311018897714]
+    meter.update(pred, label)
+    true_scores = [0.41068359261794546, 0.4106836107598449]
     assert torch.allclose(torch.tensor(meter.rmse()), torch.tensor(true_scores))
     assert torch.allclose(torch.tensor(meter.compute_metric('rmse')), torch.tensor(true_scores))
 
     meter = Meter(label_mean, label_std)
-    meter.update(label, pred, mask)
-    true_scores = [0.1337071188699867, 0.5019903799993205]
+    meter.update(pred, label, mask)
+    true_scores = [0.44433766459035057, 0.5019903799993205]
     assert torch.allclose(torch.tensor(meter.rmse()), torch.tensor(true_scores))
     assert torch.allclose(torch.tensor(meter.compute_metric('rmse')), torch.tensor(true_scores))
 
     # roc auc score
     meter = Meter()
-    meter.update(label, pred)
-    true_scores = [1.0, 0.75]
+    meter.update(pred, label)
+    true_scores = [1.0, 1.0]
     assert meter.roc_auc_score() == true_scores
     assert meter.roc_auc_score('mean') == np.mean(true_scores)
     assert meter.roc_auc_score('sum') == np.sum(true_scores)
@@ -82,7 +82,7 @@ def test_Meter():
     assert meter.compute_metric('roc_auc_score', 'sum') == np.sum(true_scores)
 
     meter = Meter()
-    meter.update(label, pred, mask)
+    meter.update(pred, label, mask)
     true_scores = [1.0, 1.0]
     assert meter.roc_auc_score() == true_scores
     assert meter.roc_auc_score('mean') == np.mean(true_scores)
@@ -91,5 +91,20 @@ def test_Meter():
     assert meter.compute_metric('roc_auc_score', 'mean') == np.mean(true_scores)
     assert meter.compute_metric('roc_auc_score', 'sum') == np.sum(true_scores)
 
+def test_cases_with_undefined_scores():
+    label = torch.tensor([[0., 1.],
+                          [0., 1.],
+                          [1., 1.]])
+    pred = torch.tensor([[0.5, 0.5],
+                         [0., 1.],
+                         [1., 0.]])
+    meter = Meter()
+    meter.update(pred, label)
+    true_scores = [1.0]
+    assert meter.roc_auc_score() == true_scores
+    assert meter.roc_auc_score('mean') == np.mean(true_scores)
+    assert meter.roc_auc_score('sum') == np.sum(true_scores)
+
 if __name__ == '__main__':
     test_Meter()
+    test_cases_with_undefined_scores()