Merge pull request #14996 from l-bat:ocv_deconv3d

* Support Deconvolution3D on IE backend

* Add test tag

* Fix tests

modules/dnn/src/layers/convolution_layer.cpp

@@ -67,7 +67,7 @@ public:
     BaseConvolutionLayerImpl(const LayerParams &params)
     {
         setParamsFrom(params);
-        getConvolutionKernelParams(params, kernel_size, pads_begin, pads_end, strides, dilations, padMode);
+        getConvolutionKernelParams(params, kernel_size, pads_begin, pads_end, strides, dilations, padMode, adjust_pads);
 
         numOutput = params.get<int>("num_output");
         int ngroups = params.get<int>("group", 1);
@@ -83,14 +83,14 @@ public:
             pad = Size(pads_begin[1], pads_begin[0]);
             dilation = Size(dilations[1], dilations[0]);
 
-            adjust_pads.push_back(params.get<int>("adj_h", 0));
-            adjust_pads.push_back(params.get<int>("adj_w", 0));
-
             adjustPad.height = adjust_pads[0];
             adjustPad.width = adjust_pads[1];
-            CV_Assert(adjustPad.width < stride.width &&
-                      adjustPad.height < stride.height);
         }
+
+        for (int i = 0; i < adjust_pads.size(); i++) {
+            CV_Assert(adjust_pads[i] < strides[i]);
+        }
+
         fusedWeights = false;
         fusedBias = false;
     }
@@ -1241,29 +1241,39 @@ public:
     virtual bool supportBackend(int backendId) CV_OVERRIDE
     {
 #ifdef HAVE_INF_ENGINE
-        const int outGroupCn = blobs[0].size[1];  // Weights are in IOHW layout
+        const int outGroupCn = blobs[0].size[1];  // Weights are in IOHW or IODHW layout
         const int group = numOutput / outGroupCn;
 
         if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
         {
-            if (kernel_size.size() == 3)
-                CV_Error(Error::StsNotImplemented, "Unsupported deconvolution3D layer");
+            if (kernel_size.size() == 3 && preferableTarget != DNN_TARGET_CPU) {
+                return false;
+            }
 
-            if (adjustPad.height || adjustPad.width)
+            if (std::accumulate(adjust_pads.begin(), adjust_pads.end(), 0, std::plus<size_t>()) > 0)
             {
                 if (padMode.empty())
                 {
                     if (preferableTarget != DNN_TARGET_CPU && group != 1)
                     {
-                        if ((adjustPad.height && pad.height) || (adjustPad.width && pad.width))
+                        for (int i = 0; i < adjust_pads.size(); i++) {
+                            if (adjust_pads[i] && pads_begin[i])
+                                return false;
+                        }
+                    }
+                    for (int i = 0; i < adjust_pads.size(); i++) {
+                        if (pads_end[i] < adjust_pads[i])
                             return false;
                     }
-                    return pad.width >= adjustPad.width && pad.height >= adjustPad.height;
+                    return true;
                 }
                 else if (padMode == "SAME")
                 {
-                    return kernel.width >= pad.width + 1 + adjustPad.width &&
-                           kernel.height >= pad.height + 1 + adjustPad.height;
+                    for (int i = 0; i < adjust_pads.size(); i++) {
+                        if (kernel_size[i] < pads_begin[i] + 1 + adjust_pads[i])
+                            return false;
+                    }
+                    return true;
                 }
                 else if (padMode == "VALID")
                     return false;
@@ -1274,7 +1284,7 @@ public:
                 return preferableTarget == DNN_TARGET_CPU;
             }
             if (preferableTarget == DNN_TARGET_OPENCL || preferableTarget == DNN_TARGET_OPENCL_FP16)
-                return dilation.width == 1 && dilation.height == 1;
+                return std::accumulate(dilations.begin(), dilations.end(), 1, std::multiplies<size_t>()) == 1;
             return true;
         }
         else
@@ -1861,11 +1871,14 @@ public:
 #ifdef HAVE_INF_ENGINE
     virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> > &) CV_OVERRIDE
     {
-        auto ieWeights = wrapToInfEngineBlob(blobs[0], InferenceEngine::Layout::OIHW);
+        InferenceEngine::Layout layout = blobs[0].dims == 5? InferenceEngine::Layout::NCDHW :
+                                                             InferenceEngine::Layout::OIHW;
+
+        auto ieWeights = wrapToInfEngineBlob(blobs[0], layout);
         if (fusedWeights)
         {
             ieWeights = InferenceEngine::make_shared_blob<float>(
-                                InferenceEngine::Precision::FP32, InferenceEngine::Layout::OIHW,
+                                InferenceEngine::Precision::FP32, layout,
                                 ieWeights->dims());
             ieWeights->allocate();
 
