Merge remote-tracking branch 'upstream/master'

Author: lvapeab

keras_wrapper/dataset.py

@@ -1528,7 +1528,7 @@ def preprocessReal(labels_list):
             labels = []
             with codecs.open(labels_list, 'r', encoding='utf-8') as list_:
                 for line in list_:
-                    labels.append(int(line.rstrip('\n')))
+                    labels.append(float(line))
         elif isinstance(labels_list, list):
             labels = labels_list
         else:

keras_wrapper/extra/callbacks.py

@@ -89,7 +89,7 @@ def __init__(self,
                  each_n_epochs=1,
                  max_eval_samples=None,
                  extra_vars=None,
-                 normalize=True,
+                 normalize=False,
                  normalization_type=None,
                  output_types=None,
                  is_text=False,
@@ -261,7 +261,7 @@ def __init__(self,
 
         else:
             # Convert min_pred_multilabel to list
-            if isinstance(self.min_pred_multilabel, list):
+            if not isinstance(self.min_pred_multilabel, list):
                 self.min_pred_multilabel = [self.min_pred_multilabel for _ in self.gt_pos]
 
         super(EvalPerformance, self).__init__()

keras_wrapper/extra/evaluation.py

@@ -253,7 +253,7 @@ def multiclass_metrics(pred_list, verbose, extra_vars, split):
     accuracy = sklearn_metrics.accuracy_score(y_gt, y_pred)
     accuracy_balanced = sklearn_metrics.accuracy_score(y_gt, y_pred, sample_weight=sample_weights)
     # Compute Precision, Recall and F1 score
-    avrg = extra_vars.get('average_mode', None)
+    avrg = extra_vars.get('average_mode', 'macro')
     precision, recall, f1, _ = sklearn_metrics.precision_recall_fscore_support(y_gt, y_pred, average=avrg)
     # Compute Confusion Matrix
     cf = sklearn_metrics.confusion_matrix(np.argmax(y_gt, -1), np.argmax(y_pred, -1))
@@ -271,7 +271,7 @@ def multiclass_metrics(pred_list, verbose, extra_vars, split):
     # Compute top 5 fp classes
     top5_fps = np.argpartition(cf * neg_identity, -5)[:, -5:][:, ::-1]
     # Compute top 5 accuracy
-    arg_top5_pred = np.argpartition(y_pred, -5)[:, -5:]
+    arg_top5_pred = np.argpartition(pred_list, -5)[:, -5:]
     arg_gt = np.argmax(y_gt, -1)
     top5_acc = np.mean(np.max(arg_top5_pred == np.repeat(np.expand_dims(arg_gt, -1), 5, -1), -1))