Chiebot-Mirror
/
opencv
mirror of https://github.com/opencv/opencv.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity

Merge pull request #24196 from dkurt:ov_backend_cleanups

Browse Source 
Use ngraph::Output in OpenVINO backend wrapper #24196

### Pull Request Readiness Checklist

resolves https://github.com/opencv/opencv/issues/24102

* Use `ngraph::Output<ngraph::Node>>` insead of `std::shared_ptr<ngraph::Node>` as a backend wrapper. It lets access to multi-output nodes: 588ddf1b18/modules/dnn/src/net_openvino.cpp (L501-L504)
* All layers can be customizable with OpenVINO >= 2022.1. nGraph reference code used for default layer implementation does not required CPU plugin also (might be tested by commenting CPU plugin at `/opt/intel/openvino/runtime/lib/intel64/plugins.xml`).
* Correct inference if only intermediate blobs requested.


See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
      Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake
pull/24228/head
Dmitry Kurtaev 1 year ago committed by GitHub
parent 2c53e3f53d
commit 178fdbbda8
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
26 changed files with 136 additions and 95 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 21
      modules/dnn/src/ie_ngraph.cpp
  2. 6
      modules/dnn/src/ie_ngraph.hpp
  3. 2
      modules/dnn/src/layers/batch_norm_layer.cpp
  4. 2
      modules/dnn/src/layers/blank_layer.cpp
  5. 4
      modules/dnn/src/layers/concat_layer.cpp
  6. 8
      modules/dnn/src/layers/convolution_layer.cpp
  7. 2
      modules/dnn/src/layers/crop_and_resize_layer.cpp
  8. 28
      modules/dnn/src/layers/elementwise_layers.cpp
  9. 15
      modules/dnn/src/layers/eltwise_layer.cpp
  10. 2
      modules/dnn/src/layers/flatten_layer.cpp
  11. 2
      modules/dnn/src/layers/fully_connected_layer.cpp
  12. 2
      modules/dnn/src/layers/lrn_layer.cpp
  13. 4
      modules/dnn/src/layers/max_unpooling_layer.cpp
  14. 2
      modules/dnn/src/layers/mvn_layer.cpp
  15. 6
      modules/dnn/src/layers/nary_eltwise_layers.cpp
  16. 8
      modules/dnn/src/layers/normalize_bbox_layer.cpp
  17. 2
      modules/dnn/src/layers/pooling_layer.cpp
  18. 4
      modules/dnn/src/layers/proposal_layer.cpp
  19. 4
      modules/dnn/src/layers/region_layer.cpp
  20. 2
      modules/dnn/src/layers/resize_layer.cpp
  21. 18
      modules/dnn/src/layers/scale_layer.cpp
  22. 2
      modules/dnn/src/layers/slice_layer.cpp
  23. 2
      modules/dnn/src/layers/softmax_layer.cpp
  24. 63
      modules/dnn/src/net_openvino.cpp
  25. 11
      modules/dnn/test/test_halide_layers.cpp
  26. 9
      modules/dnn/test/test_tflite_importer.cpp

21
modules/dnn/src/ie_ngraph.cpp
Unescape View File

@ -383,11 +383,17 @@ public:
#endif // OpenVINO >= 2022.1 #endif // OpenVINO >= 2022.1
InfEngineNgraphNode::InfEngineNgraphNode(std::shared_ptr<ngraph::Node>&& _node) InfEngineNgraphNode::InfEngineNgraphNode(ngraph::Output<ngraph::Node>&& _node)
: BackendNode(DNN_BACKEND_INFERENCE_ENGINE_NGRAPH), node(std::move(_node)) {} : BackendNode(DNN_BACKEND_INFERENCE_ENGINE_NGRAPH), node(std::move(_node)) {
CV_Assert(node.get_node());
CV_Assert(node.get_node_shared_ptr());
}
InfEngineNgraphNode::InfEngineNgraphNode(const std::shared_ptr<ngraph::Node>& _node) InfEngineNgraphNode::InfEngineNgraphNode(const ngraph::Output<ngraph::Node>& _node)
: BackendNode(DNN_BACKEND_INFERENCE_ENGINE_NGRAPH), node(_node) {} : BackendNode(DNN_BACKEND_INFERENCE_ENGINE_NGRAPH), node(_node) {
CV_Assert(node.get_node());
CV_Assert(node.get_node_shared_ptr());
}
InfEngineNgraphNode::InfEngineNgraphNode(const std::vector<Ptr<BackendNode> >& nodes, InfEngineNgraphNode::InfEngineNgraphNode(const std::vector<Ptr<BackendNode> >& nodes,
Ptr<Layer>& cvLayer_, std::vector<Mat*>& inputs, Ptr<Layer>& cvLayer_, std::vector<Mat*>& inputs,
@ -420,7 +426,7 @@ InfEngineNgraphNode::InfEngineNgraphNode(const std::vector<Ptr<BackendNode> >& n
} }
void InfEngineNgraphNode::setName(const std::string& name) { void InfEngineNgraphNode::setName(const std::string& name) {
node->set_friendly_name(name); node.get_node()->set_friendly_name(name);
} }
InfEngineNgraphNet::InfEngineNgraphNet(detail::NetImplBase& netImpl) InfEngineNgraphNet::InfEngineNgraphNet(detail::NetImplBase& netImpl)
@ -441,8 +447,7 @@ InfEngineNgraphNet::InfEngineNgraphNet(detail::NetImplBase& netImpl, InferenceEn
void InfEngineNgraphNet::addOutput(const Ptr<InfEngineNgraphNode>& node) void InfEngineNgraphNet::addOutput(const Ptr<InfEngineNgraphNode>& node)
{ {
CV_Assert(node); CV_Assert(node);
CV_Assert(node->node); const std::string& name = node->node.get_node()->get_friendly_name();
const std::string& name = node->node->get_friendly_name();
requestedOutputs.insert({name, node.get()}); requestedOutputs.insert({name, node.get()});
} }
@ -458,7 +463,7 @@ void InfEngineNgraphNet::createNet(Target targetId) {
CV_Assert(output_node_it->second); CV_Assert(output_node_it->second);
auto out = std::make_shared<ngraph::op::Result>(output_node_it->second->node); auto out = std::make_shared<ngraph::op::Result>(output_node_it->second->node);
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2022_1) #if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2022_1)
out->set_friendly_name(output_node_it->first + (output_node_it->second->node->get_output_size() == 1 ? "" : ".0")); out->set_friendly_name(output_node_it->first + (output_node_it->second->node.get_node()->get_output_size() == 1 ? "" : ".0"));
#endif #endif
outs.push_back(out); outs.push_back(out);
} }

