Added ability to remove singleton dimensions in Reshape layer.

Updated torch importer

modules/dnn/include/opencv2/dnn/blob.hpp

@@ -51,6 +51,9 @@ namespace dnn
     {
     public:
         explicit Blob();
+
+        explicit Blob(const BlobShape &shape, int type = CV_32F);
+
         /** @brief constucts 4-dimensional blob from input
          *  @param in 2-dimensional or 3-dimensional single-channel image (or vector from them)
          *  @param dstCn if specified force size of ouptut blob channel-dimension

modules/dnn/include/opencv2/dnn/blob.inl.hpp

@@ -7,15 +7,16 @@ namespace cv
 namespace dnn
 {
 
-inline BlobShape::BlobShape(int ndims, int fill) : sz( (size_t)std::max(ndims, 1) )
+inline BlobShape::BlobShape(int ndims, int fill) : sz( (size_t)std::max(ndims, 0) )
 {
+    CV_Assert(ndims >= 0);
     for (int i = 0; i < ndims; i++)
         sz[i] = fill;
 }
 
-inline BlobShape::BlobShape(int ndims, const int *sizes) : sz( (size_t)std::max(ndims, 1) )
+inline BlobShape::BlobShape(int ndims, const int *sizes) : sz( (size_t)std::max(ndims, 0) )
 {
-    CV_Assert(ndims > 0);
+    CV_Assert(ndims >= 0);
     for (int i = 0; i < ndims; i++)
         sz[i] = sizes[i];
 }
@@ -30,7 +31,6 @@ inline BlobShape::BlobShape(int num, int cn, int rows, int cols) : sz(4)
 
 inline BlobShape::BlobShape(const std::vector<int> &sizes) : sz( sizes.size() )
 {
-    CV_Assert(sizes.size() > 0);
     for (int i = 0; i < (int)sizes.size(); i++)
         sz[i] = sizes[i];
 }
@@ -81,7 +81,8 @@ inline int &BlobShape::operator[] (int axis)
 
 inline ptrdiff_t BlobShape::total()
 {
-    CV_Assert(dims() >= 1);
+    if (dims() == 0)
+        return 0;
 
     ptrdiff_t res = 1;
     for (int i = 0; i < dims(); i++)
@@ -89,7 +90,6 @@ inline ptrdiff_t BlobShape::total()
     return res;
 }
 
-
 inline const int *BlobShape::ptr() const
 {
     return sz;
@@ -119,12 +119,7 @@ inline bool operator== (const BlobShape &l, const BlobShape &r)
 inline int Blob::canonicalAxis(int axis) const
 {
     CV_Assert(-dims() <= axis && axis < dims());
-
-    if (axis < 0)
-    {
-        return dims() + axis;
-    }
-    return axis;
+    return (axis < 0) ? axis + dims() : axis;
 }
 
 inline int Blob::dims() const

modules/dnn/src/blob.cpp

@@ -114,6 +114,11 @@ namespace dnn
         }
     }
 
+    Blob::Blob(const BlobShape &shape, int type)
+    {
+        this->create(shape, type);
+    }
+
     void Blob::fill(const BlobShape &shape, int type, void *data, bool deepCopy)
     {
         CV_Assert(type == CV_32F || type == CV_64F);
@@ -129,14 +134,6 @@ namespace dnn
         }
     }
 
-    void Blob::fill(InputArray in)
-    {
-        CV_Assert(in.isMat() || in.isMatVector());
-
-        //TODO
-        *this = Blob(in);
-    }
-
     void Blob::create(const BlobShape &shape, int type)
     {
         CV_Assert(type == CV_32F || type == CV_64F);

modules/dnn/src/dnn.cpp

@@ -309,7 +309,7 @@ struct Net::Impl
         {
             LayerPin from = ld.inputBlobsId[i];
             CV_Assert(from.valid());
-            CV_Assert(layers.count(from.lid) && layers[from.lid].outputBlobs.size() > from.oid);
+            CV_DbgAssert(layers.count(from.lid) && (int)layers[from.lid].outputBlobs.size() > from.oid);
             ld.inputBlobs[i] = &layers[from.lid].outputBlobs[from.oid];
         }
 

modules/dnn/src/layers/reshape_layer.cpp

@@ -25,9 +25,20 @@ protected:
 
