Fix npairloss bug (#51092)

python/paddle/fluid/tests/unittests/test_npair_loss_op.py

@@ -196,5 +196,42 @@ class TestNpairLossOpError(unittest.TestCase):
             self.assertRaises(TypeError, test_labels_type)
 
 
+class TestNpairLossZeroError(unittest.TestCase):
+    def test_errors(self):
+        with paddle.fluid.dygraph.guard():
+
+            def test_anchor_0_size():
+                array = np.array([], dtype=np.float32)
+                anchor = paddle.to_tensor(
+                    np.reshape(array, [0, 0, 0]), dtype='float32'
+                )
+                positive = paddle.to_tensor(
+                    np.reshape(array, [0]), dtype='float32'
+                )
+                array = np.array([1, 2, 3, 4], dtype=np.float32)
+                labels = paddle.to_tensor(
+                    np.reshape(array, [4]), dtype='float32'
+                )
+                paddle.nn.functional.npair_loss(anchor, positive, labels)
+
+            def test_positive_0_size():
+                array = np.array([1], dtype=np.float32)
+                array1 = np.array([], dtype=np.float32)
+                anchor = paddle.to_tensor(
+                    np.reshape(array, [1, 1, 1]), dtype='float32'
+                )
+                positive = paddle.to_tensor(
+                    np.reshape(array1, [0]), dtype='float32'
+                )
+                array = np.array([1, 2, 3, 4], dtype=np.float32)
+                labels = paddle.to_tensor(
+                    np.reshape(array, [4]), dtype='float32'
+                )
+                paddle.nn.functional.npair_loss(anchor, positive, labels)
+
+            self.assertRaises(ValueError, test_anchor_0_size)
+            self.assertRaises(ValueError, test_positive_0_size)
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/nn/functional/loss.py

@@ -373,6 +373,10 @@ def npair_loss(anchor, positive, labels, l2_reg=0.002):
           print(npair_loss)
 
     """
+    if anchor.size == 0:
+        raise ValueError("The dims of anchor should be greater than 0.")
+    if positive.size == 0:
+        raise ValueError("The dims of positive should be greater than 0.")
     check_variable_and_dtype(
         anchor, 'anchor', ['float32', 'float64'], 'npair_loss'
     )