Chiebot-Mirror
/
opencv_contrib
8
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Merge pull request #1248 from arrybn:enable_googlenet_test

Browse Source
pull/1251/head
Vadim Pisarevsky 8 years ago
parent e551d15c2b b0d008ce03
commit 62b781bc06
2 changed files with 82 additions and 73 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 142
      modules/dnn/src/dnn.cpp
  2. 3
      modules/dnn/test/test_googlenet.cpp

142
modules/dnn/src/dnn.cpp
Unescape View File

@ -324,7 +324,6 @@ struct LayerData
//add logging info
params.name = name;
params.type = type;
skip = false;
}
int id;
@ -347,7 +346,6 @@ struct LayerData
std::map<int, bool> skipFlags;
int flag;
bool skip;
Ptr<Layer> getLayerInstance()
{
@ -666,9 +664,7 @@ struct Net::Impl
}
}
void setUpNet(const std::vector<LayerPin>& blobsToKeep_ = std::vector<LayerPin>())
{
if (!netWasAllocated || this->blobsToKeep != blobsToKeep_)
void clear()
{
MapIdToLayerData::iterator it;
for (it = layers.begin(); it != layers.end(); it++)
@ -677,7 +673,30 @@ struct Net::Impl
it->second.outputBlobs.clear();
it->second.internals.clear();
}
it->second.skipFlags.clear();
it->second.consumers.clear();
Ptr<ConvolutionLayer> convLayer = it->second.layerInstance.dynamicCast<ConvolutionLayer>();
if( !convLayer.empty() )
{
convLayer->setActivation(Ptr<ActivationLayer>());
convLayer->setBatchNorm(Ptr<BatchNormLayer>());
}
Ptr<PoolingLayer> poolingLayer = it->second.layerInstance.dynamicCast<PoolingLayer>();
if( !poolingLayer.empty() )
{
poolingLayer->computeMaxIdx = true;
}
}
}
void setUpNet(const std::vector<LayerPin>& blobsToKeep_ = std::vector<LayerPin>())
{
if (!netWasAllocated || this->blobsToKeep != blobsToKeep_)
{
clear();
allocateLayers(blobsToKeep_);
computeNetOutputLayers();
@ -1005,69 +1024,41 @@ struct Net::Impl
ld.flag = 1;
}
void allocateLayers(const std::vector<LayerPin>& blobsToKeep_)
{
MapIdToLayerData::iterator it;
for (it = layers.begin(); it != layers.end(); it++)
it->second.flag = 0;
CV_Assert(!layers[0].outputBlobs.empty());
ShapesVec inputShapes;
for(int i = 0; i < layers[0].outputBlobs.size(); i++)
{
CV_Assert(layers[0].outputBlobs[i].total());
inputShapes.push_back(shape(layers[0].outputBlobs[i]));
}
LayersShapesMap layersShapes;
getLayersShapes(inputShapes, layersShapes);
blobManager.reset();
for (it = layers.begin(); it != layers.end(); ++it)
{
const LayerData& ld = it->second;
blobManager.addReferences(ld.inputBlobsId);
}
for (int i = 0; i < blobsToKeep_.size(); i++)
{
blobManager.addReference(blobsToKeep_[i]);
}
for (it = layers.begin(); it != layers.end(); it++)
void fuseLayers(const std::vector<LayerPin>& blobsToKeep_)
{
int lid = it->first;
allocateLayer(lid, layersShapes);
}
// scan through all the layers. If there is convolution layer followed by the activation layer,
// we try to embed this activation into the convolution and disable separate execution of the activation
std::vector<String> outnames;
std::set<LayerPin> pinsToKeep(blobsToKeep_.begin(),
blobsToKeep_.end());
MapIdToLayerData::iterator it;
for (it = layers.begin(); it != layers.end(); it++)
{
int lid = it->first;
LayerData& ld = layers[lid];
if( ld.skip )
if( ld.skipFlags[DNN_BACKEND_DEFAULT] )
{
//printf("skipping %s\n", ld.layerInstance->name.c_str());
continue;
}
//printf("analyzing %s\n", ld.layerInstance->name.c_str());
if( ld.consumers.size() == 0 )
outnames.push_back(ld.layerInstance->name);
Ptr<ConvolutionLayer> convLayer = ld.layerInstance.dynamicCast<ConvolutionLayer>();
if( !convLayer.empty() && ld.consumers.size() == 1 )