@@ -1874,7 +1887,7 @@ public:
             transpose(weightsMat, newWeights);
         }
 
-        const int outGroupCn = blobs[0].size[1];  // Weights are in IOHW layout
+        const int outGroupCn = blobs[0].size[1];  // Weights are in IOHW or OIDHW layout
         const int group = numOutput / outGroupCn;
 
         InferenceEngine::Builder::DeconvolutionLayer ieLayer(name);
@@ -1886,12 +1899,19 @@ public:
 
         if (padMode.empty())
         {
-            ieLayer.setPaddingsEnd({pads_end[0] - adjust_pads[0], pads_end[1] - adjust_pads[1]});
+            std::vector<size_t> paddings_end;
+            for (int i = 0; i < pads_end.size(); i++) {
+                paddings_end.push_back(pads_end[i] - adjust_pads[i]);
+            }
+            ieLayer.setPaddingsEnd(paddings_end);
         }
         else if (padMode == "SAME")
         {
-            ieLayer.setPaddingsEnd({kernel_size[0] - pads_begin[0] - 1 - adjust_pads[0],
-                                    kernel_size[1] - pads_begin[1] - 1 - adjust_pads[1]});
+            std::vector<size_t> paddings_end;
+            for (int i = 0; i < pads_begin.size(); i++) {
+                paddings_end.push_back(kernel_size[i] - pads_begin[i] - 1 - adjust_pads[i]);
+            }
+            ieLayer.setPaddingsEnd(paddings_end);
         }
         ieLayer.setGroup((size_t)group);
         ieLayer.setOutDepth((size_t)numOutput);
@@ -1911,10 +1931,12 @@ public:
 
         float flops = 0;
         int outChannels = blobs[0].size[0];
+        size_t karea = std::accumulate(kernel_size.begin(), kernel_size.end(),
+                                       1, std::multiplies<size_t>());
 
         for (int i = 0; i < inputs.size(); i++)
         {
-            flops += CV_BIG_INT(2)*outChannels*kernel.area()*total(inputs[i]);
+            flops += CV_BIG_INT(2)*outChannels*karea*total(inputs[i]);
         }
 
         return flops;

modules/dnn/src/layers/layers_common.cpp

@@ -175,11 +175,13 @@ void getPoolingKernelParams(const LayerParams &params, std::vector<size_t>& kern
 }
 
 void getConvolutionKernelParams(const LayerParams &params, std::vector<size_t>& kernel, std::vector<size_t>& pads_begin,
-                                std::vector<size_t>& pads_end, std::vector<size_t>& strides, std::vector<size_t>& dilations, cv::String &padMode)
+                                std::vector<size_t>& pads_end, std::vector<size_t>& strides,
+                                std::vector<size_t>& dilations, cv::String &padMode, std::vector<size_t>& adjust_pads)
 {
     util::getKernelSize(params, kernel);
     util::getStrideAndPadding(params, pads_begin, pads_end, strides, padMode, kernel.size());
     util::getParameter(params, "dilation", "dilation", dilations, true, std::vector<size_t>(kernel.size(), 1));
+    util::getParameter(params, "adj", "adj", adjust_pads, true, std::vector<size_t>(kernel.size(), 0));
 
     for (int i = 0; i < dilations.size(); i++)
         CV_Assert(dilations[i] > 0);

modules/dnn/src/layers/layers_common.hpp

@@ -60,7 +60,8 @@ namespace cv
 namespace dnn
 {
 void getConvolutionKernelParams(const LayerParams &params, std::vector<size_t>& kernel, std::vector<size_t>& pads_begin,
-                                std::vector<size_t>& pads_end, std::vector<size_t>& strides, std::vector<size_t>& dilations, cv::String &padMode);
+                                std::vector<size_t>& pads_end, std::vector<size_t>& strides, std::vector<size_t>& dilations,
+                                cv::String &padMode, std::vector<size_t>& adjust_pads);
 
 void getPoolingKernelParams(const LayerParams &params, std::vector<size_t>& kernel, bool &globalPooling,
                             std::vector<size_t>& pads_begin, std::vector<size_t>& pads_end, std::vector<size_t>& strides, cv::String &padMode);

modules/dnn/src/onnx/onnx_importer.cpp

@@ -682,42 +682,37 @@ void ONNXImporter::populateNet(Net dstNet)
             layerParams.set("num_output", layerParams.blobs[0].size[1] * layerParams.get<int>("group", 1));
             layerParams.set("bias_term", node_proto.input_size() == 3);
 
