Merge pull request #12519 from l-bat:l-bat/onnx_parser

Support asymmetric padding in pooling layer (#12519)

* Add Inception_V1 support in ONNX

* Add asymmetric padding in OpenCL and Inference engine

* Refactoring

modules/dnn/include/opencv2/dnn/all_layers.hpp

@@ -234,7 +234,9 @@ CV__DNN_EXPERIMENTAL_NS_BEGIN
     {
     public:
         int type;
-        Size kernel, stride, pad;
+        Size kernel, stride;
+        int pad_l, pad_t, pad_r, pad_b;
+        CV_DEPRECATED Size pad;
         bool globalPooling;
         bool computeMaxIdx;
         String padMode;

modules/dnn/src/layers/convolution_layer.cpp

@@ -64,10 +64,17 @@ public:
     BaseConvolutionLayerImpl(const LayerParams &params)
     {
         setParamsFrom(params);
-        getConvolutionKernelParams(params, kernel.height, kernel.width, pad.height,
-                                   pad.width, stride.height, stride.width, dilation.height,
+        int pad_t = 0, pad_l = 0, pad_r = 0, pad_b = 0;
+        getConvolutionKernelParams(params, kernel.height, kernel.width, pad_t,
+                                   pad_l, pad_b, pad_r, stride.height, stride.width, dilation.height,
                                    dilation.width, padMode);
 
+        if (pad_t != pad_b || pad_l != pad_r)
+            CV_Error(Error::StsNotImplemented, "Unsupported asymmetric padding in convolution layer");
+
+        pad.width = pad_l;
+        pad.height = pad_t;
+
         numOutput = params.get<int>("num_output");
         int ngroups = params.get<int>("group", 1);
 
@@ -100,8 +107,18 @@ public:
         }
 
         Size outSize = Size(outputs[0].size[3], outputs[0].size[2]);
+
+        int pad_t = pad.height, pad_l = pad.width, pad_b = pad.height, pad_r = pad.width;
+
         getConvPoolPaddings(Size(input.size[3], input.size[2]), outSize,
-                kernel, stride, padMode, dilation, pad);
+                kernel, stride, padMode, dilation, pad_t, pad_l, pad_b, pad_r);
+
+
+        if (pad_t != pad_b || pad_l != pad_r)
+            CV_Error(Error::StsNotImplemented, "Unsupported asymmetric padding in convolution layer");
+
+        pad.width = pad_l;
+        pad.height = pad_t;
     }
 
     bool hasBias() const
@@ -1156,9 +1173,17 @@ public:
         std::vector<Mat> inputs, outputs;
         inputs_arr.getMatVector(inputs);
         outputs_arr.getMatVector(outputs);
+
+        int pad_t = pad.height, pad_l = pad.width, pad_b = pad.height, pad_r = pad.width;
         getConvPoolPaddings(Size(outputs[0].size[3], outputs[0].size[2]),
                             Size(inputs[0].size[3], inputs[0].size[2]),
-                            kernel, stride, padMode, dilation, pad);
+                            kernel, stride, padMode, dilation, pad_t, pad_l, pad_b, pad_r);
+
+        if (pad_t != pad_b || pad_l != pad_r)
+            CV_Error(Error::StsNotImplemented, "Unsupported asymmetric padding in convolution layer");
+
+        pad.width = pad_l;
+        pad.height = pad_t;
     }
 
     class MatMulInvoker : public ParallelLoopBody

modules/dnn/src/layers/layers_common.cpp

@@ -118,9 +118,19 @@ void getKernelSize(const LayerParams &params, int &kernelH, int &kernelW)
     CV_Assert(kernelH > 0 && kernelW > 0);
 }
 