 ReshapeLayer::ReshapeLayer(LayerParams &params)
 {
+    inAxis = params.get<int>("axis", 0);
+    inNumAxes = params.get<int>("num_axes", -1);
+    CV_Assert(inNumAxes >= -1);
+
+    autoAxisIdx = -1;
+
+    if (!params.has("dim"))
+    {
+        shapeDesc = BlobShape(0);
+        return;
+    }
+
     DictValue paramShape = params.get("dim");
     shapeDesc = BlobShape(paramShape.size());
-    autoAxisIdx = -1;
 
     for (int i = 0; i < paramShape.size(); i++)
     {
@@ -43,19 +54,16 @@ ReshapeLayer::ReshapeLayer(LayerParams &params)
 
         shapeDesc[i] = dim;
     }
-
-    inAxis = params.get<int>("axis", 0);
-    inNumAxes = params.get<int>("num_axes", -1);
-    CV_Assert(inNumAxes >= -1);
 }
 
 void ReshapeLayer::allocate(const std::vector<Blob*> &inputs, std::vector<Blob> &outputs)
 {
-    CV_Assert(inputs.size() == 1);
-    outputs.resize(1);
+    outputs.resize(inputs.size());
 
-    Blob &inpBlob = *inputs[0];
-    Blob &outBlob = outputs[0];
+    for (size_t i = 0; i < inputs.size(); i++)
+    {
+        Blob &inpBlob = *inputs[i];
+        Blob &outBlob = outputs[i];
         BlobShape inpShape = inpBlob.shape();
 
         int startAxis = (inAxis >= 0) ? inAxis : inpShape.dims() + 1 + inAxis;
@@ -70,6 +78,7 @@ void ReshapeLayer::allocate(const std::vector<Blob*> &inputs, std::vector<Blob>
 
         outBlob.shareFrom(inpBlob);
         outBlob.reshape(outShape);
+    }
 }
 
 void ReshapeLayer::computeOutputShape(int startAxis, int endAxis, BlobShape &inpShape, BlobShape &outShape)
@@ -84,7 +93,7 @@ void ReshapeLayer::computeOutputShape(int startAxis, int endAxis, BlobShape &inp
         {
             int inpAxisIdx = startAxis + i;
             if (inpAxisIdx < 0 || inpShape.dims() <= inpAxisIdx)
-                CV_Error(Error::StsOutOfRange, "new shape contains a 0, but there was no corresponding bottom axis to copy");
+                CV_Error(Error::StsOutOfRange, "copy dimension (which has zero size) is not presented into reshaped blob");
             outShape[idx++] = inpShape[startAxis + i];
         }
         else
@@ -113,7 +122,7 @@ void ReshapeLayer::computeOutputShape(int startAxis, int endAxis, BlobShape &inp
 
     if (inpShape.total() != outShape.total())
     {
-        CV_Error(Error::StsBadArg, "Mismatch between input and output blob elements count");
+        CV_Error(Error::StsUnmatchedSizes, "Mismatch between input and output blob elements count");
     }
 }
 

modules/dnn/src/torch/THDiskFile.cpp

@@ -1,3 +1,4 @@
+#if defined(ENABLE_TORCH_IMPORTER) && ENABLE_TORCH_IMPORTER
 #include "THGeneral.h"
 #include "THDiskFile.h"
 #include "THFilePrivate.h"
@@ -607,3 +608,4 @@ THFile *THPipeFile_new(const char *name, const char *mode, int isQuiet)
 }
 
 }
+#endif

modules/dnn/src/torch/THFile.cpp

@@ -1,3 +1,4 @@
+#if defined(ENABLE_TORCH_IMPORTER) && ENABLE_TORCH_IMPORTER
 #include "THFile.h"
 #include "THFilePrivate.h"
 
