Merge pull request #11461 from dkurt:dnn_reduce_mem_consumption

modules/dnn/src/caffe/caffe_importer.cpp

@@ -250,16 +250,13 @@ public:
         blobShapeFromProto(pbBlob, shape);
 
         dstBlob.create((int)shape.size(), &shape[0], CV_32F);
-        float *dstData = dstBlob.ptr<float>();
         if (pbBlob.data_size())
         {
             // Single precision floats.
             CV_Assert(pbBlob.data_size() == (int)dstBlob.total());
 
             CV_DbgAssert(pbBlob.GetDescriptor()->FindFieldByLowercaseName("data")->cpp_type() == FieldDescriptor::CPPTYPE_FLOAT);
-
+            Mat(dstBlob.dims, &dstBlob.size[0], CV_32F, (void*)pbBlob.data().data()).copyTo(dstBlob);
-            for (int i = 0; i < pbBlob.data_size(); i++)
-                dstData[i] = pbBlob.data(i);
         }
         else
         {
@@ -288,11 +285,18 @@ public:
         if (li == netBinary.layer_size() || netBinary.layer(li).blobs_size() == 0)
             return;
 
-        const caffe::LayerParameter &binLayer = netBinary.layer(li);
+        caffe::LayerParameter* binLayer = netBinary.mutable_layer(li);
-        layerParams.blobs.resize(binLayer.blobs_size());
+        const int numBlobs = binLayer->blobs_size();
-        for (int bi = 0; bi < binLayer.blobs_size(); bi++)
+        layerParams.blobs.resize(numBlobs);
+        for (int bi = 0; bi < numBlobs; bi++)
         {
-            blobFromProto(binLayer.blobs(bi), layerParams.blobs[bi]);
+            blobFromProto(binLayer->blobs(bi), layerParams.blobs[bi]);
+        }
+        binLayer->clear_blobs();
+        CV_Assert(numBlobs == binLayer->blobs().ClearedCount());
+        for (int bi = 0; bi < numBlobs; bi++)
+        {
+            delete binLayer->mutable_blobs()->ReleaseCleared();
         }
     }
 

modules/dnn/src/caffe/caffe_io.cpp

@@ -132,7 +132,7 @@ void UpgradeV0PaddingLayers(const NetParameter& param,
                             NetParameter* param_upgraded_pad);
 
 // Upgrade a single V0LayerConnection to the V1LayerParameter format.
-bool UpgradeV0LayerParameter(const V1LayerParameter& v0_layer_connection,
+bool UpgradeV0LayerParameter(V1LayerParameter* v0_layer_connection,
                              V1LayerParameter* layer_param);
 
 V1LayerParameter_LayerType UpgradeV0LayerType(const string& type);
@@ -149,9 +149,9 @@ bool NetNeedsV1ToV2Upgrade(const NetParameter& net_param);
 
 // Perform all necessary transformations to upgrade a NetParameter with
 // deprecated V1LayerParameters.
-bool UpgradeV1Net(const NetParameter& v1_net_param, NetParameter* net_param);
+bool UpgradeV1Net(NetParameter* net_param);
 
-bool UpgradeV1LayerParameter(const V1LayerParameter& v1_layer_param,
+bool UpgradeV1LayerParameter(V1LayerParameter* v1_layer_param,
                              LayerParameter* layer_param);
 
 const char* UpgradeV1LayerType(const V1LayerParameter_LayerType type);
@@ -194,7 +194,7 @@ bool UpgradeV0Net(const NetParameter& v0_net_param_padding_layers,
     net_param->set_name(v0_net_param.name());
   }
   for (int i = 0; i < v0_net_param.layers_size(); ++i) {
-    is_fully_compatible &= UpgradeV0LayerParameter(v0_net_param.layers(i),
+    is_fully_compatible &= UpgradeV0LayerParameter(v0_net_param.mutable_layers(i),
                                                    net_param->add_layers());
   }
   for (int i = 0; i < v0_net_param.input_size(); ++i) {
@@ -268,8 +268,10 @@ void UpgradeV0PaddingLayers(const NetParameter& param,
   }
 }
 