6
modules/dnn/src/ie_ngraph.hpp
Unescape View File

@ -93,13 +93,13 @@ public:
std::vector<Mat*>& inputs, std::vector<Mat>& outputs, std::vector<Mat*>& inputs, std::vector<Mat>& outputs,
std::vector<Mat>& internals); std::vector<Mat>& internals);
InfEngineNgraphNode(std::shared_ptr<ngraph::Node>&& _node); InfEngineNgraphNode(ngraph::Output<ngraph::Node>&& _node);
InfEngineNgraphNode(const std::shared_ptr<ngraph::Node>& _node); InfEngineNgraphNode(const ngraph::Output<ngraph::Node>& _node);
void setName(const std::string& name); void setName(const std::string& name);
// Inference Engine network object that allows to obtain the outputs of this layer. // Inference Engine network object that allows to obtain the outputs of this layer.
std::shared_ptr<ngraph::Node> node; ngraph::Output<ngraph::Node> node;
Ptr<InfEngineNgraphNet> net; Ptr<InfEngineNgraphNet> net;
Ptr<dnn::Layer> cvLayer; Ptr<dnn::Layer> cvLayer;
}; };

2
modules/dnn/src/layers/batch_norm_layer.cpp
Unescape View File

@ -457,7 +457,7 @@ public:
virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >& inputs, const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >& inputs, const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE
{ {
auto ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node; auto ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
std::vector<size_t> shape(ieInpNode->get_shape().size(), 1); std::vector<size_t> shape(ieInpNode.get_shape().size(), 1);
shape[1] = weights_.total(); shape[1] = weights_.total();
auto weight = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape(shape), weights_.data); auto weight = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape(shape), weights_.data);
auto bias = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape(shape), bias_.data); auto bias = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape(shape), bias_.data);

2
modules/dnn/src/layers/blank_layer.cpp
Unescape View File

@ -148,7 +148,7 @@ public:
virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >& inputs, virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >& inputs,
const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE
{ {
auto& ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node; auto ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
ngraph::OutputVector inp{ieInpNode}; ngraph::OutputVector inp{ieInpNode};
auto blank = std::make_shared<ngraph::op::Concat>(inp, 0); auto blank = std::make_shared<ngraph::op::Concat>(inp, 0);
return Ptr<BackendNode>(new InfEngineNgraphNode(blank)); return Ptr<BackendNode>(new InfEngineNgraphNode(blank));

4
modules/dnn/src/layers/concat_layer.cpp
Unescape View File

@ -392,7 +392,7 @@ public:
virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >& inputs, virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >& inputs,
const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE
{ {
const int numDims = nodes[0].dynamicCast<InfEngineNgraphNode>()->node->get_shape().size(); const int numDims = nodes[0].dynamicCast<InfEngineNgraphNode>()->node.get_shape().size();
const int cAxis = normalize_axis(axis, numDims); const int cAxis = normalize_axis(axis, numDims);
std::vector<size_t> maxDims(numDims, 0); std::vector<size_t> maxDims(numDims, 0);
@ -403,7 +403,7 @@ public:
auto inp = nodes[i].dynamicCast<InfEngineNgraphNode>()->node; auto inp = nodes[i].dynamicCast<InfEngineNgraphNode>()->node;
inp_nodes.push_back(inp); inp_nodes.push_back(inp);
std::vector<size_t> inpShape = inp->get_shape(); std::vector<size_t> inpShape = inp.get_shape();
for (int i = 0; i < numDims; ++i) for (int i = 0; i < numDims; ++i)
maxDims[i] = std::max(maxDims[i], inpShape[i]); maxDims[i] = std::max(maxDims[i], inpShape[i]);
} }

8
modules/dnn/src/layers/convolution_layer.cpp
Unescape View File

@ -822,13 +822,13 @@ public:
CV_Assert(!blobs.empty()); CV_Assert(!blobs.empty());
CV_Assert_N(inputs.size() >= 1, nodes.size() >= 1); CV_Assert_N(inputs.size() >= 1, nodes.size() >= 1);
auto& ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node; auto& ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
std::vector<size_t> dims = ieInpNode->get_shape(); std::vector<size_t> dims = ieInpNode.get_shape();
CV_Check(dims.size(), dims.size() >= 3 && dims.size() <= 5, ""); CV_Check(dims.size(), dims.size() >= 3 && dims.size() <= 5, "");
std::shared_ptr<ngraph::Node> ieWeights = nodes.size() > 1 ? nodes[1].dynamicCast<InfEngineNgraphNode>()->node : nullptr; ngraph::Output<ngraph::Node> ieWeights;
if (nodes.size() > 1) if (nodes.size() > 1)
CV_Assert(ieWeights); // dynamic_cast should not fail ieWeights = nodes[1].dynamicCast<InfEngineNgraphNode>()->node;
const int inpCn = dims[1]; const int inpCn = dims[1];
const int inpGroupCn = nodes.size() > 1 ? ieWeights->get_shape()[1] : blobs[0].size[1]; const int inpGroupCn = nodes.size() > 1 ? ieWeights.get_shape()[1] : blobs[0].size[1];
const int group = inpCn / inpGroupCn; const int group = inpCn / inpGroupCn;
std::vector<size_t> kernel_shape; std::vector<size_t> kernel_shape;

2
modules/dnn/src/layers/crop_and_resize_layer.cpp
Unescape View File

@ -133,7 +133,7 @@ public:
auto input = nodes[0].dynamicCast<InfEngineNgraphNode>()->node; auto input = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
auto rois = nodes[1].dynamicCast<InfEngineNgraphNode>()->node; auto rois = nodes[1].dynamicCast<InfEngineNgraphNode>()->node;
auto rois_shape = rois->get_shape(); auto rois_shape = rois.get_shape();
std::vector<int64_t> dims(rois_shape.begin(), rois_shape.end()), offsets(4, 0); std::vector<int64_t> dims(rois_shape.begin(), rois_shape.end()), offsets(4, 0);
offsets[3] = 2; offsets[3] = 2;
dims[3] = 7; dims[3] = 7;

