Element-wise subtraction from TensorFlow

modules/dnn/src/layers/eltwise_layer.cpp

@@ -98,7 +98,8 @@ public:
     {
         return backendId == DNN_BACKEND_OPENCV ||
                backendId == DNN_BACKEND_HALIDE ||
-               (backendId == DNN_BACKEND_INFERENCE_ENGINE && (op != SUM || coeffs.empty()));
+               (backendId == DNN_BACKEND_INFERENCE_ENGINE &&
+                (preferableTarget != DNN_TARGET_MYRIAD || coeffs.empty()));
     }
 
     bool getMemoryShapes(const std::vector<MatShape> &inputs,
@@ -427,6 +428,7 @@ public:
         lp.type = "Eltwise";
         lp.precision = InferenceEngine::Precision::FP32;
         std::shared_ptr<InferenceEngine::EltwiseLayer> ieLayer(new InferenceEngine::EltwiseLayer(lp));
+        ieLayer->coeff = coeffs;
         if (op == SUM)
             ieLayer->_operation = InferenceEngine::EltwiseLayer::Sum;
         else if (op == PROD)

modules/dnn/src/tensorflow/tf_importer.cpp

@@ -939,7 +939,7 @@ void TFImporter::populateNet(Net dstNet)
             if (getDataLayout(name, data_layouts) == DATA_LAYOUT_UNKNOWN)
                 data_layouts[name] = DATA_LAYOUT_NHWC;
         }
-        else if (type == "BiasAdd" || type == "Add")
+        else if (type == "BiasAdd" || type == "Add" || type == "Sub")
         {
             bool haveConst = false;
             for(int ii = 0; !haveConst && ii < layer.input_size(); ++ii)
@@ -953,6 +953,8 @@ void TFImporter::populateNet(Net dstNet)
             {
                 Mat values = getTensorContent(getConstBlob(layer, value_id));
                 CV_Assert(values.type() == CV_32FC1);
+                if (type == "Sub")
+                    values *= -1.0f;
 
                 int id;
                 if (values.total() == 1)  // is a scalar.
@@ -973,6 +975,12 @@ void TFImporter::populateNet(Net dstNet)
             else
             {
                 layerParams.set("operation", "sum");
+                if (type == "Sub")
+                {
+                    static float subCoeffs[] = {1.f, -1.f};
+                    layerParams.set("coeff", DictValue::arrayReal<float*>(subCoeffs, 2));
+                }
+
                 int id = dstNet.addLayer(name, "Eltwise", layerParams);
                 layer_id[name] = id;
 
@@ -985,36 +993,6 @@ void TFImporter::populateNet(Net dstNet)
                 }
             }
         }
-        else if (type == "Sub")
-        {
-            bool haveConst = false;
-            for(int ii = 0; !haveConst && ii < layer.input_size(); ++ii)
-            {
-                Pin input = parsePin(layer.input(ii));
-                haveConst = value_id.find(input.name) != value_id.end();
-            }
-            CV_Assert(haveConst);
-
-            Mat values = getTensorContent(getConstBlob(layer, value_id));
-            CV_Assert(values.type() == CV_32FC1);
-            values *= -1.0f;
-
-            int id;
-            if (values.total() == 1)  // is a scalar.
-            {
-                layerParams.set("shift", values.at<float>(0));
-                id = dstNet.addLayer(name, "Power", layerParams);
-            }
-            else  // is a vector
-            {
-                layerParams.blobs.resize(1, values);
-                id = dstNet.addLayer(name, "Shift", layerParams);
-            }
-            layer_id[name] = id;
-
-            // one input only
-            connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);
-        }
         else if (type == "MatMul")
         {
             CV_Assert(layer.input_size() == 2);

modules/dnn/test/test_tf_importer.cpp

@@ -139,9 +139,10 @@ TEST_P(Test_TensorFlow_layers, padding)
     runTensorFlowNet("keras_pad_concat");
 }
 
-TEST_P(Test_TensorFlow_layers, eltwise_add_mul)
+TEST_P(Test_TensorFlow_layers, eltwise)
 {
     runTensorFlowNet("eltwise_add_mul");
+    runTensorFlowNet("eltwise_sub");
 }
 
 TEST_P(Test_TensorFlow_layers, pad_and_concat)