-bool UpgradeV0LayerParameter(const V1LayerParameter& v0_layer_connection,
+bool UpgradeV0LayerParameter(V1LayerParameter* v0_layer_connection_,
                              V1LayerParameter* layer_param) {
+  CV_Assert(v0_layer_connection_ != NULL);
+  const V1LayerParameter& v0_layer_connection = *v0_layer_connection_;
   bool is_fully_compatible = true;
   layer_param->Clear();
   for (int i = 0; i < v0_layer_connection.bottom_size(); ++i) {
@@ -287,9 +289,7 @@ bool UpgradeV0LayerParameter(const V1LayerParameter& v0_layer_connection,
     if (v0_layer_param.has_type()) {
       layer_param->set_type(UpgradeV0LayerType(type));
     }
-    for (int i = 0; i < v0_layer_param.blobs_size(); ++i) {
+    layer_param->mutable_blobs()->Swap(v0_layer_connection_->mutable_blobs());
-      layer_param->add_blobs()->CopyFrom(v0_layer_param.blobs(i));
-    }
     for (int i = 0; i < v0_layer_param.blobs_lr_size(); ++i) {
       layer_param->add_blobs_lr(v0_layer_param.blobs_lr(i));
     }
@@ -770,8 +770,7 @@ bool UpgradeNetAsNeeded(const string& param_file, NetParameter* param) {
   if (NetNeedsV1ToV2Upgrade(*param)) {
     LOG(ERROR) << "Attempting to upgrade input file specified using deprecated "
                << "V1LayerParameter: " << param_file;
-    NetParameter original_param(*param);
+    if (!UpgradeV1Net(param)) {
-    if (!UpgradeV1Net(original_param, param)) {
       success = false;
       LOG(ERROR) << "Warning: had one or more problems upgrading "
           << "V1LayerParameter (see above); continuing anyway.";
@@ -791,23 +790,24 @@ bool UpgradeNetAsNeeded(const string& param_file, NetParameter* param) {
   return success;
 }
 
-bool UpgradeV1Net(const NetParameter& v1_net_param, NetParameter* net_param) {
+bool UpgradeV1Net(NetParameter* net_param) {
+  // V1LayerParameter layers -> LayerParameter layer
+  CV_Assert(net_param != NULL);
   bool is_fully_compatible = true;
-  if (v1_net_param.layer_size() > 0) {
+  if (net_param->layer_size() > 0) {
     LOG(ERROR) << "Input NetParameter to be upgraded already specifies 'layer' "
                << "fields; these will be ignored for the upgrade.";
     is_fully_compatible = false;
   }
-  net_param->CopyFrom(v1_net_param);
-  net_param->clear_layers();
   net_param->clear_layer();
-  for (int i = 0; i < v1_net_param.layers_size(); ++i) {
+  for (int i = 0; i < net_param->layers_size(); ++i) {
-    if (!UpgradeV1LayerParameter(v1_net_param.layers(i),
+    if (!UpgradeV1LayerParameter(net_param->mutable_layers(i),
                                  net_param->add_layer())) {
       LOG(ERROR) << "Upgrade of input layer " << i << " failed.";
       is_fully_compatible = false;
     }
   }
+  net_param->clear_layers();
   return is_fully_compatible;
 }
 
@@ -834,8 +834,10 @@ void UpgradeNetBatchNorm(NetParameter* net_param) {
   }
 }
 
-bool UpgradeV1LayerParameter(const V1LayerParameter& v1_layer_param,
+bool UpgradeV1LayerParameter(V1LayerParameter* v1_layer_param_,
                              LayerParameter* layer_param) {
+  CV_Assert(v1_layer_param_ != NULL);
+  const V1LayerParameter& v1_layer_param = *v1_layer_param_;
   layer_param->Clear();
   bool is_fully_compatible = true;
   for (int i = 0; i < v1_layer_param.bottom_size(); ++i) {
@@ -856,9 +858,7 @@ bool UpgradeV1LayerParameter(const V1LayerParameter& v1_layer_param,
   if (v1_layer_param.has_type()) {
     layer_param->set_type(UpgradeV1LayerType(v1_layer_param.type()));
   }
-  for (int i = 0; i < v1_layer_param.blobs_size(); ++i) {
+  layer_param->mutable_blobs()->Swap(v1_layer_param_->mutable_blobs());
-    layer_param->add_blobs()->CopyFrom(v1_layer_param.blobs(i));
-  }
   for (int i = 0; i < v1_layer_param.param_size(); ++i) {
     while (layer_param->param_size() <= i) { layer_param->add_param(); }
     layer_param->mutable_param(i)->set_name(v1_layer_param.param(i));