28
modules/dnn/src/layers/elementwise_layers.cpp
Unescape View File

@ -490,7 +490,7 @@ struct ReLUFunctor : public BaseFunctor
#endif #endif
#ifdef HAVE_DNN_NGRAPH #ifdef HAVE_DNN_NGRAPH
std::shared_ptr<ngraph::Node> initNgraphAPI(const std::shared_ptr<ngraph::Node>& node) std::shared_ptr<ngraph::Node> initNgraphAPI(const ngraph::Output<ngraph::Node>& node)
{ {
if (slope) { if (slope) {
auto param = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape{1}, &slope); auto param = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape{1}, &slope);
@ -674,7 +674,7 @@ struct ReLU6Functor : public BaseFunctor
#ifdef HAVE_DNN_NGRAPH #ifdef HAVE_DNN_NGRAPH
std::shared_ptr<ngraph::Node> initNgraphAPI(const std::shared_ptr<ngraph::Node>& node) std::shared_ptr<ngraph::Node> initNgraphAPI(const ngraph::Output<ngraph::Node>& node)
{ {
return std::make_shared<ngraph::op::Clamp>(node, minValue, maxValue); return std::make_shared<ngraph::op::Clamp>(node, minValue, maxValue);
} }
@ -796,7 +796,7 @@ struct BaseDefaultFunctor : public BaseFunctor
#endif // HAVE_CANN #endif // HAVE_CANN
#ifdef HAVE_DNN_NGRAPH #ifdef HAVE_DNN_NGRAPH
std::shared_ptr<ngraph::Node> initNgraphAPI(const std::shared_ptr<ngraph::Node>& node) std::shared_ptr<ngraph::Node> initNgraphAPI(const ngraph::Output<ngraph::Node>& node)
{ {
CV_Error(Error::StsNotImplemented, ""); CV_Error(Error::StsNotImplemented, "");
} }
@ -929,7 +929,7 @@ struct TanHFunctor : public BaseDefaultFunctor<TanHFunctor>
#endif // HAVE_CANN #endif // HAVE_CANN
#ifdef HAVE_DNN_NGRAPH #ifdef HAVE_DNN_NGRAPH
std::shared_ptr<ngraph::Node> initNgraphAPI(const std::shared_ptr<ngraph::Node>& node) std::shared_ptr<ngraph::Node> initNgraphAPI(const ngraph::Output<ngraph::Node>& node)
{ {
return std::make_shared<ngraph::op::Tanh>(node); return std::make_shared<ngraph::op::Tanh>(node);
} }
@ -998,7 +998,7 @@ struct SwishFunctor : public BaseDefaultFunctor<SwishFunctor>
#endif // HAVE_CANN #endif // HAVE_CANN
#ifdef HAVE_DNN_NGRAPH #ifdef HAVE_DNN_NGRAPH
std::shared_ptr<ngraph::Node> initNgraphAPI(const std::shared_ptr<ngraph::Node>& node) std::shared_ptr<ngraph::Node> initNgraphAPI(const ngraph::Output<ngraph::Node>& node)
{ {
auto sigmoid = std::make_shared<ngraph::op::Sigmoid>(node); auto sigmoid = std::make_shared<ngraph::op::Sigmoid>(node);
return std::make_shared<ngraph::op::v1::Multiply>(node, sigmoid); return std::make_shared<ngraph::op::v1::Multiply>(node, sigmoid);
@ -1074,7 +1074,7 @@ struct MishFunctor : public BaseDefaultFunctor<MishFunctor>
#endif // HAVE_CANN #endif // HAVE_CANN
#ifdef HAVE_DNN_NGRAPH #ifdef HAVE_DNN_NGRAPH
std::shared_ptr<ngraph::Node> initNgraphAPI(const std::shared_ptr<ngraph::Node>& node) std::shared_ptr<ngraph::Node> initNgraphAPI(const ngraph::Output<ngraph::Node>& node)
{ {
float one = 1.0f; float one = 1.0f;
auto constant = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape{1}, &one); auto constant = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape{1}, &one);
@ -1157,7 +1157,7 @@ struct SigmoidFunctor : public BaseDefaultFunctor<SigmoidFunctor>
#endif // HAVE_CANN #endif // HAVE_CANN
#ifdef HAVE_DNN_NGRAPH #ifdef HAVE_DNN_NGRAPH
std::shared_ptr<ngraph::Node> initNgraphAPI(const std::shared_ptr<ngraph::Node>& node) std::shared_ptr<ngraph::Node> initNgraphAPI(const ngraph::Output<ngraph::Node>& node)
{ {
return std::make_shared<ngraph::op::Sigmoid>(node); return std::make_shared<ngraph::op::Sigmoid>(node);
} }
@ -1237,7 +1237,7 @@ struct ELUFunctor : public BaseDefaultFunctor<ELUFunctor>
#endif // HAVE_CANN #endif // HAVE_CANN
#ifdef HAVE_DNN_NGRAPH #ifdef HAVE_DNN_NGRAPH
std::shared_ptr<ngraph::Node> initNgraphAPI(const std::shared_ptr<ngraph::Node>& node) std::shared_ptr<ngraph::Node> initNgraphAPI(const ngraph::Output<ngraph::Node>& node)
{ {
return std::make_shared<ngraph::op::Elu>(node, alpha); return std::make_shared<ngraph::op::Elu>(node, alpha);
} }
@ -1307,7 +1307,7 @@ struct AbsValFunctor : public BaseDefaultFunctor<AbsValFunctor>
#endif // HAVE_CANN #endif // HAVE_CANN
#ifdef HAVE_DNN_NGRAPH #ifdef HAVE_DNN_NGRAPH
std::shared_ptr<ngraph::Node> initNgraphAPI(const std::shared_ptr<ngraph::Node>& node) std::shared_ptr<ngraph::Node> initNgraphAPI(const ngraph::Output<ngraph::Node>& node)
{ {
float coeff = -0.999999f; float coeff = -0.999999f;
// float coeff = preferableTarget == DNN_TARGET_MYRIAD ? -0.999f : -0.999999f; // float coeff = preferableTarget == DNN_TARGET_MYRIAD ? -0.999f : -0.999999f;
@ -1603,7 +1603,7 @@ struct SqrtFunctor : public BaseDefaultFunctor<SqrtFunctor>
#endif // HAVE_HALIDE #endif // HAVE_HALIDE
#ifdef HAVE_DNN_NGRAPH #ifdef HAVE_DNN_NGRAPH
std::shared_ptr<ngraph::Node> initNgraphAPI(const std::shared_ptr<ngraph::Node>& node) std::shared_ptr<ngraph::Node> initNgraphAPI(const ngraph::Output<ngraph::Node>& node)
{ {
return std::make_shared<ngraph::op::v0::Sqrt>(node); return std::make_shared<ngraph::op::v0::Sqrt>(node);
} }
@ -2329,7 +2329,7 @@ struct PowerFunctor : public BaseFunctor
#endif // HAVE_CANN #endif // HAVE_CANN
#ifdef HAVE_DNN_NGRAPH #ifdef HAVE_DNN_NGRAPH
std::shared_ptr<ngraph::Node> initNgraphAPI(const std::shared_ptr<ngraph::Node>& node) std::shared_ptr<ngraph::Node> initNgraphAPI(const ngraph::Output<ngraph::Node>& node)
{ {
auto scale_node = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, auto scale_node = std::make_shared<ngraph::op::Constant>(ngraph::element::f32,
ngraph::Shape{1}, &scale); ngraph::Shape{1}, &scale);
@ -2439,7 +2439,7 @@ struct ExpFunctor : public BaseDefaultFunctor<ExpFunctor>
#endif // HAVE_HALIDE #endif // HAVE_HALIDE
#ifdef HAVE_DNN_NGRAPH #ifdef HAVE_DNN_NGRAPH
std::shared_ptr<ngraph::Node> initNgraphAPI(const std::shared_ptr<ngraph::Node>& node) std::shared_ptr<ngraph::Node> initNgraphAPI(const ngraph::Output<ngraph::Node>& node)
{ {
auto scale_node = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, auto scale_node = std::make_shared<ngraph::op::Constant>(ngraph::element::f32,
ngraph::Shape{1}, &normScale); ngraph::Shape{1}, &normScale);
@ -2598,7 +2598,7 @@ struct ChannelsPReLUFunctor : public BaseFunctor
#endif // HAVE_CANN #endif // HAVE_CANN
#ifdef HAVE_DNN_NGRAPH #ifdef HAVE_DNN_NGRAPH
std::shared_ptr<ngraph::Node> initNgraphAPI(const std::shared_ptr<ngraph::Node>& node) std::shared_ptr<ngraph::Node> initNgraphAPI(const ngraph::Output<ngraph::Node>& node)
{ {
const size_t numChannels = scale.total(); const size_t numChannels = scale.total();
auto slope = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape{numChannels}, scale.data); auto slope = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape{numChannels}, scale.data);
@ -2678,7 +2678,7 @@ struct PReLUFunctor : public ChannelsPReLUFunctor
} }
#ifdef HAVE_DNN_NGRAPH #ifdef HAVE_DNN_NGRAPH
std::shared_ptr<ngraph::Node> initNgraphAPI(const std::shared_ptr<ngraph::Node>& node) std::shared_ptr<ngraph::Node> initNgraphAPI(const ngraph::Output<ngraph::Node>& node)
{ {
auto shape = getShape<size_t>(scale); auto shape = getShape<size_t>(scale);
auto slope = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, shape, scale.ptr<float>()); auto slope = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, shape, scale.ptr<float>());