-void getStrideAndPadding(const LayerParams &params, int &padH, int &padW, int &strideH, int &strideW, cv::String& padMode)
+void getStrideAndPadding(const LayerParams &params, int &padT, int &padL, int &padB, int &padR, int &strideH, int &strideW, cv::String& padMode)
 {
-    util::getParameter(params, "pad", "pad", padH, padW, true, 0);
+    if (params.has("pad_l") && params.has("pad_t") && params.has("pad_r") && params.has("pad_b")) {
+        padT = params.get<int>("pad_t");
+        padL = params.get<int>("pad_l");
+        padB = params.get<int>("pad_b");
+        padR = params.get<int>("pad_r");
+    }
+    else {
+        util::getParameter(params, "pad", "pad", padT, padL, true, 0);
+        padB = padT;
+        padR = padL;
+    }
     util::getParameter(params, "stride", "stride", strideH, strideW, true, 1);
 
     padMode = "";
@@ -129,15 +139,15 @@ void getStrideAndPadding(const LayerParams &params, int &padH, int &padW, int &s
         padMode = params.get<String>("pad_mode");
     }
 
-    CV_Assert(padH >= 0 && padW >= 0 && strideH > 0 && strideW > 0);
+    CV_Assert(padT >= 0 && padL >= 0 && padB >= 0 && padR >= 0 && strideH > 0 && strideW > 0);
 }
 }
 
 
 void getPoolingKernelParams(const LayerParams &params, int &kernelH, int &kernelW, bool &globalPooling,
-                            int &padH, int &padW, int &strideH, int &strideW, cv::String &padMode)
+                            int &padT, int &padL, int &padB, int &padR, int &strideH, int &strideW, cv::String &padMode)
 {
-    util::getStrideAndPadding(params, padH, padW, strideH, strideW, padMode);
+    util::getStrideAndPadding(params, padT, padL, padB, padR, strideH, strideW, padMode);
 
     globalPooling = params.has("global_pooling") &&
                     params.get<bool>("global_pooling");
@@ -148,9 +158,9 @@ void getPoolingKernelParams(const LayerParams &params, int &kernelH, int &kernel
         {
             CV_Error(cv::Error::StsBadArg, "In global_pooling mode, kernel_size (or kernel_h and kernel_w) cannot be specified");
         }
-        if(padH != 0 || padW != 0 || strideH != 1 || strideW != 1)
+        if(padT != 0 || padL != 0 || padB != 0 || padR != 0 || strideH != 1 || strideW != 1)
         {
-            CV_Error(cv::Error::StsBadArg, "In global_pooling mode, pad_h and pad_w must be = 0, and stride_h and stride_w must be = 1");
+            CV_Error(cv::Error::StsBadArg, "In global_pooling mode, pads must be = 0, and stride_h and stride_w must be = 1");
         }
     }
     else
@@ -159,12 +169,11 @@ void getPoolingKernelParams(const LayerParams &params, int &kernelH, int &kernel
     }
 }
 
-void getConvolutionKernelParams(const LayerParams &params, int &kernelH, int &kernelW, int &padH, int &padW,
+void getConvolutionKernelParams(const LayerParams &params, int &kernelH, int &kernelW, int &padT, int &padL, int &padB, int &padR,
                                 int &strideH, int &strideW, int &dilationH, int &dilationW, cv::String &padMode)
 {
     util::getKernelSize(params, kernelH, kernelW);
-    util::getStrideAndPadding(params, padH, padW, strideH, strideW, padMode);
-
+    util::getStrideAndPadding(params, padT, padL, padB, padR, strideH, strideW, padMode);
     util::getParameter(params, "dilation", "dilation", dilationH, dilationW, true, 1);
 
     CV_Assert(dilationH > 0 && dilationW > 0);
@@ -201,11 +210,11 @@ void getConvPoolOutParams(const Size& inp, const Size &kernel,
 
 void getConvPoolPaddings(const Size& inp, const Size& out,
                          const Size &kernel, const Size &stride,
-                         const String &padMode, const Size &dilation, Size &pad)
+                         const String &padMode, const Size &dilation, int &padT, int &padL, int &padB, int &padR)
 {
     if (padMode == "VALID")
     {
-        pad = cv::Size(0,0);
+        padT = padL = padB = padR = 0;
     }
     else if (padMode == "SAME")
     {
@@ -213,7 +222,8 @@ void getConvPoolPaddings(const Size& inp, const Size& out,
         int Pw = std::max(0, (out.width - 1) * stride.width + (dilation.width * (kernel.width - 1) + 1) - inp.width);
         // For odd values of total padding, add more padding at the 'right'
         // side of the given dimension.
-        pad = cv::Size(Pw / 2, Ph / 2);
+        padT= padB = Ph / 2;
+        padL = padR = Pw / 2;
     }
 }
 

