Fix mIoU calculatiton range (open-mmlab#471)

* Fix fence(IoU) = 0 when training on PascalContextDataset59;

* Add a test case in test_metrics() of tests/test_metrics.py to test the bug caused by torch.histc;

* Update tests/test_metrics.py

Co-authored-by: Jerry Jiarui XU <[email protected]>

Co-authored-by: Jerry Jiarui XU <[email protected]>

Author: clownrat6

mmseg/core/evaluation/metrics.py

@@ -57,11 +57,11 @@ def intersect_and_union(pred_label,
 
     intersect = pred_label[pred_label == label]
     area_intersect = torch.histc(
-        intersect.float(), bins=(num_classes), min=0, max=num_classes)
+        intersect.float(), bins=(num_classes), min=0, max=num_classes - 1)
     area_pred_label = torch.histc(
-        pred_label.float(), bins=(num_classes), min=0, max=num_classes)
+        pred_label.float(), bins=(num_classes), min=0, max=num_classes - 1)
     area_label = torch.histc(
-        label.float(), bins=(num_classes), min=0, max=num_classes)
+        label.float(), bins=(num_classes), min=0, max=num_classes - 1)
     area_union = area_pred_label + area_label - area_intersect
     return area_intersect, area_union, area_pred_label, area_label
 

tests/test_metrics.py

@@ -64,30 +64,36 @@ def test_metrics():
     ignore_index = 255
     results = np.random.randint(0, num_classes, size=pred_size)
     label = np.random.randint(0, num_classes, size=pred_size)
+
+    # Test the availability of arg: ignore_index.
     label[:, 2, 5:10] = ignore_index
+
+    # Test the correctness of the implementation of mIoU calculation.
     all_acc, acc, iou = eval_metrics(
         results, label, num_classes, ignore_index, metrics='mIoU')
     all_acc_l, acc_l, iou_l = legacy_mean_iou(results, label, num_classes,
                                               ignore_index)
     assert all_acc == all_acc_l
     assert np.allclose(acc, acc_l)
     assert np.allclose(iou, iou_l)
-
+    # Test the correctness of the implementation of mDice calculation.
     all_acc, acc, dice = eval_metrics(
         results, label, num_classes, ignore_index, metrics='mDice')
     all_acc_l, acc_l, dice_l = legacy_mean_dice(results, label, num_classes,
                                                 ignore_index)
     assert all_acc == all_acc_l
     assert np.allclose(acc, acc_l)
     assert np.allclose(dice, dice_l)
-
+    # Test the correctness of the implementation of joint calculation.
     all_acc, acc, iou, dice = eval_metrics(
         results, label, num_classes, ignore_index, metrics=['mIoU', 'mDice'])
     assert all_acc == all_acc_l
     assert np.allclose(acc, acc_l)
     assert np.allclose(iou, iou_l)
     assert np.allclose(dice, dice_l)
 
+    # Test the correctness of calculation when arg: num_classes is larger
+    # than the maximum value of input maps.
     results = np.random.randint(0, 5, size=pred_size)
     label = np.random.randint(0, 4, size=pred_size)
     all_acc, acc, iou = eval_metrics(
@@ -121,6 +127,17 @@ def test_metrics():
     assert dice[-1] == -1
     assert iou[-1] == -1
 
+    # Test the bug which is caused by torch.histc.
+    # torch.histc:  https://pytorch.org/docs/stable/generated/torch.histc.html
+    # When the arg:bins is set to be same as arg:max,
+    # some channels of mIoU may be nan.
+    results = np.array([np.repeat(31, 59)])
+    label = np.array([np.arange(59)])
+    num_classes = 59
+    all_acc, acc, iou = eval_metrics(
+        results, label, num_classes, ignore_index=255, metrics='mIoU')
+    assert not np.any(np.isnan(iou))
+
 
 def test_mean_iou():
     pred_size = (10, 30, 30)
@@ -182,7 +199,7 @@ def save_arr(input_arrays: list, title: str, is_image: bool, dir: str):
             filenames.append(filename)
         return filenames
 
-    pred_size = (10, 512, 1024)
+    pred_size = (10, 30, 30)
     num_classes = 19
     ignore_index = 255
     results = np.random.randint(0, num_classes, size=pred_size)