potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 07:43:03 +04:00
Code Packages Releases Wiki Activity

Merge pull request #12082 from dkurt:dnn_ie_faster_rcnn

Browse Source
This commit is contained in:
Vadim Pisarevsky
2018-08-06 14:28:58 +00:00
parent 23022f3ffb 8e034053af
commit e0c93bcf6c
11 changed files with 134 additions and 119 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/perf/perf_net.cpp
+11
View File
@@ -258,6 +258,17 @@ PERF_TEST_P_(DNNTestNetwork, FastNeuralStyle_eccv16)
processNet("dnn/fast_neural_style_eccv16_starry_night.t7", "", "", Mat(cv::Size(320, 240), CV_32FC3));
}
PERF_TEST_P_(DNNTestNetwork, Inception_v2_Faster_RCNN)
{
if (backend == DNN_BACKEND_HALIDE ||
(backend == DNN_BACKEND_INFERENCE_ENGINE && target != DNN_TARGET_CPU) ||
(backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16))
throw SkipTestException("");
processNet("dnn/faster_rcnn_inception_v2_coco_2018_01_28.pb",
"dnn/faster_rcnn_inception_v2_coco_2018_01_28.pbtxt", "",
Mat(cv::Size(800, 600), CV_32FC3));
}
const tuple<DNNBackend, DNNTarget> testCases[] = {
#ifdef HAVE_HALIDE
tuple<DNNBackend, DNNTarget>(DNN_BACKEND_HALIDE, DNN_TARGET_CPU),
modules/dnn/src/dnn.cpp
+20 -19
View File
@@ -1408,7 +1408,7 @@ struct Net::Impl
bool fused = ld.skip;
Ptr<Layer> layer = ld.layerInstance;
if (!layer->supportBackend(preferableBackend))
if (!fused && !layer->supportBackend(preferableBackend))
{
addInfEngineNetOutputs(ld);
net = Ptr<InfEngineBackendNet>();
@@ -2050,10 +2050,10 @@ struct Net::Impl
TickMeter tm;
tm.start();
if (preferableBackend == DNN_BACKEND_OPENCV ||
!layer->supportBackend(preferableBackend))
if( !ld.skip )
{
if( !ld.skip )
std::map<int, Ptr<BackendNode> >::iterator it = ld.backendNodes.find(preferableBackend);
if (preferableBackend == DNN_BACKEND_OPENCV || it == ld.backendNodes.end() || it->second.empty())
{
if (preferableBackend == DNN_BACKEND_OPENCV && IS_DNN_OPENCL_TARGET(preferableTarget))
{
@@ -2196,24 +2196,25 @@ struct Net::Impl
}
}
else
tm.reset();
}
else if (!ld.skip)
{
Ptr<BackendNode> node = ld.backendNodes[preferableBackend];
if (preferableBackend == DNN_BACKEND_HALIDE)
{
forwardHalide(ld.outputBlobsWrappers, node);
}
else if (preferableBackend == DNN_BACKEND_INFERENCE_ENGINE)
{
forwardInfEngine(node);
}
else
{
CV_Error(Error::StsNotImplemented, "Unknown backend identifier");
Ptr<BackendNode> node = it->second;
CV_Assert(!node.empty());
if (preferableBackend == DNN_BACKEND_HALIDE)
{
forwardHalide(ld.outputBlobsWrappers, node);
}
else if (preferableBackend == DNN_BACKEND_INFERENCE_ENGINE)
{
forwardInfEngine(node);
}
else
{
CV_Error(Error::StsNotImplemented, "Unknown backend identifier");
}
}
}
else
tm.reset();
tm.stop();
layersTimings[ld.id] = tm.getTimeTicks();
modules/dnn/src/layers/detection_output_layer.cpp
+1 -1
View File
@@ -196,7 +196,7 @@ public:
virtual bool supportBackend(int backendId) CV_OVERRIDE
{
return backendId == DNN_BACKEND_OPENCV ||
backendId == DNN_BACKEND_INFERENCE_ENGINE && !_locPredTransposed && _bboxesNormalized;
backendId == DNN_BACKEND_INFERENCE_ENGINE && !_locPredTransposed && _bboxesNormalized && !_clip;