@@ -159,3 +160,4 @@ IMPLEMENT_THFILE_STORAGE(Double, double)
 */
 
 }
+#endif

modules/dnn/src/torch/THGeneral.cpp

@@ -1,4 +1,6 @@
+#if defined(ENABLE_TORCH_IMPORTER) && ENABLE_TORCH_IMPORTER
 #include "THGeneral.h"
+#include <opencv2/core.hpp>
 
 extern "C"
 {
@@ -15,10 +17,9 @@ extern "C"
 #endif
 
 /* Torch Error Handling */
-static void defaultTorchErrorHandlerFunction(const char *msg, void *data)
+static void defaultTorchErrorHandlerFunction(const char *msg, void*)
 {
-  printf("$ Error: %s\n", msg);
-  exit(-1);
+  CV_Error(cv::Error::StsError, cv::String("Torch Error: ") + msg);
 }
 
 static __thread void (*torchErrorHandlerFunction)(const char *msg, void *data) = defaultTorchErrorHandlerFunction;
@@ -61,13 +62,12 @@ void THSetErrorHandler( void (*torchErrorHandlerFunction_)(const char *msg, void
 }
 
 /* Torch Arg Checking Handling */
-static void defaultTorchArgErrorHandlerFunction(int argNumber, const char *msg, void *data)
+static void defaultTorchArgErrorHandlerFunction(int argNumber, const char *msg, void*)
 {
   if(msg)
-    printf("$ Invalid argument %d: %s\n", argNumber, msg);
+    CV_Error(cv::Error::StsError, cv::format("Torch invalid argument %d: %s", argNumber, msg));
   else
-    printf("$ Invalid argument %d\n", argNumber);
-  exit(-1);
+    CV_Error(cv::Error::StsError, cv::format("Invalid argument %d", argNumber));
 }
 
 static __thread void (*torchArgErrorHandlerFunction)(int argNumber, const char *msg, void *data) = defaultTorchArgErrorHandlerFunction;
@@ -252,3 +252,4 @@ double THLog1p(const double x)
 }
 
 }
+#endif

modules/dnn/src/torch/torch_importer.cpp

@@ -8,7 +8,7 @@
 namespace cv {
 namespace dnn {
 
-#if ENABLE_TORCH_IMPORTER || 1
+#if defined(ENABLE_TORCH_IMPORTER) && ENABLE_TORCH_IMPORTER
 #include "THDiskFile.h"
 
 enum LuaType
@@ -575,32 +575,44 @@ struct TorchImporter : public ::cv::dnn::Importer
                 }
                 return prevLayerId;
             }
-            else if (module->thName == "Parallel" || module->thName == "Concat")
+            else if (module->thName == "Concat")
             {
-                int splitId, mergeId, newId;
+                int newId, splitId, mergeId;
+                LayerParams mergeParams, splitParams;
+                mergeParams.set("axis", module->params.get<int>("dimension") - 1);
+
+                splitId = net.addLayer(generateLayerName("torchSplit"), "Split", splitParams);
+                mergeId = net.addLayer(generateLayerName("torchMerge"), "Concat", mergeParams);
+                net.connect(prevLayerId, prevOutNum, splitId, 0);
 
-                String splitType;
-                LayerParams splitParams, mergeParams;
-                if (module->thName == "Parallel")
+                for (int i = 0; i < (int)module->modules.size(); i++)
                 {
-                    splitType = "Slice";
-                    splitParams.set("axis", module->params.get<int>("inputDimension") - 1);
-                    mergeParams.set("axis", module->params.get<int>("outputDimension") - 1);
+                    newId = fill(module->modules[i], splitId, i);
+                    net.connect(newId, 0, mergeId, i);
                 }
-                else
-                {
-                    splitType = "Split";
-                    mergeParams.set("axis", module->params.get<int>("dimension") - 1);
+
+                return mergeId;
             }
+            else if (module->thName == "Parallel")
+            {
+                int newId, splitId, mergeId, reshapeId;
 
-                splitId = net.addLayer(generateLayerName("torchSplit"), splitType, splitParams);
+                LayerParams splitParams, mergeParams, reshapeParams;
+                splitParams.set("axis", module->params.get<int>("inputDimension") - 1);
+                mergeParams.set("axis", module->params.get<int>("outputDimension") - 1);
+                reshapeParams.set("axis", splitParams.get<int>("axis"));
+                reshapeParams.set("num_axes", 1);
+
+                splitId = net.addLayer(generateLayerName("torchSplit"), "Slice", splitParams);
                 mergeId = net.addLayer(generateLayerName("torchMerge"), "Concat", mergeParams);
+                reshapeId = net.addLayer(generateLayerName("torchReshape"), "Reshape", reshapeParams);
                 net.connect(prevLayerId, prevOutNum, splitId, 0);
 
