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

Enable some tests for clDNN plugin from Intel's Inference Engine

Browse Source
pull/11351/head
Dmitry Kurtaev 7 years ago
parent 7ea5029ae5
commit bd77d100e1
6 changed files with 81 additions and 69 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. 4
      cmake/OpenCVDetectInferenceEngine.cmake
  2. 6
      modules/dnn/perf/perf_net.cpp
  3. 9
      modules/dnn/src/dnn.cpp
  4. 60
      modules/dnn/src/layers/prior_box_layer.cpp
  5. 12
      modules/dnn/src/op_inf_engine.cpp
  6. 59
      modules/dnn/test/test_backends.cpp

4
cmake/OpenCVDetectInferenceEngine.cmake
Unescape View File

@ -15,6 +15,10 @@ macro(ie_fail)
return()
endmacro()
if(NOT HAVE_CXX11)
ie_fail()
endif()
if(NOT INF_ENGINE_ROOT_DIR OR NOT EXISTS "${INF_ENGINE_ROOT_DIR}/include/inference_engine.hpp")
set(ie_root_paths "${INF_ENGINE_ROOT_DIR}")
if(DEFINED ENV{INTEL_CVSDK_DIR})

6
modules/dnn/perf/perf_net.cpp
Unescape View File

@ -95,24 +95,18 @@ PERF_TEST_P_(DNNTestNetwork, AlexNet)
PERF_TEST_P_(DNNTestNetwork, GoogLeNet)
{
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16)
throw SkipTestException("");
processNet("dnn/bvlc_googlenet.caffemodel", "dnn/bvlc_googlenet.prototxt",
"", Mat(cv::Size(224, 224), CV_32FC3));
}
PERF_TEST_P_(DNNTestNetwork, ResNet_50)
{
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16)
throw SkipTestException("");
processNet("dnn/ResNet-50-model.caffemodel", "dnn/ResNet-50-deploy.prototxt",
"resnet_50.yml", Mat(cv::Size(224, 224), CV_32FC3));
}
PERF_TEST_P_(DNNTestNetwork, SqueezeNet_v1_1)
{
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16)
throw SkipTestException("");
processNet("dnn/squeezenet_v1.1.caffemodel", "dnn/squeezenet_v1.1.prototxt",
"squeezenet_v1_1.yml", Mat(cv::Size(227, 227), CV_32FC3));
}

9
modules/dnn/src/dnn.cpp
Unescape View File

@ -1255,6 +1255,15 @@ struct Net::Impl
if (weightableLayer->_biases)
weightableLayer->_biases = convertFp16(weightableLayer->_biases);
}
else
{
for (const auto& weights : {"weights", "biases"})
{
auto it = ieNode->layer->blobs.find(weights);
if (it != ieNode->layer->blobs.end())
it->second = convertFp16(it->second);
}
}
}
ieNode->connect(ld.inputBlobsWrappers, ld.outputBlobsWrappers);