15
modules/dnn/src/layers/eltwise_layer.cpp
Unescape View File

@ -896,12 +896,14 @@ public:
virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >& inputs, virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >& inputs,
const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE
{ {
CV_Assert(nodes.size() >= 2);
auto curr_node = nodes[0].dynamicCast<InfEngineNgraphNode>()->node; auto curr_node = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
if (!coeffs.empty()) { if (!coeffs.empty()) {
auto coeff = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape{1}, &coeffs[0]); auto coeff = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape{1}, &coeffs[0]);
curr_node = std::make_shared<ngraph::op::v1::Multiply>(curr_node, coeff, ngraph::op::AutoBroadcastType::NUMPY); curr_node = std::make_shared<ngraph::op::v1::Multiply>(curr_node, coeff, ngraph::op::AutoBroadcastType::NUMPY);
} }
std::shared_ptr<ngraph::Node> res;
for (size_t i = 1; i < nodes.size(); i++) for (size_t i = 1; i < nodes.size(); i++)
{ {
auto next_node = nodes[i].dynamicCast<InfEngineNgraphNode>()->node; auto next_node = nodes[i].dynamicCast<InfEngineNgraphNode>()->node;
@ -910,15 +912,16 @@ public:
next_node = std::make_shared<ngraph::op::v1::Multiply>(next_node, coeff, ngraph::op::AutoBroadcastType::NUMPY); next_node = std::make_shared<ngraph::op::v1::Multiply>(next_node, coeff, ngraph::op::AutoBroadcastType::NUMPY);
} }
switch (op) { switch (op) {
case SUM: curr_node = std::make_shared<ngraph::op::v1::Add>(curr_node, next_node); break; case SUM: res = std::make_shared<ngraph::op::v1::Add>(curr_node, next_node); break;
case PROD: curr_node = std::make_shared<ngraph::op::v1::Multiply>(curr_node, next_node); break; case PROD: res = std::make_shared<ngraph::op::v1::Multiply>(curr_node, next_node); break;
case DIV: curr_node = std::make_shared<ngraph::op::v1::Divide>(curr_node, next_node); break; case DIV: res = std::make_shared<ngraph::op::v1::Divide>(curr_node, next_node); break;
case MAX: curr_node = std::make_shared<ngraph::op::v1::Maximum>(curr_node, next_node); break; case MAX: res = std::make_shared<ngraph::op::v1::Maximum>(curr_node, next_node); break;
case MIN: curr_node = std::make_shared<ngraph::op::v1::Minimum>(curr_node, next_node); break; case MIN: res = std::make_shared<ngraph::op::v1::Minimum>(curr_node, next_node); break;
default: CV_Error(Error::StsNotImplemented, "Unsupported eltwise operation"); default: CV_Error(Error::StsNotImplemented, "Unsupported eltwise operation");
} }
curr_node = res;
} }
return Ptr<BackendNode>(new InfEngineNgraphNode(curr_node)); return Ptr<BackendNode>(new InfEngineNgraphNode(res));
} }
#endif // HAVE_DNN_NGRAPH #endif // HAVE_DNN_NGRAPH

2
modules/dnn/src/layers/flatten_layer.cpp
Unescape View File

@ -209,7 +209,7 @@ public:
const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE
{ {
auto& ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node; auto& ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
std::vector<size_t> dims = ieInpNode->get_shape(); std::vector<size_t> dims = ieInpNode.get_shape();
int numAxes = dims.size(); int numAxes = dims.size();
int startAxis = normalize_axis(_startAxis, numAxes); int startAxis = normalize_axis(_startAxis, numAxes);

2
modules/dnn/src/layers/fully_connected_layer.cpp
Unescape View File

@ -803,7 +803,7 @@ public:
} }
else else
{ {
std::vector<int> shape(1 + normalize_axis(axis, ieInpNode->get_shape().size()), 0); std::vector<int> shape(1 + normalize_axis(axis, ieInpNode.get_shape().size()), 0);
shape[shape.size() - 1] = -1; shape[shape.size() - 1] = -1;
auto inp = std::make_shared<ngraph::op::v1::Reshape>( auto inp = std::make_shared<ngraph::op::v1::Reshape>(
ieInpNode, ieInpNode,