LayerPin lp(lid, 0);
if( !convLayer.empty() && ld.consumers.size() == 1 &&
pinsToKeep.count(lp) == 0 )
{
LayerData* nextData = &layers[ld.consumers[0].lid];
Ptr<BatchNormLayer> nextBNormLayer =
nextData->layerInstance.dynamicCast<BatchNormLayer>();
if( !nextBNormLayer.empty() )
LayerPin lpNext(ld.consumers[0].lid, 0);
if( !nextBNormLayer.empty() && pinsToKeep.count(lpNext) == 0 )
{
LayerData* bnormData = nextData;
nextData = 0;
if( convLayer->setBatchNorm(nextBNormLayer) )
{
//printf("fused convolution (%s) and batch norm (%s)\n", convLayer->name.c_str(), nextBNormLayer->name.c_str());
bnormData->skip = true;
bnormData->skipFlags[DNN_BACKEND_DEFAULT] = true;
ld.outputBlobs = layers[lpNext.lid].outputBlobs;
if( bnormData->consumers.size() == 1 )
nextData = &layers[bnormData->consumers[0].lid];
}
@ -1079,8 +1070,8 @@ struct Net::Impl
if( !nextActivLayer.empty() && convLayer->setActivation(nextActivLayer) )
{
//printf("fused convolution (%s) and activation (%s)\n", convLayer->name.c_str(), nextActivLayer->name.c_str());
nextData->skip = true;
nextData->skipFlags[DNN_BACKEND_DEFAULT] = true;
ld.outputBlobs = layers[lpNext.lid].outputBlobs;
}
}
Ptr<PoolingLayer> poolingLayer = ld.layerInstance.dynamicCast<PoolingLayer>();
@ -1096,10 +1087,43 @@ struct Net::Impl
poolingLayer->computeMaxIdx = false;
}
}
/*printf("outputs: ");
for( size_t j = 0; j < outnames.size(); j++ )
printf("%s ", outnames[j].c_str());
printf("\n");*/
}
void allocateLayers(const std::vector<LayerPin>& blobsToKeep_)
{
MapIdToLayerData::iterator it;
for (it = layers.begin(); it != layers.end(); it++)
it->second.flag = 0;
CV_Assert(!layers[0].outputBlobs.empty());
ShapesVec inputShapes;
for(int i = 0; i < layers[0].outputBlobs.size(); i++)
{
CV_Assert(layers[0].outputBlobs[i].total());
inputShapes.push_back(shape(layers[0].outputBlobs[i]));
}
LayersShapesMap layersShapes;
getLayersShapes(inputShapes, layersShapes);
blobManager.reset();
for (it = layers.begin(); it != layers.end(); ++it)
{
const LayerData& ld = it->second;
blobManager.addReferences(ld.inputBlobsId);
}
for (int i = 0; i < blobsToKeep_.size(); i++)
{
blobManager.addReference(blobsToKeep_[i]);
}
for (it = layers.begin(); it != layers.end(); it++)
{
int lid = it->first;
allocateLayer(lid, layersShapes);
}
fuseLayers(blobsToKeep_);
}
void forwardLayer(LayerData &ld)
@ -1109,7 +1133,7 @@ struct Net::Impl
if (preferableBackend == DNN_BACKEND_DEFAULT ||
!layer->supportBackend(preferableBackend))
{
if( !ld.skip )
if( !ld.skipFlags[DNN_BACKEND_DEFAULT] )
layer->forward(ld.inputBlobs, ld.outputBlobs, ld.internals);
}
else if (!ld.skipFlags[preferableBackend])
@ -1300,20 +1324,6 @@ void Net::connect(String _outPin, String _inPin)
impl->connect(outPin.lid, outPin.oid, inpPin.lid, inpPin.oid);
}
//void Net::forward(LayerId toLayer)
//{
// if (!impl->netWasAllocated)
// {
// impl->setUpNet();
// }
// if (toLayer.isString() && toLayer.get<String>().empty())
// impl->forwardAll();
// else
// impl->forwardLayer(impl->getLayerData(toLayer));
//}
Mat Net::forward(const String& outputName)
{
String layerName = outputName;

3
modules/dnn/test/test_googlenet.cpp
Unescape View File

@ -95,8 +95,7 @@ static void launchGoogleNetTest()
std::replace( filename.begin(), filename.end(), '/', '#');
Mat ref = blobFromNPY(_tf("googlenet_" + filename + ".npy"));
// TODO: disabled the check for now, because it conflicts with the layer fusion
// normAssert(outs[i], ref, "", 1E-4, 1E-2);
normAssert(outs[i], ref, "", 1E-4, 1E-2);
}
}

Write Preview
Loading…
Cancel
Save