fuse onnx unsqueeze-batchnorm-squeeze as batchnorm

tools/onnx/onnx2ncnn.cpp

@@ -610,6 +610,55 @@ static void fuse_hardsigmoid(onnx::GraphProto* mutable_graph, std::map<std::stri
     }
 }
 
+static void fuse_batchnorm1d_squeeze_unsqueeze(onnx::GraphProto* mutable_graph, std::map<std::string, onnx::TensorProto>& weights, std::map<std::string, onnx::TensorProto>& binaryop_weights, std::map<std::string, int>& node_reference, std::set<std::string>& blob_names, int& reduced_node_count, std::vector<std::string>& reduced_binaryop_weights)
+{
+    int node_count = mutable_graph->node_size();
+    for (int i=0; i<node_count; i++)
+    {
+        onnx::NodeProto* node = mutable_graph->mutable_node(i);
+
+        // BatchNormalization <= Unsqueeze - BatchNormalization - Squeeze
+        if (node->op_type() == "Unsqueeze")
+        {
+            if (node_reference.find(node->output(0)) == node_reference.end() || node_reference[node->output(0)] != 1)
+                continue;
+
+            if (i+2 >= node_count)
+                continue;
+
+            onnx::NodeProto* node2 = mutable_graph->mutable_node(i+1);
+            onnx::NodeProto* node3 = mutable_graph->mutable_node(i+2);
+
+            if (node2->op_type() != "BatchNormalization" || node3->op_type() != "Squeeze")
+                continue;
+
+            if (node_reference.find(node2->output(0)) == node_reference.end() || node_reference[node2->output(0)] != 1)
+                continue;
+
+            if (node_reference.find(node3->output(0)) == node_reference.end() || node_reference[node3->output(0)] != 1)
+                continue;
+
+            if (node2->input(0) != node->output(0) || node3->input(0) != node2->output(0))
+                continue;
+
+            // reduce
+            node->set_op_type("noop_reducedncnn");
+            node3->set_op_type("noop_reducedncnn");
+
+            node_reference.erase(node_reference.find(node->output(0)));
+            node_reference.erase(node_reference.find(node2->output(0)));
+            blob_names.erase(node->output(0));
+            blob_names.erase(node2->output(0));
+
+            node2->set_input(0, node->input(0));
+            node2->set_output(0, node3->output(0));
+
+            reduced_node_count += 2;
+            i += 2;
+        }
+    }
+}
+
 int main(int argc, char** argv)
 {
     const char* onnxpb = argv[1];
@@ -802,6 +851,7 @@ int main(int argc, char** argv)
     fuse_shufflechannel (mutable_graph, weights, binaryop_weights, node_reference, blob_names, reduced_node_count, reduced_binaryop_weights);
     fuse_hardsigmoid    (mutable_graph, weights, binaryop_weights, node_reference, blob_names, reduced_node_count, reduced_binaryop_weights);
     fuse_hardswish      (mutable_graph, weights, binaryop_weights, node_reference, blob_names, reduced_node_count, reduced_binaryop_weights);
+    fuse_batchnorm1d_squeeze_unsqueeze(mutable_graph, weights, binaryop_weights, node_reference, blob_names, reduced_node_count, reduced_binaryop_weights);
 
     // remove node_reference entry with reference equals to one
     int splitncnn_blob_count = 0;