60
modules/dnn/src/layers/prior_box_layer.cpp
Unescape View File

@ -295,6 +295,19 @@ public:
return false;
}
void finalize(const std::vector<Mat*> &inputs, std::vector<Mat> &outputs) CV_OVERRIDE
{
CV_Assert(inputs.size() > 1, inputs[0]->dims == 4, inputs[1]->dims == 4);
int layerWidth = inputs[0]->size[3];
int layerHeight = inputs[0]->size[2];
int imageWidth = inputs[1]->size[3];
int imageHeight = inputs[1]->size[2];
_stepY = _stepY == 0 ? (static_cast<float>(imageHeight) / layerHeight) : _stepY;
_stepX = _stepX == 0 ? (static_cast<float>(imageWidth) / layerWidth) : _stepX;
}
#ifdef HAVE_OPENCL
bool forward_ocl(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
{
@ -310,16 +323,6 @@ public:
int _imageWidth = inputs[1].size[3];
int _imageHeight = inputs[1].size[2];
float stepX, stepY;
if (_stepX == 0 || _stepY == 0)
{
stepX = static_cast<float>(_imageWidth) / _layerWidth;
stepY = static_cast<float>(_imageHeight) / _layerHeight;
} else {
stepX = _stepX;
stepY = _stepY;
}
if (umat_offsetsX.empty())
{
Mat offsetsX(1, _offsetsX.size(), CV_32FC1, &_offsetsX[0]);
@ -339,8 +342,8 @@ public:
ocl::Kernel kernel("prior_box", ocl::dnn::prior_box_oclsrc);
kernel.set(0, (int)nthreads);
kernel.set(1, (float)stepX);
kernel.set(2, (float)stepY);
kernel.set(1, (float)_stepX);
kernel.set(2, (float)_stepY);
kernel.set(3, ocl::KernelArg::PtrReadOnly(umat_offsetsX));
kernel.set(4, ocl::KernelArg::PtrReadOnly(umat_offsetsY));
kernel.set(5, (int)_offsetsX.size());
@ -410,15 +413,6 @@ public:
int _imageWidth = inputs[1]->size[3];
int _imageHeight = inputs[1]->size[2];
float stepX, stepY;
if (_stepX == 0 || _stepY == 0) {
stepX = static_cast<float>(_imageWidth) / _layerWidth;
stepY = static_cast<float>(_imageHeight) / _layerHeight;
} else {
stepX = _stepX;
stepY = _stepY;
}
float* outputPtr = outputs[0].ptr<float>();
float _boxWidth, _boxHeight;
for (size_t h = 0; h < _layerHeight; ++h)
@ -431,8 +425,8 @@ public:
_boxHeight = _boxHeights[i];
for (int j = 0; j < _offsetsX.size(); ++j)
{
float center_x = (w + _offsetsX[j]) * stepX;
float center_y = (h + _offsetsY[j]) * stepY;
float center_x = (w + _offsetsX[j]) * _stepX;
float center_y = (h + _offsetsY[j]) * _stepY;
outputPtr = addPrior(center_x, center_y, _boxWidth, _boxHeight, _imageWidth,
_imageHeight, _bboxesNormalized, outputPtr);
}
@ -495,7 +489,7 @@ public:
ieLayer->params["aspect_ratio"] += format(",%f", _aspectRatios[i]);
}
ieLayer->params["flip"] = _flip ? "1" : "0";
ieLayer->params["flip"] = "0"; // We already flipped aspect ratios.
ieLayer->params["clip"] = _clip ? "1" : "0";
CV_Assert(!_variance.empty());
@ -503,12 +497,20 @@ public:
for (int i = 1; i < _variance.size(); ++i)
ieLayer->params["variance"] += format(",%f", _variance[i]);
ieLayer->params["step"] = _stepX == _stepY ? format("%f", _stepX) : "0";
ieLayer->params["step_h"] = _stepY;
ieLayer->params["step_w"] = _stepX;
if (_stepX == _stepY)
{
ieLayer->params["step"] = format("%f", _stepX);
ieLayer->params["step_h"] = "0.0";
ieLayer->params["step_w"] = "0.0";
}
else
{
ieLayer->params["step"] = "0.0";
ieLayer->params["step_h"] = format("%f", _stepY);
ieLayer->params["step_w"] = format("%f", _stepX);
}
CV_Assert(_offsetsX.size() == 1, _offsetsY.size() == 1, _offsetsX[0] == _offsetsY[0]);
ieLayer->params["offset"] = format("%f", _offsetsX[0]);;
ieLayer->params["offset"] = format("%f", _offsetsX[0]);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif // HAVE_INF_ENGINE

