Completed coverage. Fixes #1

tests/test_seg_metrics.py

@@ -4,6 +4,49 @@
 from sklearn.metrics import accuracy_score, f1_score, recall_score, precision_score, jaccard_score, cohen_kappa_score
 
 
+def test_segmentation_metrics():
+    shape = (256, 256, 3)
+    # case all pixels belong to the positive class and are detected
+    y_true = np.random.randint(0, 2, shape)
+    y_pred = np.random.randint(0, 2, shape)
+    valid_mask = np.random.randint(0, 2, shape)
+    tpfptnfn = segm.tpfptnfn(y_true, y_pred, valid_mask)
+    decimals = 6
+    metrics = segm.segmentation_metrics(tpfptnfn, decimals=decimals)
+    # acc
+    assert (
+        round(segm._accuracy(tpfptnfn["tp"], tpfptnfn["fp"], tpfptnfn["tn"], tpfptnfn["fn"]), decimals)
+        == metrics["acc"]
+    )
+    # recall
+    assert round(segm._recall(tpfptnfn["tp"], tpfptnfn["fn"]), decimals) == metrics["recall"]
+    # precision
+    assert round(segm._precision(tpfptnfn["tp"], tpfptnfn["fp"]), decimals) == metrics["precision"]
+    # F1
+    assert (
+        round(
+            segm._f1_score(
+                segm._recall(tpfptnfn["tp"], tpfptnfn["fn"]), segm._precision(tpfptnfn["tp"], tpfptnfn["fp"])
+            ),
+            decimals,
+        )
+        == metrics["F1"]
+    )
+    # IoU
+    assert (
+        round(segm._intersection_over_union(tpfptnfn["tp"], tpfptnfn["fp"], tpfptnfn["fn"], tpfptnfn["tn"]), decimals)
+        == metrics["iou"]
+    )
+    # kappa
+    assert (
+        round(
+            segm._kappa(tpfptnfn["tp"], tpfptnfn["fp"], tpfptnfn["tn"], tpfptnfn["fn"], tpfptnfn["n_valid_pixel"]),
+            decimals,
+        )
+        == metrics["kappa"]
+    )
+
+
 def test_tpfptnfn():
     shape = (256, 256, 3)
     # case all pixels belong to the positive class and are detected
@@ -49,6 +92,14 @@ def test_tpfptnfn():
         == np.sum([tpfptnfn["tp"], tpfptnfn["fp"], tpfptnfn["tn"], tpfptnfn["fn"]])
         == tpfptnfn["n_valid_pixel"]
     )
+    # test automatic generation off all-is-valid mask, if no valid-mask is provided
+    tpfptnfn = segm.tpfptnfn(y_true, y_pred, None)
+    assert (
+        np.sum([tpfptnfn["tp"], tpfptnfn["fp"], tpfptnfn["tn"], tpfptnfn["fn"]])
+        == tpfptnfn["n_valid_pixel"]
+        == y_true.size
+        == y_pred.size
+    )
 
 
 def test_accuracy():
@@ -108,9 +159,7 @@ def test_f1_score():
     tpfptnfn = segm.tpfptnfn(y_true, y_pred, valid_mask)
     rec = segm._recall(tpfptnfn["tp"], tpfptnfn["fn"])
     prec = segm._precision(tpfptnfn["tp"], tpfptnfn["fp"])
-    assert np.round(segm._f1_score(rec, prec), decimals=6) == np.round(
-        f1_score(y_true_sk.flatten(), y_pred_sk.flatten()), decimals=6
-    )
+    assert round(segm._f1_score(rec, prec), 6) == round(f1_score(y_true_sk.flatten(), y_pred_sk.flatten()), 6)
 
 
 def test_intersection_over_union():
@@ -127,6 +176,17 @@ def test_intersection_over_union():
     assert segm._intersection_over_union(
         tpfptnfn["tp"], tpfptnfn["fp"], tpfptnfn["fn"], tpfptnfn["tn"]
     ) == jaccard_score(y_true_sk.flatten(), y_pred_sk.flatten())
+    # test the special case were we have only true negatives and no wrong predictions
+    y_true = np.zeros(shape)
+    y_pred = np.zeros(shape)
+    y_true_sk = y_true[valid_mask == 1]
+    y_pred_sk = y_pred[valid_mask == 1]
+    tpfptnfn = segm.tpfptnfn(y_true, y_pred, valid_mask)
+    assert (
+        segm._intersection_over_union(tpfptnfn["tp"], tpfptnfn["fp"], tpfptnfn["fn"], tpfptnfn["tn"])
+        == 1.0
+        == jaccard_score(y_true_sk.flatten(), y_pred_sk.flatten(), zero_division=1.0)
+    )
 
 
 def test_kappa():
@@ -141,9 +201,9 @@ def test_kappa():
     y_pred_sk = y_pred[valid_mask == 1]
     tpfptnfn = segm.tpfptnfn(y_true, y_pred, valid_mask)
     assert (
-        np.round(
+        round(
             segm._kappa(tpfptnfn["tp"], tpfptnfn["fp"], tpfptnfn["tn"], tpfptnfn["fn"], tpfptnfn["n_valid_pixel"]),
-            decimals=6,
+            6,
         )
-        == np.round(cohen_kappa_score(y_true_sk.flatten(), y_pred_sk.flatten()), decimals=6)
+        == round(cohen_kappa_score(y_true_sk.flatten(), y_pred_sk.flatten()), 6)
     )