!837 add onnx op converter: ReLU6 and DepthwiseConv2dNative

Merge pull request !837 from chenzupeng/master

mindspore/ccsrc/onnx/onnx_exporter.cc

@@ -330,6 +330,10 @@ class OnnxExporter {
                       onnx::GraphProto *graph_proto);
   void ExportPrimPReLU(const FuncGraphPtr &func_graph, const CNodePtr &node, std::map<AnfNodePtr, size_t> *node_map_ptr,
                        onnx::GraphProto *graph_proto);
+  void ExportPrimReLU6(const FuncGraphPtr &func_graph, const CNodePtr &node, std::map<AnfNodePtr, size_t> *node_map_ptr,
+                       onnx::GraphProto *graph_proto);
+  void ExportPrimDepthwiseConv2d(const FuncGraphPtr &func_graph, const CNodePtr &node,
+                                 std::map<AnfNodePtr, size_t> *node_map_ptr, onnx::GraphProto *graph_proto);
 
   void ExportMergeConv(const FuncGraphPtr &func_graph, const CNodePtr &node, std::map<AnfNodePtr, size_t> *node_map_ptr,
                        onnx::GraphProto *graph_proto);
@@ -711,6 +715,115 @@ void OnnxExporter::ExportPrimPReLU(const FuncGraphPtr & /*func_graph*/, const CN
   node_proto->add_input(input_slope);
 }
 