+            if (!layerParams.has("kernel_size"))
+                CV_Error(Error::StsNotImplemented,
+                         "Required attribute 'kernel_size' is not present.");
+
             if (layerParams.has("output_shape"))
             {
                 const DictValue& outShape = layerParams.get("output_shape");
+                DictValue strides = layerParams.get("stride");
+                DictValue kernel = layerParams.get("kernel_size");
 
-                if (outShape.size() != 4)
-                    CV_Error(Error::StsNotImplemented, "Output shape must have 4 elements.");
-
-                DictValue stride = layerParams.get("stride");
-                const int strideY = stride.getIntValue(0);
-                const int strideX = stride.getIntValue(1);
-                const int outH = outShape.getIntValue(2);
-                const int outW = outShape.getIntValue(3);
-
-                if (layerParams.get<String>("pad_mode") == "SAME")
+                String padMode;
+                std::vector<int> adjust_pads;
+                if (layerParams.has("pad_mode"))
                 {
-                    layerParams.set("adj_w", (outW - 1) % strideX);
-                    layerParams.set("adj_h", (outH - 1) % strideY);
-                }
-                else if (layerParams.get<String>("pad_mode") == "VALID")
-                {
-                    if (!layerParams.has("kernel_size"))
-                        CV_Error(Error::StsNotImplemented,
-                                 "Required attribute 'kernel_size' is not present.");
-
-                    DictValue kernel = layerParams.get("kernel_size");
-                    layerParams.set("adj_h", (outH - kernel.getIntValue(0)) % strideY);
-                    layerParams.set("adj_w", (outW - kernel.getIntValue(1)) % strideX);
+                    padMode = toUpperCase(layerParams.get<String>("pad_mode"));
+                    if (padMode != "SAME" && padMode != "VALID")
+                        CV_Error(Error::StsError, "Unsupported padding mode " + padMode);
+
+                    for (int i = 0; i < strides.size(); i++)
+                    {
+                        int sz = outShape.get<int>(2 + i);
+                        int stride = strides.get<int>(i);
+                        adjust_pads.push_back(padMode == "SAME"? (sz - 1) % stride :
+                                                                 (sz - kernel.get<int>(i)) % stride);
+                    }
+                    layerParams.set("adj", DictValue::arrayInt(&adjust_pads[0], adjust_pads.size()));
                 }
             }
             else if (layerParams.has("output_padding"))
             {
-                const DictValue& adj_pad = layerParams.get("output_padding");
-                if (adj_pad.size() != 2)
-                    CV_Error(Error::StsNotImplemented, "Deconvolution3D layer is not supported");
-                layerParams.set("adj_w", adj_pad.get<int>(1));
-                layerParams.set("adj_h", adj_pad.get<int>(0));
+                replaceLayerParam(layerParams, "output_padding", "adj");
             }
         }
         else if (layer_type == "Transpose")

modules/dnn/test/test_onnx_importer.cpp

@@ -127,6 +127,19 @@ TEST_P(Test_ONNX_layers, Deconvolution)
     testONNXModels("deconv_adjpad_2d", npy, 0, 0, false, false);
 }
 
+TEST_P(Test_ONNX_layers, Deconvolution3D)
+{
+#if defined(INF_ENGINE_RELEASE)
+    applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_2018R5);
+#endif
+    if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU)
+        throw SkipTestException("Only DLIE backend on CPU is supported");
+    testONNXModels("deconv3d");
+    testONNXModels("deconv3d_bias");
+    testONNXModels("deconv3d_pad");
+    testONNXModels("deconv3d_adjpad");
+}
+
 TEST_P(Test_ONNX_layers, Dropout)
 {
     testONNXModels("dropout");