modules/dnn/src/layers/layers_common.hpp

@@ -60,19 +60,20 @@ namespace cv
 namespace dnn
 {
 
-void getConvolutionKernelParams(const LayerParams &params, int &kernelH, int &kernelW, int &padH, int &padW,
+void getConvolutionKernelParams(const LayerParams &params, int &kernelH, int &kernelW, int &padT, int &padL, int &padB, int &padR,
                                 int &strideH, int &strideW, int &dilationH, int &dilationW, cv::String& padMode);
 
 void getPoolingKernelParams(const LayerParams &params, int &kernelH, int &kernelW, bool &globalPooling,
-                            int &padH, int &padW, int &strideH, int &strideW, cv::String& padMode);
+                            int &padT, int &padL, int &padB, int &padR, int &strideH, int &strideW, cv::String& padMode);
 
 void getConvPoolOutParams(const Size& inp, const Size &kernel,
                           const Size &stride, const String &padMode,
                           const Size &dilation, Size& out);
 
+
 void getConvPoolPaddings(const Size& inp, const Size& out,
                          const Size &kernel, const Size &stride,
-                         const String &padMode, const Size &dilation, Size &pad);
+                         const String &padMode, const Size &dilation, int &padT, int &padL, int &padB, int &padR);
 
 }
 }

modules/dnn/src/ocl4dnn/include/ocl4dnn.hpp

@@ -345,7 +345,7 @@ struct OCL4DNNPoolConfig
 {
     OCL4DNNPoolConfig() :
         kernel(1, 1),
-        pad(0, 0),
+        pad_l(0), pad_t(0), pad_r(0), pad_b(0),
         stride(1, 1),
         dilation(1, 1),
         channels(0),
@@ -358,7 +358,7 @@ struct OCL4DNNPoolConfig
     MatShape in_shape;
     MatShape out_shape;
     Size kernel;
-    Size pad;
+    int pad_l, pad_t, pad_r, pad_b;
     Size stride;
     Size dilation;
 
@@ -381,7 +381,6 @@ class OCL4DNNPool
                      UMat& top_mask);
     private:
         // Pooling parameters
-        std::vector<int32_t> pad_;
         std::vector<int32_t> stride_;
         std::vector<int32_t> kernel_shape_;
         std::vector<int32_t> im_in_shape_;
@@ -394,8 +393,10 @@ class OCL4DNNPool
         int32_t kernel_w_;
         int32_t stride_h_;
         int32_t stride_w_;
-        int32_t pad_h_;
-        int32_t pad_w_;
+        int32_t pad_t_;
+        int32_t pad_l_;
+        int32_t pad_b_;
+        int32_t pad_r_;
         int32_t height_;
         int32_t width_;
         int32_t pooled_height_;

modules/dnn/src/ocl4dnn/src/ocl4dnn_pool.cpp

@@ -62,7 +62,6 @@ OCL4DNNPool<Dtype>::OCL4DNNPool(OCL4DNNPoolConfig config)
     for (int i = 0; i < spatial_dims; ++i)
     {
         kernel_shape_.push_back(i == 0 ? config.kernel.height : config.kernel.width);
-        pad_.push_back(i == 0 ? config.pad.height : config.pad.width);
         stride_.push_back(i == 0 ? config.stride.height : config.stride.width);
         im_in_shape_.push_back(config.in_shape[dims - spatial_dims + i]);
         im_out_shape_.push_back(config.out_shape[dims - spatial_dims + i]);
@@ -72,8 +71,10 @@ OCL4DNNPool<Dtype>::OCL4DNNPool(OCL4DNNPoolConfig config)
     kernel_w_ = kernel_shape_[1];
     stride_h_ = stride_[0];
     stride_w_ = stride_[1];
-    pad_h_ = pad_[0];
-    pad_w_ = pad_[1];
+    pad_t_ = config.pad_t;
+    pad_l_ = config.pad_l;
+    pad_r_ = config.pad_r;
+    pad_b_ = config.pad_b;
     height_ = im_in_shape_[0];
     width_ = im_in_shape_[1];
     pooled_height_ = im_out_shape_[0];
@@ -113,14 +114,13 @@ bool OCL4DNNPool<Dtype>::Forward(const UMat& bottom,
                 ocl::dnn::ocl4dnn_pooling_oclsrc,
                 format(" -D Dtype=%s -D KERNEL_MAX_POOL=1 -D KERNEL_W=%d -D KERNEL_H=%d"
                        " -D STRIDE_W=%d -D STRIDE_H=%d"
-                       " -D PAD_W=%d -D PAD_H=%d%s",
+                       " -D PAD_L=%d -D PAD_T=%d -D PAD_R=%d -D PAD_B=%d%s",
                        (use_half) ? "half" : "float",
                        kernel_w_, kernel_h_,
                        stride_w_, stride_h_,
-                       pad_w_, pad_h_,
+                       pad_l_, pad_t_, pad_r_, pad_b_,
                        computeMaxIdx ? " -D HAVE_MASK=1" : ""
                 ));
-
             if (oclk_max_pool_forward.empty())
                 return false;
 
