[MXNET-1399] multiclass-mcc metric enhancements (#14874)

* multiclass-mcc metric enhancements

 * Rename metric from "PCC" to "mMCC" because though the math is derived
   from Pearson CC, it's utility is as a multiclass extension of Mathews CC.
 * Harden mx.metric.mMCC.update to more variations of input format, similar to
   mx.metric.Accuracy.update.

* Harden mx.metric.PCC.update to more variations of input format, similar to mx.metric.Accuracy.update.

* Enhance testcases for mx.metric.PCC.

Author: tlby

python/mxnet/metric.py

@@ -1572,7 +1572,11 @@ def update(self, labels, preds):
         # update the confusion matrix
         for label, pred in zip(labels, preds):
             label = label.astype('int32', copy=False).asnumpy()
-            pred = pred.asnumpy().argmax(axis=1)
+            pred = pred.asnumpy()
+            if pred.shape != label.shape:
+                pred = pred.argmax(axis=1)
+            else:
+                pred = pred.astype('int32', copy=False)
             n = max(pred.max(), label.max())
             if n >= self.k:
                 self._grow(n + 1 - self.k)

tests/python/unittest/test_metric.py

@@ -304,6 +304,13 @@ def test_pcc():
     _, pear = met_pear.get()
     np.testing.assert_almost_equal(pcc, pear)
 
+    # pcc should also accept pred as scalar rather than softmax vector
+    # like acc does
+    met_pcc.reset()
+    met_pcc.update(labels, [p.argmax(axis=1) for p in preds])
+    _, chk = met_pcc.get()
+    np.testing.assert_almost_equal(pcc, chk)
+
     # check multiclass case against reference implementation
     CM = [
         [ 23, 13,  3 ],