modules/dnn/src/layers/convolution_layer.cpp

@@ -169,7 +169,8 @@ class ConvolutionLayerImpl CV_FINAL : public BaseConvolutionLayerImpl
 {
 public:
     enum { VEC_ALIGN = 8, DFT_TYPE = CV_32F };
-    Mat weightsMat, weightsMat_doubles;
+    Mat weightsMat;
+    std::vector<double> weightsMultipliers;
     std::vector<float> biasvec;
     std::vector<float> reluslope;
     Ptr<ActivationLayer> activ;
@@ -259,7 +260,7 @@ public:
             wm = wm_aligned;
         }
         weightsMat = wm;
-        weightsMat.convertTo(weightsMat_doubles, CV_64F);
+        weightsMultipliers.assign(outCn, 1.0);
 
         Mat biasMat = hasBias() ? blobs[1].reshape(1, outCn) : Mat();
         biasvec.resize(outCn+2);
@@ -335,13 +336,14 @@ public:
 
         if (!w.empty())
         {
+            Mat originWeights = blobs[0].reshape(1, outCn);
             for (int i = 0; i < outCn; ++i)
             {
                 double wi = w.at<float>(i);
-                cv::multiply(slice(weightsMat_doubles, i), wi, slice(weightsMat_doubles, i));
+                weightsMultipliers[i] *= wi;
+                cv::multiply(originWeights.row(i), weightsMultipliers[i], weightsMat.row(i));
                 biasvec[i] *= wi;
             }
-            weightsMat_doubles.convertTo(weightsMat, weightsMat.type());
         }
 
         if (!b.empty())

modules/dnn/src/tensorflow/tf_graph_simplifier.cpp

@@ -612,7 +612,7 @@ void RemoveIdentityOps(tensorflow::GraphDef& net)
 
 Mat getTensorContent(const tensorflow::TensorProto &tensor)
 {
-    std::string content = tensor.tensor_content();
+    const std::string& content = tensor.tensor_content();
     switch (tensor.dtype())
     {
         case tensorflow::DT_FLOAT:
@@ -681,6 +681,14 @@ Mat getTensorContent(const tensorflow::TensorProto &tensor)
     return Mat();
 }
 
+void releaseTensor(tensorflow::TensorProto* tensor)
+{
+    if (!tensor->mutable_tensor_content()->empty())
+    {
+        delete tensor->release_tensor_content();
+    }
+}
+
 CV__DNN_EXPERIMENTAL_NS_END
 }}  // namespace dnn, namespace cv
 

modules/dnn/src/tensorflow/tf_graph_simplifier.hpp

@@ -23,6 +23,8 @@ void simplifySubgraphs(tensorflow::GraphDef& net);
 
 Mat getTensorContent(const tensorflow::TensorProto &tensor);
 
+void releaseTensor(tensorflow::TensorProto* tensor);
+
 CV__DNN_EXPERIMENTAL_NS_END
 }}  // namespace dnn, namespace cv
 

modules/dnn/src/tensorflow/tf_importer.cpp

@@ -677,7 +677,9 @@ void TFImporter::populateNet(Net dstNet)
                 layers_to_ignore.insert(next_layers[0].first);
             }
 
-            kernelFromTensor(getConstBlob(layer, value_id), layerParams.blobs[0]);
+            const tensorflow::TensorProto& kernelTensor = getConstBlob(layer, value_id);
+            kernelFromTensor(kernelTensor, layerParams.blobs[0]);
+            releaseTensor(const_cast<tensorflow::TensorProto*>(&kernelTensor));
             int* kshape = layerParams.blobs[0].size.p;
             if (type == "DepthwiseConv2dNative")
             {
@@ -788,7 +790,9 @@ void TFImporter::populateNet(Net dstNet)
             }
 
             int kernel_blob_index = -1;
-            blobFromTensor(getConstBlob(layer, value_id, -1, &kernel_blob_index), layerParams.blobs[0]);
+            const tensorflow::TensorProto& kernelTensor = getConstBlob(layer, value_id, -1, &kernel_blob_index);
+            blobFromTensor(kernelTensor, layerParams.blobs[0]);
+            releaseTensor(const_cast<tensorflow::TensorProto*>(&kernelTensor));
 
             if (kernel_blob_index == 1) { // In this case output is computed by x*W formula - W should be transposed
                 Mat data = layerParams.blobs[0].t();