diff --git a/tools/onnx2bnn/OnnxConverter.cpp b/tools/onnx2bnn/OnnxConverter.cpp
index 72b48aa..842be24 100644
--- a/tools/onnx2bnn/OnnxConverter.cpp
+++ b/tools/onnx2bnn/OnnxConverter.cpp
@@ -289,6 +289,17 @@ std::vector<std::string> OnnxConverter::Convert(
                  expected_binary_conv_outputs.end());
             if (binary_conv) {
                 binary_conv_outputs.push_back(node.output(0));
+                bool precede_bn = false;
+                for (const auto &node2 : model_proto_.graph().node()) {
+                    if (node2.op_type() == "BatchNormalization" &&
+                        node2.input(0) == node.output(0)) {
+                        precede_bn = true;
+                        break;
+                    }
+                }
+                if (!precede_bn) {
+                    throw std::invalid_argument("Binary convolutions should precede BatchNorm");
+                }
             }
             AddConv(m(node.input(0)), strides, pads, dilations, group,
                     ori_weight_name, bias_name, m(node.output(0)), binary_conv);
@@ -556,6 +567,13 @@ void OnnxConverter::CalculateCoeff(const ONNX_NAMESPACE::NodeProto &node,
                                                             height *
                                                             coeff_a_data[i];
                 }
+                if (node2.input_size() == 2) {
+                    const auto &bias = onnx_float_tensors_[node2.input(2)];
+                    
+                    FORZ(i, coeff_b_data.size()) {
+                        coeff_b_data[i] += coeff_a_data[i] * bias.data[i];
+                    }
+                }
             }
             {
                 FORZ(i, coeff_a_data.size()) { coeff_a_data[i] *= -2; }