diff --git a/python_module/megengine/functional/nn.py b/python_module/megengine/functional/nn.py
index b7a8140f..93fd66b7 100644
--- a/python_module/megengine/functional/nn.py
+++ b/python_module/megengine/functional/nn.py
@@ -427,7 +427,7 @@ def batch_norm2d(
     return output
 
 
-def one_hot(inp: Tensor, num_classes: int = -1) -> Tensor:
+def one_hot(inp: Tensor, num_classes: int) -> Tensor:
     r"""
     Perform one-hot encoding for the input tensor.
 
@@ -457,8 +457,6 @@ def one_hot(inp: Tensor, num_classes: int = -1) -> Tensor:
     """
     comp_node, comp_graph = _decide_comp_node_and_comp_graph(inp)
 
-    if num_classes == -1:
-        num_classes = inp.max() + 1
     zeros = mgb.make_immutable(value=0, comp_node=comp_node, comp_graph=comp_graph)
     zeros_symvar = zeros.broadcast(inp.shapeof(), num_classes)
 
diff --git a/python_module/test/unit/functional/test_onehot.py b/python_module/test/unit/functional/test_onehot.py
index 808323c1..3edbe5de 100644
--- a/python_module/test/unit/functional/test_onehot.py
+++ b/python_module/test/unit/functional/test_onehot.py
@@ -8,7 +8,7 @@ from megengine.test import assertTensorClose
 
 def test_onehot_low_dimension():
     inp = tensor(np.arange(1, 4, dtype=np.int32))
-    out = F.one_hot(inp)
+    out = F.one_hot(inp, num_classes=4)
 
     assertTensorClose(
         out.numpy(), np.eye(4, dtype=np.int32)[np.arange(1, 4, dtype=np.int32)]