diff --git a/dygraph/paddleseg/cvlibs/param_init.py b/dygraph/paddleseg/cvlibs/param_init.py
index 567399c0a0c7d2310931b1c0ccae13cd0d5422b1..495a47c189b8204a260014de4ca23300b7cbcc23 100644
--- a/dygraph/paddleseg/cvlibs/param_init.py
+++ b/dygraph/paddleseg/cvlibs/param_init.py
@@ -23,3 +23,8 @@ def constant_init(param, **kwargs):
 def normal_init(param, **kwargs):
     initializer = fluid.initializer.Normal(**kwargs)
     initializer(param, param.block)
+
+
+def msra_init(param, **kwargs):
+    initializer = fluid.initializer.MSRA(**kwargs)
+    initializer(param, param.block)
diff --git a/dygraph/paddleseg/models/bisenet.py b/dygraph/paddleseg/models/bisenet.py
index 5c1964932b8fc601eabc813910597e914daac33d..a1ae897ef264812dbc3cc623317d290e36e37ff5 100644
--- a/dygraph/paddleseg/models/bisenet.py
+++ b/dygraph/paddleseg/models/bisenet.py
@@ -259,7 +259,7 @@ class BiSeNet(nn.Layer):
         else:
             for sublayer in self.sublayers():
                 if isinstance(sublayer, nn.Conv2d):
-                    param_init.normal_init(sublayer.weight, scale=0.001)
+                    param_init.msra_init(sublayer.weight)
                 elif isinstance(sublayer, nn.SyncBatchNorm):
                     param_init.constant_init(sublayer.weight, value=1.0)
                     param_init.constant_init(sublayer.bias, value=0.0)