2
modules/dnn/src/layers/lrn_layer.cpp
Unescape View File

@ -480,7 +480,7 @@ public:
if (type != SPATIAL_NRM) { if (type != SPATIAL_NRM) {
axes = {1}; axes = {1};
} else { } else {
axes.resize(ieInpNode->get_shape().size() - 2); axes.resize(ieInpNode.get_shape().size() - 2);
std::iota(axes.begin(), axes.end(), 2); std::iota(axes.begin(), axes.end(), 2);
} }
auto ngraph_axes = std::make_shared<ngraph::op::Constant>(ngraph::element::i64, ngraph::Shape{axes.size()}, axes.data()); auto ngraph_axes = std::make_shared<ngraph::op::Constant>(ngraph::element::i64, ngraph::Shape{axes.size()}, axes.data());

4
modules/dnn/src/layers/max_unpooling_layer.cpp
Unescape View File

@ -194,7 +194,7 @@ public:
std::vector<MatShape> inpShapes(nodes.size()); std::vector<MatShape> inpShapes(nodes.size());
std::vector<MatShape> outShapes, internals; std::vector<MatShape> outShapes, internals;
for (int i = 0; i < nodes.size(); ++i) { for (int i = 0; i < nodes.size(); ++i) {
std::vector<size_t> shape = nodes[i].dynamicCast<InfEngineNgraphNode>()->node->get_shape(); std::vector<size_t> shape = nodes[i].dynamicCast<InfEngineNgraphNode>()->node.get_shape();
inpShapes[i] = std::vector<int>(shape.begin(), shape.end()); inpShapes[i] = std::vector<int>(shape.begin(), shape.end());
} }
getMemoryShapes(inpShapes, 1, outShapes, internals); getMemoryShapes(inpShapes, 1, outShapes, internals);
@ -213,7 +213,7 @@ public:
std::make_shared<ngraph::op::Constant>(ngraph::element::i32, ngraph::Shape{1}, &newShape), std::make_shared<ngraph::op::Constant>(ngraph::element::i32, ngraph::Shape{1}, &newShape),
true true
); );
if (indices->get_element_type() != ngraph::element::i32 && indices->get_element_type() != ngraph::element::i64) { if (indices.get_element_type() != ngraph::element::i32 && indices.get_element_type() != ngraph::element::i64) {
indices = std::make_shared<ngraph::op::Convert>(indices, ngraph::element::i64); indices = std::make_shared<ngraph::op::Convert>(indices, ngraph::element::i64);
} }

2
modules/dnn/src/layers/mvn_layer.cpp
Unescape View File

@ -390,7 +390,7 @@ public:
auto mvn = std::make_shared<ngraph::op::MVN>(ieInpNode, acrossChannels, normVariance, eps); auto mvn = std::make_shared<ngraph::op::MVN>(ieInpNode, acrossChannels, normVariance, eps);
#else #else
int64_t start_axis = acrossChannels ? 1 : 2; int64_t start_axis = acrossChannels ? 1 : 2;
std::vector<int64_t> axes_v(ieInpNode->get_shape().size() - start_axis); std::vector<int64_t> axes_v(ieInpNode.get_shape().size() - start_axis);
std::iota(axes_v.begin(), axes_v.end(), start_axis); std::iota(axes_v.begin(), axes_v.end(), start_axis);
auto axes = std::make_shared<ngraph::op::Constant>(ngraph::element::i64, ngraph::Shape{axes_v.size()}, axes_v.data()); auto axes = std::make_shared<ngraph::op::Constant>(ngraph::element::i64, ngraph::Shape{axes_v.size()}, axes_v.data());
auto mvn = std::make_shared<ngraph::op::v6::MVN>(ieInpNode, axes, normVariance, eps, ngraph::op::MVNEpsMode::INSIDE_SQRT); auto mvn = std::make_shared<ngraph::op::v6::MVN>(ieInpNode, axes, normVariance, eps, ngraph::op::MVNEpsMode::INSIDE_SQRT);

6
modules/dnn/src/layers/nary_eltwise_layers.cpp
Unescape View File

@ -900,12 +900,12 @@ public:
auto& inp0 = nodes[0].dynamicCast<InfEngineNgraphNode>()->node; auto& inp0 = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
auto& inp1 = nodes[1].dynamicCast<InfEngineNgraphNode>()->node; auto& inp1 = nodes[1].dynamicCast<InfEngineNgraphNode>()->node;
if (inp0->get_element_type() != inp1->get_element_type()) { if (inp0.get_element_type() != inp1.get_element_type()) {
auto dtype = preferableTarget == DNN_TARGET_OPENCL_FP16 || preferableTarget == DNN_TARGET_MYRIAD ? auto dtype = preferableTarget == DNN_TARGET_OPENCL_FP16 || preferableTarget == DNN_TARGET_MYRIAD ?
ngraph::element::f16 : ngraph::element::f32; ngraph::element::f16 : ngraph::element::f32;
if (inp0->get_element_type() != dtype) if (inp0.get_element_type() != dtype)
inp0 = std::make_shared<ngraph::op::v0::Convert>(inp0, dtype); inp0 = std::make_shared<ngraph::op::v0::Convert>(inp0, dtype);
if (inp1->get_element_type() != dtype) if (inp1.get_element_type() != dtype)
inp1 = std::make_shared<ngraph::op::v0::Convert>(inp1, dtype); inp1 = std::make_shared<ngraph::op::v0::Convert>(inp1, dtype);
} }

8
modules/dnn/src/layers/normalize_bbox_layer.cpp
Unescape View File

@ -273,21 +273,21 @@ public:
const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE
{ {
auto& ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node; auto& ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
const size_t batch = ieInpNode->get_shape()[0]; const size_t batch = ieInpNode.get_shape()[0];
const size_t numChannels = ieInpNode->get_shape()[1]; const size_t numChannels = ieInpNode.get_shape()[1];
std::vector<int64_t> axes_data; std::vector<int64_t> axes_data;
if (!acrossSpatial) { if (!acrossSpatial) {
axes_data.push_back(1); axes_data.push_back(1);
} else { } else {
axes_data.resize(ieInpNode->get_shape().size() - 1); axes_data.resize(ieInpNode.get_shape().size() - 1);
std::iota(axes_data.begin(), axes_data.end(), 1); std::iota(axes_data.begin(), axes_data.end(), 1);
} }
auto axes = std::make_shared<ngraph::op::Constant>(ngraph::element::i64, ngraph::Shape{axes_data.size()}, axes_data); auto axes = std::make_shared<ngraph::op::Constant>(ngraph::element::i64, ngraph::Shape{axes_data.size()}, axes_data);
auto norm = std::make_shared<ngraph::op::v0::NormalizeL2>(ieInpNode, axes, epsilon, ngraph::op::EpsMode::ADD); auto norm = std::make_shared<ngraph::op::v0::NormalizeL2>(ieInpNode, axes, epsilon, ngraph::op::EpsMode::ADD);
CV_Assert(blobs.empty() || numChannels == blobs[0].total()); CV_Assert(blobs.empty() || numChannels == blobs[0].total());
std::vector<size_t> shape(ieInpNode->get_shape().size(), 1); std::vector<size_t> shape(ieInpNode.get_shape().size(), 1);
shape[0] = blobs.empty() ? 1 : batch; shape[0] = blobs.empty() ? 1 : batch;
shape[1] = numChannels; shape[1] = numChannels;
if (!blobs.empty()) if (!blobs.empty())