@@ -150,11 +150,11 @@ bool OCL4DNNPool<Dtype>::Forward(const UMat& bottom,
                 ocl::dnn::ocl4dnn_pooling_oclsrc,
                 format(" -D Dtype=%s -D KERNEL_AVE_POOL=1 -D KERNEL_W=%d -D KERNEL_H=%d"
                        " -D STRIDE_W=%d -D STRIDE_H=%d"
-                       " -D PAD_W=%d -D PAD_H=%d%s",
+                       " -D PAD_L=%d -D PAD_T=%d -D PAD_R=%d -D PAD_B=%d%s",
                        (use_half) ? "half" : "float",
                        kernel_w_, kernel_h_,
                        stride_w_, stride_h_,
-                       pad_w_, pad_h_,
+                       pad_l_, pad_t_, pad_r_, pad_b_,
                        avePoolPaddedArea ? " -D AVE_POOL_PADDING_AREA" : ""
                 ));
 

modules/dnn/src/onnx/onnx_importer.cpp

@@ -174,9 +174,8 @@ LayerParams ONNXImporter::getLayerParams(const opencv_onnx::NodeProto& node_prot
         else if(attribute_name == "pads")
         {
             CV_Assert(attribute_proto.ints_size() == 4);
-            lp.set("pad_h", saturate_cast<int32_t>(attribute_proto.ints(0)));
-            lp.set("pad_w", saturate_cast<int32_t>(attribute_proto.ints(1)));
-            // push pad_b and pad_r for compute ceil_mode
+            lp.set("pad_t", saturate_cast<int32_t>(attribute_proto.ints(0)));
+            lp.set("pad_l", saturate_cast<int32_t>(attribute_proto.ints(1)));
             lp.set("pad_b", saturate_cast<int32_t>(attribute_proto.ints(2)));
             lp.set("pad_r", saturate_cast<int32_t>(attribute_proto.ints(3)));
         }
@@ -306,6 +305,7 @@ void ONNXImporter::populateNet(Net dstNet)
         std::string layer_type = node_proto.op_type();
         layerParams.type = layer_type;
 
+
         if (layer_type == "MaxPool")
         {
             layerParams.type = "Pooling";
@@ -551,7 +551,6 @@ void ONNXImporter::populateNet(Net dstNet)
 
          for (int j = 0; j < node_proto.input_size(); j++) {
              layerId = layer_id.find(node_proto.input(j));
-
              if (layerId != layer_id.end()) {
                  dstNet.connect(layerId->second.layerId, layerId->second.outputId, id, j);
              }