-                for (size_t i = 0; i < module->modules.size(); i++)
+                for (int i = 0; i < (int)module->modules.size(); i++)
                 {
-                    newId = fill(module->modules[i], splitId, (int)i);
-                    net.connect(newId, 0, mergeId, (int)i);
+                    net.connect(splitId, i, reshapeId, i);
+                    newId = fill(module->modules[i], reshapeId, i);
+                    net.connect(newId, 0, mergeId, i);
                 }
 
                 return mergeId;

modules/dnn/test/npy_blob.hpp

@@ -8,7 +8,7 @@ inline cv::dnn::Blob blobFromNPY(const cv::String &path)
     cnpy::NpyArray npyBlob = cnpy::npy_load(path.c_str());
     cv::dnn::BlobShape shape((int)npyBlob.shape.size(), (int*)&npyBlob.shape[0]);
 
-    cv::dnn::Blob blob;
+    cv::dnn::Blob blob(shape);
     blob.fill(shape, CV_32F, npyBlob.data);
 
     npyBlob.destruct();

modules/dnn/test/test_layers.cpp

@@ -75,6 +75,23 @@ TEST(Layer_LRN_channels_Test, Accuracy)
      testLayer("lrn_channels.prototxt");
 }
 
+TEST(Layer_Reshape_squeeze, Accuracy)
+{
+    LayerParams params;
+    params.set("axis", 2);
+    params.set("num_axes", 1);
+
+    Blob inp(BlobShape(4, 3, 1, 2));
+    std::vector<Blob*> inpVec(1, &inp);
+    std::vector<Blob> outVec;
+
+    Ptr<Layer> rl = LayerRegister::createLayerInstance("Reshape", params);
+    rl->allocate(inpVec, outVec);
+    rl->forward(inpVec, outVec);
+
+    EXPECT_EQ(outVec[0].shape(), BlobShape(Vec3i(4, 3, 2)));
+}
+
 TEST(Layer_Reshape_Split_Slice_Test, Accuracy)
 {
     Net net;

modules/dnn/test/test_torch_importer.cpp

@@ -1,5 +1,5 @@
-#if 1 || defined(ENABLE_TORCH_IMPORTER) && ENABLE_TORCH_IMPORTER
-#if 1 || defined(ENABLE_TORCH_TESTS) && ENABLE_TORCH_TESTS
+#if defined(ENABLE_TORCH_IMPORTER) && ENABLE_TORCH_IMPORTER
+#if defined(ENABLE_TORCH_TESTS) && ENABLE_TORCH_TESTS
 #include "test_precomp.hpp"
 
 namespace cvtest
@@ -31,10 +31,8 @@ static void runTorchNet(String prefix, String outLayerName, bool isBinary)
     String suffix = (isBinary) ? ".dat" : ".txt";
 
     Net net;
-    Ptr<Importer> importer;
-    ASSERT_NO_THROW( importer = createTorchImporter(_tf(prefix + "_net" + suffix), isBinary) );
+    Ptr<Importer> importer = createTorchImporter(_tf(prefix + "_net" + suffix), isBinary);
     ASSERT_TRUE(importer != NULL);
-    //ASSERT_NO_THROW( importer->populateNet(net) );
     importer->populateNet(net);
 
     Blob inp, outRef;
@@ -45,10 +43,6 @@ static void runTorchNet(String prefix, String outLayerName, bool isBinary)
     net.forward();
     Blob out = net.getBlob(outLayerName);
 
-    std::cout << "inp " << inp.shape() << "\n";
-    std::cout << "out " << out.shape() << "\n";
-    std::cout << "ref " << outRef.shape() << "\n";
-
     normAssert(outRef, out);
 }
 
@@ -81,8 +75,7 @@ TEST(Torch_Importer, run_linear)
 
 TEST(Torch_Importer, run_paralel)
 {
-    //TODO: fix and add Reshape
-    //runTorchNet("net_parallel", "l2_torchMerge", false);
+    runTorchNet("net_parallel", "l2_torchMerge", false);
 }
 
 TEST(Torch_Importer, run_concat)