2
modules/dnn/src/layers/pooling_layer.cpp
Unescape View File

@ -601,7 +601,7 @@ public:
return Ptr<BackendNode>(new InfEngineNgraphNode(ave_pool)); return Ptr<BackendNode>(new InfEngineNgraphNode(ave_pool));
} }
else if (type == SUM) { else if (type == SUM) {
ngraph::Shape inpShape = ieInpNode->get_shape(); ngraph::Shape inpShape = ieInpNode.get_shape();
CV_Assert(inpShape.size() == 2 + kernel_size.size()); CV_Assert(inpShape.size() == 2 + kernel_size.size());
std::vector<int64_t> axes; std::vector<int64_t> axes;
for (size_t i = 0; i < kernel_size.size(); i++) for (size_t i = 0; i < kernel_size.size(); i++)

4
modules/dnn/src/layers/proposal_layer.cpp
Unescape View File

@ -366,10 +366,10 @@ public:
auto& class_logits = nodes[1].dynamicCast<InfEngineNgraphNode>()->node; auto& class_logits = nodes[1].dynamicCast<InfEngineNgraphNode>()->node;
auto& image_shape = nodes[2].dynamicCast<InfEngineNgraphNode>()->node; auto& image_shape = nodes[2].dynamicCast<InfEngineNgraphNode>()->node;
CV_Assert_N(image_shape->get_shape().size() == 2, image_shape->get_shape().front() == 1); CV_Assert_N(image_shape.get_shape().size() == 2, image_shape.get_shape().front() == 1);
auto shape = std::make_shared<ngraph::op::Constant>(ngraph::element::i64, auto shape = std::make_shared<ngraph::op::Constant>(ngraph::element::i64,
ngraph::Shape{1}, ngraph::Shape{1},
std::vector<int64_t>{(int64_t)image_shape->get_shape().back()}); std::vector<int64_t>{(int64_t)image_shape.get_shape().back()});
auto reshape = std::make_shared<ngraph::op::v1::Reshape>(image_shape, shape, true); auto reshape = std::make_shared<ngraph::op::v1::Reshape>(image_shape, shape, true);
auto proposal = std::make_shared<ngraph::op::Proposal>(class_probs, class_logits, reshape, attr); auto proposal = std::make_shared<ngraph::op::Proposal>(class_probs, class_logits, reshape, attr);

4
modules/dnn/src/layers/region_layer.cpp
Unescape View File

@ -466,7 +466,7 @@ public:
const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE
{ {
auto& input = nodes[0].dynamicCast<InfEngineNgraphNode>()->node; auto& input = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
auto parent_shape = input->get_shape(); auto parent_shape = input.get_shape();
int64_t b = parent_shape[0]; int64_t b = parent_shape[0];
int64_t h = parent_shape[1]; int64_t h = parent_shape[1];
int64_t w = parent_shape[2]; int64_t w = parent_shape[2];
@ -567,7 +567,7 @@ public:
int hNorm, wNorm; int hNorm, wNorm;
if (nodes.size() > 1) if (nodes.size() > 1)
{ {
auto node_1_shape = nodes[1].dynamicCast<InfEngineNgraphNode>()->node->get_shape(); auto node_1_shape = nodes[1].dynamicCast<InfEngineNgraphNode>()->node.get_shape();
hNorm = node_1_shape[2]; hNorm = node_1_shape[2];
wNorm = node_1_shape[3]; wNorm = node_1_shape[3];
} }

2
modules/dnn/src/layers/resize_layer.cpp
Unescape View File

@ -443,7 +443,7 @@ public:
std::vector<int64_t> shape = {outHeight, outWidth}; std::vector<int64_t> shape = {outHeight, outWidth};
auto out_shape = std::make_shared<ngraph::op::Constant>(ngraph::element::i64, ngraph::Shape{2}, shape.data()); auto out_shape = std::make_shared<ngraph::op::Constant>(ngraph::element::i64, ngraph::Shape{2}, shape.data());
auto& input_shape = ieInpNode->get_shape(); auto& input_shape = ieInpNode.get_shape();
CV_Assert_N(input_shape[2] != 0, input_shape[3] != 0); CV_Assert_N(input_shape[2] != 0, input_shape[3] != 0);
std::vector<float> scales = {static_cast<float>(outHeight) / input_shape[2], static_cast<float>(outWidth) / input_shape[3]}; std::vector<float> scales = {static_cast<float>(outHeight) / input_shape[2], static_cast<float>(outWidth) / input_shape[3]};
auto scales_shape = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape{2}, scales.data()); auto scales_shape = std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape{2}, scales.data());