modules/dnn/src/opencl/ocl4dnn_pooling.cl

@@ -73,8 +73,8 @@ __kernel void
   const int xx = index / pooled_width;
   const int ph = xx % pooled_height;
   const int ch = xx / pooled_height;
-  int hstart = ph * STRIDE_H - PAD_H;
-  int wstart = pw * STRIDE_W - PAD_W;
+  int hstart = ph * STRIDE_H - PAD_T;
+  int wstart = pw * STRIDE_W - PAD_L;
   Dtype maxval = -FLT_MAX;
   int maxidx = -1;
   int in_offset = ch * height * width;
@@ -117,10 +117,10 @@ __kernel void TEMPLATE(ave_pool_forward, Dtype)(
   const int xx = index / pooled_width;
   const int ph = xx % pooled_height;
   const int ch = xx / pooled_height;
-  int hstart = ph * STRIDE_H - PAD_H;
-  int wstart = pw * STRIDE_W - PAD_W;
-  int hend = min(hstart + KERNEL_H, height + PAD_H);
-  int wend = min(wstart + KERNEL_W, width + PAD_W);
+  int hstart = ph * STRIDE_H - PAD_T;
+  int wstart = pw * STRIDE_W - PAD_L;
+  int hend = min(hstart + KERNEL_H, height + PAD_B);
+  int wend = min(wstart + KERNEL_W, width + PAD_R);
   int pool_size;
 #ifdef AVE_POOL_PADDING_AREA
   pool_size = (hend - hstart) * (wend - wstart);

modules/dnn/src/opencl/pooling.cl

@@ -27,7 +27,7 @@
 __kernel void MaxPoolForward(const int nthreads,
     __global T* bottom_data, const int num, const int channels, const int height, const int width,
     const int pooled_height, const int pooled_width, const int kernel_h, const int kernel_w,
-    const int stride_h, const int stride_w, const int pad_h, const int pad_w,
+    const int stride_h, const int stride_w, const int pad_t, const int pad_l, const int pad_b, const int pad_r,
     __global T* top_data
 #ifdef MASK
     , __global float* mask
@@ -41,8 +41,8 @@ __kernel void MaxPoolForward(const int nthreads,
     int ph = (index / pooled_width) % pooled_height;
     int c = (index / pooled_width / pooled_height) % channels;
     int n = index / pooled_width / pooled_height / channels;
-    int hstart = ph * stride_h - pad_h;
-    int wstart = pw * stride_w - pad_w;
+    int hstart = ph * stride_h - pad_t;
+    int wstart = pw * stride_w - pad_l;
     const int hend = min(hstart + kernel_h, height);
     const int wend = min(wstart + kernel_w, width);
     hstart = max(hstart, 0);
@@ -71,7 +71,7 @@ __kernel void MaxPoolForward(const int nthreads,
 __kernel void AvePoolForward(const int nthreads,
     __global T* bottom_data, const int num, const int channels, const int height, const int width,
     const int pooled_height, const int pooled_width, const int kernel_h, const int kernel_w,
-    const int stride_h, const int stride_w, const int pad_h, const int pad_w,
+    const int stride_h, const int stride_w, const int pad_t, const int pad_l, const int pad_b, const int pad_r,
     __global T* top_data
 #ifdef MASK
     , __global float* mask // NOT USED
@@ -84,9 +84,9 @@ __kernel void AvePoolForward(const int nthreads,
     int pw = index % pooled_width;
     int ph = (index / pooled_width) % pooled_height;
     int c = (index / pooled_width / pooled_height) % channels;
-    int n = index / pooled_width / pooled_height / channels; int hstart = ph * stride_h - pad_h; int wstart = pw * stride_w - pad_w;
-    int hend = min(hstart + kernel_h, height + pad_h);
-    int wend = min(wstart + kernel_w, width + pad_w);
+    int n = index / pooled_width / pooled_height / channels; int hstart = ph * stride_h - pad_t; int wstart = pw * stride_w - pad_l;
+    int hend = min(hstart + kernel_h, height + pad_b);
+    int wend = min(wstart + kernel_w, width + pad_r);
     const int pool_size = (hend - hstart) * (wend - wstart);
     hstart = max(hstart, 0);
     wstart = max(wstart, 0);

modules/dnn/test/test_onnx_importer.cpp

@@ -346,6 +346,10 @@ TEST_P(Test_ONNX_nets, DenseNet121)
     testONNXModels("densenet121", pb, l1, lInf);
 }
 
+TEST_P(Test_ONNX_nets, Inception_v1)
+{
+    testONNXModels("inception_v1", pb);
+}
 
 INSTANTIATE_TEST_CASE_P(/**/, Test_ONNX_nets, dnnBackendsAndTargets());