+void OnnxExporter::ExportPrimReLU6(const FuncGraphPtr & /*func_graph*/, const CNodePtr &node,
+                                   std::map<AnfNodePtr, size_t> *node_map_ptr, onnx::GraphProto *const graph_proto) {
+  auto input_x = GetNodeInputName(node->input(1), node_map_ptr, graph_proto);
+  auto node_idx = AllocateNodeIndex();
+  (*node_map_ptr)[node] = node_idx;
+  onnx::NodeProto *node_proto = graph_proto->add_node();
+  node_proto->set_op_type("Clip");
+  node_proto->add_output(std::to_string(node_idx));
+  node_proto->add_input(input_x);
+  onnx::AttributeProto *attr_proto = node_proto->add_attribute();
+  attr_proto->set_type(onnx::AttributeProto_AttributeType_FLOAT);
+  attr_proto->set_name("min");
+  attr_proto->set_f(0.f);
+  attr_proto = node_proto->add_attribute();
+  attr_proto->set_type(onnx::AttributeProto_AttributeType_FLOAT);
+  attr_proto->set_name("max");
+  attr_proto->set_f(6.f);
+}
+
+void OnnxExporter::ExportPrimDepthwiseConv2d(const FuncGraphPtr & /*func_graph*/, const CNodePtr &node,
+                                             std::map<AnfNodePtr, size_t> *node_map_ptr,
+                                             onnx::GraphProto *const graph_proto) {
+  auto input_x = GetNodeInputName(node->input(1), node_map_ptr, graph_proto);
+  auto input_w = GetNodeInputName(node->input(2), node_map_ptr, graph_proto);
+  auto x_shape = dyn_cast<abstract::Shape>(node->input(1)->Shape());
+  auto w_shape = dyn_cast<abstract::Shape>(node->input(2)->Shape());
+  MS_EXCEPTION_IF_NULL(x_shape);
+  MS_EXCEPTION_IF_NULL(w_shape);
+  if (x_shape->shape().size() != 4 || w_shape->shape().size() != 4) {
+    MS_LOG(EXCEPTION) << "DepthwiseConv2d input shape should be 4d.";
+  }
+  if (w_shape->shape()[0] != 1 && w_shape->shape()[1] != 1) {
+    MS_LOG(EXCEPTION) << "DepthwiseConv2d weight shape[0] != 1 and shape[1] != 1, cannot reshape";
+  }
+  // create w_shape constant node
+  auto node_idx = AllocateNodeIndex();
+  onnx::NodeProto *node_proto = graph_proto->add_node();
+  std::string name_w_shape = std::to_string(node_idx);
+  node_proto->add_output(name_w_shape);
+  node_proto->set_op_type("Constant");
+  // create Value Tensor
+  onnx::AttributeProto *attr_proto = node_proto->add_attribute();
+  attr_proto->set_name("value");
+  attr_proto->set_type(onnx::AttributeProto_AttributeType_TENSOR);
+  onnx::TensorProto *tensor_proto = attr_proto->mutable_t();
+  tensor_proto->add_dims(static_cast<::google::protobuf::int64>(w_shape->shape().size()));
+  tensor_proto->set_data_type(onnx::TensorProto_DataType_INT64);
+  // reshape
+  tensor_proto->add_int64_data(w_shape->shape()[1]);
+  tensor_proto->add_int64_data(w_shape->shape()[0]);
+  tensor_proto->add_int64_data(w_shape->shape()[2]);
+  tensor_proto->add_int64_data(w_shape->shape()[3]);
+
+  // add reshape node
+  node_idx = AllocateNodeIndex();
+  node_proto = graph_proto->add_node();
+  node_proto->set_op_type(prim::kPrimReshape->name());
+  node_proto->add_input(input_w);
+  node_proto->add_input(name_w_shape);
+  input_w = std::to_string(node_idx);
+  node_proto->add_output(input_w);
+
+  // add conv node
+  node_idx = AllocateNodeIndex();
+  (*node_map_ptr)[node] = node_idx;
+  node_proto = graph_proto->add_node();
+  node_proto->set_op_type("Conv");
+  node_proto->add_input(input_x);
+  node_proto->add_input(input_w);
+  node_proto->add_output(std::to_string(node_idx));
+  // set attributes
+  AnfNodePtr op = node->input(0);
+  auto op_value = dyn_cast<ValueNode>(op);
+  auto prim = dyn_cast<Primitive>(op_value->value());
+  // set dilations
+  onnx::AttributeProto *onnx_attr_proto = node_proto->add_attribute();
+  onnx_attr_proto->set_name("dilations");
+  SetAttrTupleValueToProto<2>(prim->GetAttr("dilation"), onnx::AttributeProto_AttributeType_INTS, onnx_attr_proto,
+                              prim);
+  // set group
+  onnx_attr_proto = node_proto->add_attribute();
+  onnx_attr_proto->set_name("group");
+  onnx_attr_proto->set_type(onnx::AttributeProto_AttributeType_INT);
+  onnx_attr_proto->set_i(x_shape->shape()[1]);
+  // set kernel_shape
+  onnx_attr_proto = node_proto->add_attribute();
+  onnx_attr_proto->set_name("kernel_shape");
+  SetAttrTupleValueToProto<0>(prim->GetAttr("kernel_size"), onnx::AttributeProto_AttributeType_INTS, onnx_attr_proto,
+                              prim);
+
+  // set pad
+  onnx_attr_proto = node_proto->add_attribute();
+  auto attr_value = GetValue<std::string>(prim->GetAttr("pad_mode"));
+  onnx_attr_proto->set_name("auto_pad");
+  onnx_attr_proto->set_type(onnx::AttributeProto_AttributeType_STRING);
+  if (attr_value == "valid") {
+    onnx_attr_proto->set_s("VALID");
+  } else if (attr_value == "same") {
+    onnx_attr_proto->set_s("SAME_UPPER");
+  } else {
+    onnx_attr_proto->set_name("pads");
+    SetAttrTupleValueToProto(prim->GetAttr("pads"), onnx::AttributeProto_AttributeType_INTS, onnx_attr_proto, prim);
+  }
+  // set strides
+  onnx_attr_proto = node_proto->add_attribute();
+  onnx_attr_proto->set_name("strides");
+  SetAttrTupleValueToProto<2>(prim->GetAttr("stride"), onnx::AttributeProto_AttributeType_INTS, onnx_attr_proto, prim);
+}
+
 void OnnxExporter::ExportCNode(const FuncGraphPtr &func_graph, const CNodePtr &node,
                                std::map<AnfNodePtr, size_t> *node_map_ptr, onnx::GraphProto *const graph_proto) {
   // Type of the 2nd input of 'Reshape' of MindSpore is tuple, but ONNX's is tensor, need to do some convert
@@ -732,6 +845,16 @@ void OnnxExporter::ExportCNode(const FuncGraphPtr &func_graph, const CNodePtr &n
     return ExportPrimPReLU(func_graph, node, node_map_ptr, graph_proto);
   }
 