18
modules/dnn/src/layers/scale_layer.cpp
Unescape View File

@ -331,34 +331,36 @@ public:
virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >& inputs, const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >& inputs, const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE
{ {
auto ieInpNode0 = nodes[0].dynamicCast<InfEngineNgraphNode>()->node; auto ieInpNode0 = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
auto ieInpNode1 = nodes.size() > 1 ? nodes[1].dynamicCast<InfEngineNgraphNode>()->node : nullptr; ngraph::Output<ngraph::Node> ieInpNode1;
if (nodes.size() > 1)
ieInpNode1 = nodes[1].dynamicCast<InfEngineNgraphNode>()->node;
size_t numChannels = 1; size_t numChannels = 1;
if (blobs.empty()) if (blobs.empty())
for (const size_t& dim : ieInpNode1->get_shape()) for (const size_t& dim : ieInpNode1.get_shape())
numChannels *= dim; numChannels *= dim;
else else
numChannels = blobs[0].total(); numChannels = blobs[0].total();
std::vector<size_t> shape(ieInpNode0->get_shape().size(), 1); std::vector<size_t> shape(ieInpNode0.get_shape().size(), 1);
int cAxis = normalize_axis(axis, shape.size()); int cAxis = normalize_axis(axis, shape.size());
shape[cAxis] = numChannels; shape[cAxis] = numChannels;
auto node = ieInpNode0; std::shared_ptr<ngraph::Node> node;
if (hasWeights) if (hasWeights)
{ {
auto weight = blobs.empty() ? ieInpNode1 : ngraph::Output<ngraph::Node> weight = blobs.empty() ? ieInpNode1 :
std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape(shape), blobs[0].data); std::make_shared<ngraph::op::Constant>(ngraph::element::f32, ngraph::Shape(shape), blobs[0].data);
#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2021_2) #if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2021_2)
node = std::make_shared<ngraph::op::v1::Multiply>(node, weight, ngraph::op::AutoBroadcastType::NUMPY); node = std::make_shared<ngraph::op::v1::Multiply>(ieInpNode0, weight, ngraph::op::AutoBroadcastType::NUMPY);
#else #else
node = std::make_shared<ngraph::op::v0::Multiply>(node, weight, ngraph::op::AutoBroadcastType::NUMPY); node = std::make_shared<ngraph::op::v0::Multiply>(ieInpNode0, weight, ngraph::op::AutoBroadcastType::NUMPY);
#endif #endif
} }
if (hasBias || !hasWeights) if (hasBias || !hasWeights)
{ {
std::shared_ptr<ngraph::Node> bias; ngraph::Output<ngraph::Node> bias;
if (hasBias) if (hasBias)
{ {
bias = blobs.empty() ? ieInpNode1 : bias = blobs.empty() ? ieInpNode1 :

2
modules/dnn/src/layers/slice_layer.cpp
Unescape View File

@ -759,7 +759,7 @@ public:
{ {
CV_Assert_N(nodes.size() <= 2); CV_Assert_N(nodes.size() <= 2);
auto& ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node; auto& ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
CV_Assert(finalSliceRanges[0].size() == ieInpNode->get_shape().size()); CV_Assert(finalSliceRanges[0].size() == ieInpNode.get_shape().size());
std::vector<int64_t> offsets, dims; std::vector<int64_t> offsets, dims;
for (int i = 0; i < finalSliceRanges[0].size(); ++i) for (int i = 0; i < finalSliceRanges[0].size(); ++i)

2
modules/dnn/src/layers/softmax_layer.cpp
Unescape View File

@ -385,7 +385,7 @@ public:
const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE
{ {
auto& ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node; auto& ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
int axis = normalize_axis(axisRaw, ieInpNode->get_shape().size()); int axis = normalize_axis(axisRaw, ieInpNode.get_shape().size());
auto softmax = std::make_shared<ngraph::op::v1::Softmax>(ieInpNode, axis); auto softmax = std::make_shared<ngraph::op::v1::Softmax>(ieInpNode, axis);
if (logSoftMax) if (logSoftMax)
return Ptr<BackendNode>(new InfEngineNgraphNode(std::make_shared<ngraph::op::v0::Log>(softmax))); return Ptr<BackendNode>(new InfEngineNgraphNode(std::make_shared<ngraph::op::v0::Log>(softmax)));