12
modules/dnn/src/op_inf_engine.cpp
Unescape View File

@ -233,8 +233,16 @@ InferenceEngine::StatusCode
InfEngineBackendNet::getLayerByName(const char *layerName, InferenceEngine::CNNLayerPtr &out,
InferenceEngine::ResponseDesc *resp) noexcept
{
CV_Error(Error::StsNotImplemented, "");
return InferenceEngine::StatusCode::OK;
for (auto& l : layers)
{
if (l->name == layerName)
{
out = l;
return InferenceEngine::StatusCode::OK;
}
}
CV_Error(Error::StsObjectNotFound, cv::format("Cannot find a layer %s", layerName));
return InferenceEngine::StatusCode::NOT_FOUND;
}
void InfEngineBackendNet::setTargetDevice(InferenceEngine::TargetDevice device) noexcept

59
modules/dnn/test/test_backends.cpp
Unescape View File

@ -23,9 +23,9 @@ public:
}
void processNet(const std::string& weights, const std::string& proto,
Size inpSize, const std::string& outputLayer,
Size inpSize, const std::string& outputLayer = "",
const std::string& halideScheduler = "",
double l1 = 1e-5, double lInf = 1e-4)
double l1 = 0.0, double lInf = 0.0)
{
// Create a common input blob.
int blobSize[] = {1, 3, inpSize.height, inpSize.width};
@ -36,9 +36,9 @@ public:
}
void processNet(std::string weights, std::string proto,
Mat inp, const std::string& outputLayer,
Mat inp, const std::string& outputLayer = "",
std::string halideScheduler = "",
double l1 = 1e-5, double lInf = 1e-4)
double l1 = 0.0, double lInf = 0.0)
{
if (backend == DNN_BACKEND_DEFAULT && target == DNN_TARGET_OPENCL)
{
@ -49,6 +49,16 @@ public:
throw SkipTestException("OpenCL is not available/disabled in OpenCV");
}
}
if (target == DNN_TARGET_OPENCL_FP16)
{
l1 = l1 == 0.0 ? 4e-3 : l1;
lInf = lInf == 0.0 ? 2e-2 : lInf;
}
else
{
l1 = l1 == 0.0 ? 1e-5 : l1;
lInf = lInf == 0.0 ? 1e-4 : lInf;
}
weights = findDataFile(weights, false);
if (!proto.empty())
proto = findDataFile(proto, false);
@ -71,31 +81,28 @@ public:
Mat out = net.forward(outputLayer).clone();
if (outputLayer == "detection_out")
normAssertDetections(outDefault, out, "First run", 0.2);
normAssertDetections(outDefault, out, "First run", 0.2, l1, lInf);
else
normAssert(outDefault, out, "First run", l1, lInf);
// Test 2: change input.
inp *= 0.1f;
float* inpData = (float*)inp.data;
for (int i = 0; i < inp.size[0] * inp.size[1]; ++i)
{
Mat slice(inp.size[2], inp.size[3], CV_32F, inpData);
cv::flip(slice, slice, 1);
inpData += slice.total();
}
netDefault.setInput(inp);
net.setInput(inp);
outDefault = netDefault.forward(outputLayer).clone();
out = net.forward(outputLayer).clone();
if (outputLayer == "detection_out")
checkDetections(outDefault, out, "Second run", l1, lInf);
normAssertDetections(outDefault, out, "Second run", 0.2, l1, lInf);
else
normAssert(outDefault, out, "Second run", l1, lInf);
}
void checkDetections(const Mat& out, const Mat& ref, const std::string& msg,
float l1, float lInf, int top = 5)
{
top = std::min(std::min(top, out.size[2]), out.size[3]);
std::vector<cv::Range> range(4, cv::Range::all());