}
bool getMemoryShapes(const std::vector<MatShape> &inputs,
modules/dnn/src/layers/scale_layer.cpp
+2 -3
View File
@@ -48,9 +48,8 @@ public:
virtual bool supportBackend(int backendId) CV_OVERRIDE
{
return backendId == DNN_BACKEND_OPENCV ||
backendId == DNN_BACKEND_HALIDE && haveHalide() ||
backendId == DNN_BACKEND_INFERENCE_ENGINE && haveInfEngine();
return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE ||
backendId == DNN_BACKEND_INFERENCE_ENGINE && axis == 1;
}
void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE
modules/dnn/src/layers/slice_layer.cpp
+1 -1
View File
@@ -111,7 +111,7 @@ public:
virtual bool supportBackend(int backendId) CV_OVERRIDE
{
return backendId == DNN_BACKEND_OPENCV ||
backendId == DNN_BACKEND_INFERENCE_ENGINE && sliceRanges.size() == 1;
backendId == DNN_BACKEND_INFERENCE_ENGINE && sliceRanges.size() == 1 && sliceRanges[0].size() == 4;
}
bool getMemoryShapes(const std::vector<MatShape> &inputs,
modules/dnn/src/layers/softmax_layer.cpp
+4 -2
View File
@@ -307,15 +307,17 @@ public:
return Ptr<BackendNode>();
}
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >& inputs) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
InferenceEngine::DataPtr input = infEngineDataNode(inputs[0]);
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "SoftMax";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::SoftMaxLayer> ieLayer(new InferenceEngine::SoftMaxLayer(lp));
ieLayer->axis = axisRaw;
ieLayer->axis = clamp(axisRaw, input->dims.size());
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
modules/dnn/src/tensorflow/tf_importer.cpp
+30 -36
View File
@@ -954,6 +954,13 @@ void TFImporter::populateNet(Net dstNet)
{
CV_Assert(layer.input_size() == 2);
// For the object detection networks, TensorFlow Object Detection API
// predicts deltas for bounding boxes in yxYX (ymin, xmin, ymax, xmax)
// order. We can manage it at DetectionOutput layer parsing predictions
// or shuffle last Faster-RCNN's matmul weights.
bool locPredTransposed = hasLayerAttr(layer, "loc_pred_transposed") &&
getLayerAttr(layer, "loc_pred_transposed").b();
layerParams.set("bias_term", false);
layerParams.blobs.resize(1);
@@ -970,6 +977,17 @@ void TFImporter::populateNet(Net dstNet)
blobFromTensor(getConstBlob(net.node(weights_layer_index), value_id), layerParams.blobs[1]);
ExcludeLayer(net, weights_layer_index, 0, false);
layers_to_ignore.insert(next_layers[0].first);
if (locPredTransposed)
{
const int numWeights = layerParams.blobs[1].total();
float* biasData = reinterpret_cast<float*>(layerParams.blobs[1].data);
CV_Assert(numWeights % 4 == 0);
for (int i = 0; i < numWeights; i += 2)
{
std::swap(biasData[i], biasData[i + 1]);
}
}
}
int kernel_blob_index = -1;
@@ -983,6 +1001,16 @@ void TFImporter::populateNet(Net dstNet)
}
layerParams.set("num_output", layerParams.blobs[0].size[0]);
if (locPredTransposed)
{
CV_Assert(layerParams.blobs[0].dims == 2);
for (int i = 0; i < layerParams.blobs[0].size[0]; i += 2)
{
cv::Mat src = layerParams.blobs[0].row(i);
cv::Mat dst = layerParams.blobs[0].row(i + 1);
