fix math lrn op numerical problem

Author: lei-Qiao

tests/test_backend.py

@@ -489,13 +489,20 @@ def test_conv2d_with_input_transpose(self):
                             process_args={"inputs_as_nchw": [_INPUT]},
                             onnx_feed_dict={_INPUT: x_val_for_onnx})
 
-    @unittest.skip("")
-    def test_lrn(self):
-        # FIXME: numerical results are not correct
+    def test_lrn_default(self):
         x_shape = [1, 3, 4, 3]
         x_val = np.arange(1, 1 + np.prod(x_shape)).astype("float32").reshape(x_shape)
-        _ = tf.placeholder(tf.float32, shape=x_val.shape, name=_TFINPUT)
-        op = tf.nn.local_response_normalization(x_val)
+        x_ = tf.placeholder(tf.float32, shape=x_shape, name=_TFINPUT)
+        op = tf.nn.local_response_normalization(x_)
+        _ = tf.identity(op, name=_TFOUTPUT)
+        self._run_test_case([_OUTPUT], {_INPUT: x_val}, rtol=1e-05)
+
+    def test_lrn(self):
+        # can't set bias = 0
+        x_shape = [1, 2, 2, 8]
+        x_val = np.arange(1, 1 + np.prod(x_shape)).astype("float32").reshape(x_shape)
+        x_ = tf.placeholder(tf.float32, shape=x_shape, name=_TFINPUT)
+        op = tf.nn.local_response_normalization(x_, depth_radius=4, bias=2, alpha=2, beta=1)
         _ = tf.identity(op, name=_TFOUTPUT)
         self._run_test_case([_OUTPUT], {_INPUT: x_val}, rtol=1e-05)
 

tf2onnx/onnx_opset/math.py

@@ -253,10 +253,16 @@ def version_1(cls, ctx, node, **kwargs):
         #     output = input / (bias + alpha * sqr_sum) ** beta
         depth_radius = node.get_attr("depth_radius")
         if depth_radius:
-            size = depth_radius.i
+            size = depth_radius.i * 2 + 1
         else:
-            size = 5
+            size = 5 * 2 + 1
+
         node.set_attr("size", size)
+        node.set_attr("alpha", size * node.get_attr("alpha").f)
+
+        ctx.insert_new_node_on_input(node, "Transpose", node.input[0], perm=[0, 3, 1, 2])
+        op_name = utils.make_name(node.name)
+        ctx.insert_new_node_on_output("Transpose", node.output[0], perm=[0, 2, 3, 1],name=op_name)
 
 
 @tf_op(["MatMul", "BatchMatMul"])