range[2] = cv::Range(0, top);
normAssert(out(range), ref(range));
}
};
TEST_P(DNNTestNetwork, AlexNet)
@ -110,8 +117,6 @@ TEST_P(DNNTestNetwork, AlexNet)
TEST_P(DNNTestNetwork, ResNet_50)
{
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16)
throw SkipTestException("");
processNet("dnn/ResNet-50-model.caffemodel", "dnn/ResNet-50-deploy.prototxt",
Size(224, 224), "prob",
target == DNN_TARGET_OPENCL ? "dnn/halide_scheduler_opencl_resnet_50.yml" :
@ -120,8 +125,6 @@ TEST_P(DNNTestNetwork, ResNet_50)
TEST_P(DNNTestNetwork, SqueezeNet_v1_1)
{
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16)
throw SkipTestException("");
processNet("dnn/squeezenet_v1.1.caffemodel", "dnn/squeezenet_v1.1.prototxt",
Size(227, 227), "prob",
target == DNN_TARGET_OPENCL ? "dnn/halide_scheduler_opencl_squeezenet_v1_1.yml" :
@ -130,8 +133,6 @@ TEST_P(DNNTestNetwork, SqueezeNet_v1_1)
TEST_P(DNNTestNetwork, GoogLeNet)
{
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16)
throw SkipTestException("");
processNet("dnn/bvlc_googlenet.caffemodel", "dnn/bvlc_googlenet.prototxt",
Size(224, 224), "prob");
}
@ -180,7 +181,7 @@ TEST_P(DNNTestNetwork, SSD_VGG16)
{
if (backend == DNN_BACKEND_DEFAULT && target == DNN_TARGET_OPENCL ||
backend == DNN_BACKEND_HALIDE && target == DNN_TARGET_CPU ||
backend == DNN_BACKEND_INFERENCE_ENGINE)
backend == DNN_BACKEND_INFERENCE_ENGINE && target != DNN_TARGET_CPU)
throw SkipTestException("");
processNet("dnn/VGG_ILSVRC2016_SSD_300x300_iter_440000.caffemodel",
"dnn/ssd_vgg16.prototxt", Size(300, 300), "detection_out");
@ -189,30 +190,24 @@ TEST_P(DNNTestNetwork, SSD_VGG16)
TEST_P(DNNTestNetwork, OpenPose_pose_coco)
{
if (backend == DNN_BACKEND_HALIDE) throw SkipTestException("");
double l1 = target == DNN_TARGET_OPENCL_FP16 ? 3e-5 : 1e-5;
double lInf = target == DNN_TARGET_OPENCL_FP16 ? 3e-3 : 1e-4;
processNet("dnn/openpose_pose_coco.caffemodel", "dnn/openpose_pose_coco.prototxt",
Size(368, 368), "", "", l1, lInf);
Size(368, 368));
}
TEST_P(DNNTestNetwork, OpenPose_pose_mpi)
{
if (backend == DNN_BACKEND_HALIDE) throw SkipTestException("");
double l1 = target == DNN_TARGET_OPENCL_FP16 ? 4e-5 : 1e-5;
double lInf = target == DNN_TARGET_OPENCL_FP16 ? 7e-3 : 1e-4;
processNet("dnn/openpose_pose_mpi.caffemodel", "dnn/openpose_pose_mpi.prototxt",
Size(368, 368), "", "", l1, lInf);
Size(368, 368));
}
TEST_P(DNNTestNetwork, OpenPose_pose_mpi_faster_4_stages)
{
if (backend == DNN_BACKEND_HALIDE) throw SkipTestException("");
double l1 = target == DNN_TARGET_OPENCL_FP16 ? 5e-5 : 1e-5;
double lInf = target == DNN_TARGET_OPENCL_FP16 ? 5e-3 : 1e-4;
// The same .caffemodel but modified .prototxt
// See https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/src/openpose/pose/poseParameters.cpp
processNet("dnn/openpose_pose_mpi.caffemodel", "dnn/openpose_pose_mpi_faster_4_stages.prototxt",
Size(368, 368), "", "", l1, lInf);
Size(368, 368));
}
TEST_P(DNNTestNetwork, OpenFace)

Write Preview
Loading…
Cancel
Save