std::swap_ranges(src.begin<float>(), src.end<float>(), dst.begin<float>());
}
}
int id = dstNet.addLayer(name, "InnerProduct", layerParams);
layer_id[name] = id;
@@ -1010,6 +1038,7 @@ void TFImporter::populateNet(Net dstNet)
layer_id[permName] = permId;
connect(layer_id, dstNet, inpId, permId, 0);
inpId = Pin(permName);
inpLayout = DATA_LAYOUT_NCHW;
}
else if (newShape.total() == 4 && inpLayout == DATA_LAYOUT_NHWC)
{
@@ -1024,7 +1053,7 @@ void TFImporter::populateNet(Net dstNet)
// one input only
connect(layer_id, dstNet, inpId, id, 0);
data_layouts[name] = newShape.total() == 2 ? DATA_LAYOUT_PLANAR : DATA_LAYOUT_UNKNOWN;
data_layouts[name] = newShape.total() == 2 ? DATA_LAYOUT_PLANAR : inpLayout;
}
else if (type == "Flatten" || type == "Squeeze")
{
@@ -1696,41 +1725,6 @@ void TFImporter::populateNet(Net dstNet)
connect(layer_id, dstNet, parsePin(layer.input(1)), id, 1);
data_layouts[name] = DATA_LAYOUT_UNKNOWN;
}
else if (type == "DetectionOutput")
{
// op: "DetectionOutput"
// input_0: "locations"
// input_1: "classifications"
// input_2: "prior_boxes"
if (hasLayerAttr(layer, "num_classes"))
layerParams.set("num_classes", getLayerAttr(layer, "num_classes").i());
if (hasLayerAttr(layer, "share_location"))
layerParams.set("share_location", getLayerAttr(layer, "share_location").b());
if (hasLayerAttr(layer, "background_label_id"))
layerParams.set("background_label_id", getLayerAttr(layer, "background_label_id").i());
if (hasLayerAttr(layer, "nms_threshold"))
layerParams.set("nms_threshold", getLayerAttr(layer, "nms_threshold").f());
if (hasLayerAttr(layer, "top_k"))
layerParams.set("top_k", getLayerAttr(layer, "top_k").i());
if (hasLayerAttr(layer, "code_type"))
layerParams.set("code_type", getLayerAttr(layer, "code_type").s());
if (hasLayerAttr(layer, "keep_top_k"))
layerParams.set("keep_top_k", getLayerAttr(layer, "keep_top_k").i());
if (hasLayerAttr(layer, "confidence_threshold"))
layerParams.set("confidence_threshold", getLayerAttr(layer, "confidence_threshold").f());
if (hasLayerAttr(layer, "loc_pred_transposed"))
layerParams.set("loc_pred_transposed", getLayerAttr(layer, "loc_pred_transposed").b());
if (hasLayerAttr(layer, "clip"))
layerParams.set("clip", getLayerAttr(layer, "clip").b());
if (hasLayerAttr(layer, "variance_encoded_in_target"))
layerParams.set("variance_encoded_in_target", getLayerAttr(layer, "variance_encoded_in_target").b());
int id = dstNet.addLayer(name, "DetectionOutput", layerParams);
layer_id[name] = id;
for (int i = 0; i < 3; ++i)
connect(layer_id, dstNet, parsePin(layer.input(i)), id, i);
data_layouts[name] = DATA_LAYOUT_UNKNOWN;
}
else if (type == "Softmax")
{
if (hasLayerAttr(layer, "axis"))
modules/dnn/test/test_tf_importer.cpp
+1 -1
View File
@@ -323,7 +323,7 @@ TEST_P(Test_TensorFlow_nets, Inception_v2_SSD)
TEST_P(Test_TensorFlow_nets, Inception_v2_Faster_RCNN)
{
checkBackend();
if (backend == DNN_BACKEND_INFERENCE_ENGINE ||
if ((backend == DNN_BACKEND_INFERENCE_ENGINE && target != DNN_TARGET_CPU) ||
(backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16))
throw SkipTestException("");