63
modules/dnn/src/net_openvino.cpp
Unescape View File

@ -321,8 +321,10 @@ void NetImplOpenVINO::initBackend(const std::vector<LayerPin>& blobsToKeep_)
return; return;
} }
#if INF_ENGINE_VER_MAJOR_LT(INF_ENGINE_RELEASE_2022_1)
bool supportsCPUFallback = !isArmComputePlugin() && (preferableTarget == DNN_TARGET_CPU || bool supportsCPUFallback = !isArmComputePlugin() && (preferableTarget == DNN_TARGET_CPU ||
openvino::checkTarget(DNN_TARGET_CPU)); openvino::checkTarget(DNN_TARGET_CPU));
#endif
// Build Inference Engine networks from sets of layers that support this // Build Inference Engine networks from sets of layers that support this
// backend. Split a whole model on several Inference Engine networks if // backend. Split a whole model on several Inference Engine networks if
@ -341,6 +343,10 @@ void NetImplOpenVINO::initBackend(const std::vector<LayerPin>& blobsToKeep_)
bool fused = ld.skip; bool fused = ld.skip;
Ptr<Layer> layer = ld.layerInstance; Ptr<Layer> layer = ld.layerInstance;
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2022_1)
if (ld.id == 0)
continue;
#else
if (!fused && !layer->supportBackend(preferableBackend)) if (!fused && !layer->supportBackend(preferableBackend))
{ {
CV_LOG_DEBUG(NULL, "DNN/IE: NOT supported!"); CV_LOG_DEBUG(NULL, "DNN/IE: NOT supported!");
@ -355,17 +361,6 @@ void NetImplOpenVINO::initBackend(const std::vector<LayerPin>& blobsToKeep_)
} }
} }
// TODO: fix these workarounds
if (preferableTarget == DNN_TARGET_MYRIAD ||
preferableTarget == DNN_TARGET_HDDL ||
preferableTarget == DNN_TARGET_OPENCL ||
preferableTarget == DNN_TARGET_OPENCL_FP16)
customizable &= ld.type != "Concat";
if (preferableTarget == DNN_TARGET_OPENCL ||
preferableTarget == DNN_TARGET_OPENCL_FP16)
customizable &= ld.type != "Power";
if (preferableTarget == DNN_TARGET_OPENCL) if (preferableTarget == DNN_TARGET_OPENCL)
customizable &= ld.type != "Eltwise"; customizable &= ld.type != "Eltwise";
@ -390,6 +385,7 @@ void NetImplOpenVINO::initBackend(const std::vector<LayerPin>& blobsToKeep_)
continue; continue;
} }
} }
#endif
ld.skip = true; // Initially skip all Inference Engine supported layers. ld.skip = true; // Initially skip all Inference Engine supported layers.
// Create a new network if one of inputs from different Inference Engine graph. // Create a new network if one of inputs from different Inference Engine graph.
@ -478,7 +474,7 @@ void NetImplOpenVINO::initBackend(const std::vector<LayerPin>& blobsToKeep_)
int oid = ld.inputBlobsId[i].oid; int oid = ld.inputBlobsId[i].oid;
auto ieInpNode = inputNodes[i].dynamicCast<InfEngineNgraphNode>(); auto ieInpNode = inputNodes[i].dynamicCast<InfEngineNgraphNode>();
const auto& ngraph_input_node = ieInpNode->node; const auto& ngraph_input_node = ieInpNode->node.get_node_shared_ptr();
CV_LOG_DEBUG(NULL, "DNN/IE: bind output port " << lid << ":" << oid << " (" << ngraph_input_node->get_friendly_name() << ":" << ngraph_input_node->get_type_info().name << ")"); CV_LOG_DEBUG(NULL, "DNN/IE: bind output port " << lid << ":" << oid << " (" << ngraph_input_node->get_friendly_name() << ":" << ngraph_input_node->get_type_info().name << ")");
if ((oid == 0 && ngraph_input_node->get_output_size() == 1) || lid == 0) if ((oid == 0 && ngraph_input_node->get_output_size() == 1) || lid == 0)
@ -498,10 +494,7 @@ void NetImplOpenVINO::initBackend(const std::vector<LayerPin>& blobsToKeep_)
} }
CV_CheckLT((size_t)oid, ngraph_input_node->get_output_size(), ""); CV_CheckLT((size_t)oid, ngraph_input_node->get_output_size(), "");
#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2020_4) #if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2020_4)
// FIXIT refactor ".initNgraph()" API to use Output<Node> inputNodes[i] = new InfEngineNgraphNode(ngraph_input_node->output(oid));
// WA: use Concat to emulate Identity operation with requested output port
auto oid_node = std::make_shared<ngraph::op::Concat>(ngraph::OutputVector { ngraph_input_node->output(oid) }, 0);
inputNodes[i] = Ptr<BackendNode>(new InfEngineNgraphNode(oid_node));
#elif INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2020_3) #elif INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2020_3)
inputNodes[i] = Ptr<BackendNode>(new InfEngineNgraphNode(ieInpNode->node->get_output_as_single_output_node(oid))); inputNodes[i] = Ptr<BackendNode>(new InfEngineNgraphNode(ieInpNode->node->get_output_as_single_output_node(oid)));
#else #else
@ -556,6 +549,36 @@ void NetImplOpenVINO::initBackend(const std::vector<LayerPin>& blobsToKeep_)
addNgraphOutputs(ld); addNgraphOutputs(ld);
} }
// User may choose to return only intermediate blobs but not network's result (see Test_TFLite.max_unpooling)
// Such layers should not be skipped when forwardLayer is called.
// Also, perform a sanity check that there is no double inferred networks (a single skip=false per unique net instance)
std::set<Ptr<InfEngineNgraphNet>> uniqueNets;
if (!blobsToKeep_.empty())
{
LayerPin latestLayerPin = getLatestLayerPin(blobsToKeep_);
for (MapIdToLayerData::iterator it = layers.begin(); it != layers.end(); ++it)
{
LayerData& ld = it->second;
auto iter = ld.backendNodes.find(preferableBackend);
if (iter == ld.backendNodes.end())
continue;
Ptr<BackendNode>& node = iter->second;
if (node.empty())
continue;
Ptr<InfEngineNgraphNode> ieNode = node.dynamicCast<InfEngineNgraphNode>();
if (ieNode.empty())
continue;
if (ld.id == latestLayerPin.lid) {
ld.skip = false;
uniqueNets.insert(ieNode->net);
break;
}
}
}
// Initialize all networks. // Initialize all networks.
for (MapIdToLayerData::reverse_iterator it = layers.rbegin(); it != layers.rend(); ++it) for (MapIdToLayerData::reverse_iterator it = layers.rbegin(); it != layers.rend(); ++it)
{ {
@ -578,9 +601,15 @@ void NetImplOpenVINO::initBackend(const std::vector<LayerPin>& blobsToKeep_)
{ {
ieNode->net->addOutput(ieNode); ieNode->net->addOutput(ieNode);
ieNode->net->createNet((Target)preferableTarget); ieNode->net->createNet((Target)preferableTarget);
ld.skip = false; if (uniqueNets.find(ieNode->net) == uniqueNets.end()) {
ld.skip = false;
uniqueNets.insert(ieNode->net);
}
} }
} }
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2022_1)
CV_Assert(uniqueNets.size() == 1);
#endif
} }

11
modules/dnn/test/test_halide_layers.cpp
Unescape View File

@ -425,6 +425,13 @@ TEST_P(FullyConnected, Accuracy)
applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X); applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X);
} }
#endif #endif
// https://github.com/openvinotoolkit/openvino/issues/19436
if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && targetId == DNN_TARGET_OPENCL_FP16 && batch == 16)
applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL_FP16);
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_EQ(2023000000)
if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && targetId == DNN_TARGET_OPENCL && batch == 16)
applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL);
#endif
Mat weights(outChannels, inChannels * inSize.height * inSize.width, CV_32F); Mat weights(outChannels, inChannels * inSize.height * inSize.width, CV_32F);
randu(weights, -1.0f, 1.0f); randu(weights, -1.0f, 1.0f);
@ -454,11 +461,13 @@ TEST_P(FullyConnected, Accuracy)
if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && targetId == DNN_TARGET_OPENCL_FP16) if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && targetId == DNN_TARGET_OPENCL_FP16)
{ {
l1 = 0.01; l1 = 0.01;
if (INF_ENGINE_VER_MAJOR_GE(2023000000))
lInf = 0.016;
} }
if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && targetId == DNN_TARGET_OPENCL) if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && targetId == DNN_TARGET_OPENCL)
{ {
l1 = 5e-3; l1 = 5e-3;
lInf = 7e-3; lInf = INF_ENGINE_VER_MAJOR_GE(2023000000) ? 0.016 : 7e-3;
} }
#endif #endif
if (targetId == DNN_TARGET_CUDA_FP16) if (targetId == DNN_TARGET_CUDA_FP16)

9
modules/dnn/test/test_tflite_importer.cpp
Unescape View File

@ -157,14 +157,7 @@ TEST_P(Test_TFLite, max_unpooling)
net.setInput(input); net.setInput(input);
std::vector<std::vector<Mat> > outs; std::vector<std::vector<Mat> > outs;
if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) { net.forward(outs, {"p_re_lu_1", "max_pooling_with_argmax2d", "conv2d_86", "max_unpooling2d_2"});
// TODO: seems like a bug with a retrieving intermediate tensors
net.forward(outs, {"conv2d_transpose_4", "p_re_lu_1", "max_pooling_with_argmax2d", "conv2d_86", "max_unpooling2d_2"});
outs.erase(outs.begin());
}
else {
net.forward(outs, {"p_re_lu_1", "max_pooling_with_argmax2d", "conv2d_86", "max_unpooling2d_2"});
}
ASSERT_EQ(outs.size(), 4); ASSERT_EQ(outs.size(), 4);
ASSERT_EQ(outs[0].size(), 1); ASSERT_EQ(outs[0].size(), 1);

Write Preview
Loading…
Cancel
Save