+  // MindSpore ReLU6(x) --> ONNX Clip[min=0.f, max=6.f](x)
+  if (node->IsApply(std::make_shared<Primitive>("ReLU6"))) {
+    return ExportPrimReLU6(func_graph, node, node_map_ptr, graph_proto);
+  }
+
+  // MindSpore DepthwiseConv2dNative --> ONNX Conv(x, reshape(w))
+  if (node->IsApply(std::make_shared<Primitive>("DepthwiseConv2dNative"))) {
+    return ExportPrimDepthwiseConv2d(func_graph, node, node_map_ptr, graph_proto);
+  }
+
   auto inputs = node->inputs();
   if (inputs.size() < 1) {
     MS_LOG(EXCEPTION) << "Inputs of apply node is empty";

mindspore/model_zoo/mobilenet.py

@@ -89,7 +89,8 @@ class DepthwiseConv(nn.Cell):
         self.channel_multiplier = channel_multiplier
         self.out_channels = in_planes * channel_multiplier
         self.kernel_size = (kernel_size, kernel_size)
-        self.depthwise_conv = P.DepthwiseConv2dNative(channel_multiplier=channel_multiplier, kernel_size=kernel_size,
+        self.depthwise_conv = P.DepthwiseConv2dNative(channel_multiplier=channel_multiplier,
+                                                      kernel_size=self.kernel_size,
                                                       stride=stride, pad_mode=pad_mode, pad=pad)
         self.bias_add = P.BiasAdd()
         weight_shape = [channel_multiplier, in_planes, *self.kernel_size]

tests/ut/python/utils/test_serialize.py

@@ -323,6 +323,22 @@ class BatchNormTester(nn.Cell):
         return self.bn(x)
 
 
+class DepthwiseConv2dAndReLU6(nn.Cell):
+    "Net for testing DepthwiseConv2d and ReLU6"
+
+    def __init__(self, input_channel, kernel_size):
+        super(DepthwiseConv2dAndReLU6, self).__init__()
+        weight_shape = [1, input_channel, kernel_size, kernel_size]
+        from mindspore.common.initializer import initializer
+        self.weight = Parameter(initializer('ones', weight_shape), name='weight')
+        self.depthwise_conv = P.DepthwiseConv2dNative(channel_multiplier=1, kernel_size=(kernel_size, kernel_size))
+        self.relu6 = nn.ReLU6()
+
+    def construct(self, x):
+        x = self.depthwise_conv(x, self.weight)
+        x = self.relu6(x)
+        return x
+
 def test_batchnorm_train_onnx_export():
     input = Tensor(np.ones([1, 3, 32, 32]).astype(np.float32) * 0.01)
     net = BatchNormTester(3)
@@ -421,6 +437,38 @@ def test_lenet5_onnx_load_run():
     print(outputs[0])
 
 
+@run_on_onnxruntime
+def test_depthwiseconv_relu6_onnx_load_run():
+    onnx_file = 'depthwiseconv_relu6.onnx'
+    input_channel = 3
+    input = Tensor(np.ones([1, input_channel, 32, 32]).astype(np.float32) * 0.01)
+    net = DepthwiseConv2dAndReLU6(input_channel, kernel_size=3)
+    export(net, input, file_name=onnx_file, file_format='ONNX')
+
+    import onnx
+    import onnxruntime as ort
+
+    print('--------------------- onnx load ---------------------')
+    # Load the ONNX model
+    model = onnx.load(onnx_file)
+    # Check that the IR is well formed
+    onnx.checker.check_model(model)
+    # Print a human readable representation of the graph
+    g = onnx.helper.printable_graph(model.graph)
+    print(g)
+
+    print('------------------ onnxruntime run ------------------')
+    ort_session = ort.InferenceSession(onnx_file)
+    input_map = {'x' : input.asnumpy()}
+    # provide only input x to run model
+    outputs = ort_session.run(None, input_map)
+    print(outputs[0])
+    # overwrite default weight to run model
+    for item in net.trainable_params():
+        input_map[item.name] = np.ones(item.default_input.asnumpy().shape, dtype=np.float32)
+    outputs = ort_session.run(None, input_map)
+    print(outputs[0])
+
 def teardown_module():
     files = ['parameters.ckpt', 'new_ckpt.ckpt', 'lenet5.onnx', 'batch_norm.onnx', 'empty.ckpt']
     for item in files: