1 // This file is part of OpenCV project.
2 // It is subject to the license terms in the LICENSE file found in the top-level directory
3 // of this distribution and at http://opencv.org/license.html.
5 // Copyright (C) 2018, Intel Corporation, all rights reserved.
6 // Third party copyrights are property of their respective owners.
8 #include "../precomp.hpp"
9 #include <opencv2/dnn/shape_utils.hpp>
11 #include <opencv2/dnn/layer_reg.private.hpp>
13 #include <opencv2/core/utils/fp_control_utils.hpp>
15 #include <opencv2/core/utils/logger.defines.hpp>
16 #undef CV_LOG_STRIP_LEVEL
17 #define CV_LOG_STRIP_LEVEL CV_LOG_LEVEL_VERBOSE + 1
18 #include <opencv2/core/utils/logger.hpp>
20 #include <opencv2/core/utils/configuration.private.hpp>
31 #if defined _MSC_VER && _MSC_VER < 1910/*MSVS 2017*/
33 #pragma warning(disable: 4503) // decorated name length exceeded, name was truncated
36 #if defined(__GNUC__) && __GNUC__ >= 5
37 #pragma GCC diagnostic push
38 #pragma GCC diagnostic ignored "-Wsuggest-override"
40 #include "opencv-onnx.pb.h"
41 #if defined(__GNUC__) && __GNUC__ >= 5
42 #pragma GCC diagnostic pop
45 #include "onnx_graph_simplifier.hpp"
49 CV__DNN_INLINE_NS_BEGIN
51 extern bool DNN_DIAGNOSTICS_RUN;
53 class ONNXLayerHandler;
57 FPDenormalsIgnoreHintScope fp_denormals_ignore_scope;
59 opencv_onnx::ModelProto model_proto;
63 LayerInfo(int _layerId = 0, int _outputId = 0) : layerId(_layerId), outputId(_outputId) {}
66 std::map<std::string, Mat> getGraphTensors(
67 const opencv_onnx::GraphProto& graph_proto);
68 Mat getBlob(const opencv_onnx::NodeProto& node_proto, int index);
69 Mat getBlob(const std::string& input_name);
71 LayerParams getLayerParams(const opencv_onnx::NodeProto& node_proto);
73 void addConstant(const std::string& name, const Mat& blob);
74 void addLayer(LayerParams& layerParams,
75 const opencv_onnx::NodeProto& node_proto);
76 void handleQuantizedNode(LayerParams& layerParams,
77 const opencv_onnx::NodeProto& node_proto);
79 void expandMid(const std::string& prefix, opencv_onnx::NodeProto& node_proto,
80 const std::string& input, size_t n);
81 void addNegation(const LayerParams& layerParams, opencv_onnx::NodeProto& node_proto, int input_id);
82 void lstm_extractConsts(LayerParams& layerParams, const opencv_onnx::NodeProto& lstm_proto, size_t idx, int* blobShape_, int size);
83 void lstm_add_reshape(const std::string& input_name, const std::string& output_name, int* layerShape, size_t n);
84 std::string lstm_add_slice(int index, const std::string& input_name, int* begin, int* end, size_t n);
85 std::string lstm_fix_dims(LayerParams& layerParams, const opencv_onnx::NodeProto& lstm_proto,
86 int batch_size, int num_directions, int hidden_size, bool need_y, const std::string& y_name,
88 void lstm_add_transform(int num_directions, int batch_size, int hidden_size,
89 int index, const std::string& input_name, const std::string& output_name);
91 ONNXImporter(Net& net, const char *onnxFile);
92 ONNXImporter(Net& net, const char* buffer, size_t sizeBuffer);
97 std::unique_ptr<ONNXLayerHandler> layerHandler;
100 opencv_onnx::GraphProto graph_proto;
101 std::string framework_name;
103 std::map<std::string, Mat> constBlobs;
105 std::map<std::string, MatShape> outShapes; // List of internal blobs shapes.
106 bool hasDynamicShapes; // Whether the model has inputs with dynamic shapes
107 typedef std::map<std::string, MatShape>::iterator IterShape_t;
109 std::map<std::string, LayerInfo> layer_id;
110 typedef std::map<std::string, LayerInfo>::iterator IterLayerId_t;
111 typedef std::map<std::string, LayerInfo>::const_iterator ConstIterLayerId_t;
113 void handleNode(const opencv_onnx::NodeProto& node_proto);
116 friend class ONNXLayerHandler;
117 typedef void (ONNXImporter::*ONNXImporterNodeParser)(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
118 typedef std::map<std::string, ONNXImporterNodeParser> DispatchMap;
119 typedef std::map<std::string, DispatchMap> DomainDispatchMap;
121 DomainDispatchMap domain_dispatch_map;
122 std::string getLayerTypeDomain(const opencv_onnx::NodeProto& node_proto);
123 const DispatchMap& getDispatchMap(const opencv_onnx::NodeProto& node_proto);
124 void buildDispatchMap_ONNX_AI(int opset_version);
125 void buildDispatchMap_COM_MICROSOFT(int opset_version);
127 // Domain: 'ai.onnx' (default)
128 // URL: https://github.com/onnx/onnx/blob/master/docs/Operators.md
129 void parseArg (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
130 void parseMaxUnpool (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
131 void parseMaxPool (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
132 void parseAveragePool (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
133 void parseGlobalPool (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
134 void parseReduce (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
135 void parseSlice (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
136 void parseSplit (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
137 void parseBias (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
138 void parsePow (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
139 void parseMinMax (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
140 void parseNeg (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
141 void parseConstant (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
142 void parseLSTM (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
143 void parseGRU (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
144 void parseImageScaler (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
145 void parseClip (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
146 void parseLeakyRelu (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
147 void parseRelu (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
148 void parseElu (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
149 void parseTanh (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
150 void parseAbs (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
151 void parseCompare (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
152 void parsePRelu (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
153 void parseLRN (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
154 void parseInstanceNormalization(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
155 void parseBatchNormalization (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
156 void parseGemm (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
157 void parseMatMul (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
158 void parseMul (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
159 void parseConv (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
160 void parseConvTranspose (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
161 void parseTranspose (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
162 void parseSqueeze (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
163 void parseFlatten (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
164 void parseUnsqueeze (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
165 void parseExpand (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
166 void parseReshape (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
167 void parsePad (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
168 void parseShape (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
169 void parseCast (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
170 void parseConstantFill (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
171 void parseGather (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
172 void parseConcat (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
173 void parseResize (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
174 void parseUpsample (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
175 void parseSoftMax (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
176 void parseDetectionOutput (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
177 void parseCumSum (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
178 void parseDepthToSpace (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
179 void parseSimpleLayers (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
181 // Domain: com.microsoft
182 // URL: https://github.com/microsoft/onnxruntime/blob/master/docs/ContribOperators.md
183 void parseQuantDequant (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
184 void parseQConv (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
185 void parseQMatMul (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
186 void parseQEltwise (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
187 void parseQLeakyRelu (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
188 void parseQSigmoid (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
189 void parseQAvgPool (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
190 void parseQConcat (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
192 // '???' domain or '???' layer type
193 void parseCustomLayer (LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto);
195 int onnx_opset; // OperatorSetIdProto for 'onnx' domain
196 std::map<std::string, int> onnx_opset_map; // map from OperatorSetIdProto
197 void parseOperatorSet();
199 const std::string str_domain_ai_onnx = "ai.onnx";
203 bool getParamUseLegacyNames()
205 bool param = utils::getConfigurationParameterBool("OPENCV_DNN_ONNX_USE_LEGACY_NAMES", false);
208 std::string extractNodeName(const opencv_onnx::NodeProto& node_proto);
212 class ONNXLayerHandler : public detail::LayerHandler
215 explicit ONNXLayerHandler(ONNXImporter* importer_);
217 void fillRegistry(const opencv_onnx::GraphProto& net);
220 ONNXImporter* importer;
223 ONNXLayerHandler::ONNXLayerHandler(ONNXImporter* importer_) : importer(importer_){}
225 void ONNXLayerHandler::fillRegistry(const opencv_onnx::GraphProto &net)
227 int layersSize = net.node_size();
228 for (int li = 0; li < layersSize; li++) {
229 const opencv_onnx::NodeProto &node_proto = net.node(li);
230 const std::string& name = node_proto.output(0);
231 const std::string& type = node_proto.op_type();
232 const std::string& layer_type_domain = importer->getLayerTypeDomain(node_proto);
233 const auto& dispatch = importer->getDispatchMap(node_proto);
234 if (dispatch.find(type) == dispatch.end())
236 addMissing(name, cv::format("%s.%s", layer_type_domain.c_str(), type.c_str()));
242 ONNXImporter::ONNXImporter(Net& net, const char *onnxFile)
243 : layerHandler(DNN_DIAGNOSTICS_RUN ? new ONNXLayerHandler(this) : nullptr)
246 , useLegacyNames(getParamUseLegacyNames())
248 hasDynamicShapes = false;
250 CV_LOG_DEBUG(NULL, "DNN/ONNX: processing ONNX model from file: " << onnxFile);
252 std::fstream input(onnxFile, std::ios::in | std::ios::binary);
255 CV_Error(Error::StsBadArg, cv::format("Can't read ONNX file: %s", onnxFile));
258 if (!model_proto.ParseFromIstream(&input))
260 CV_Error(Error::StsUnsupportedFormat, cv::format("Failed to parse ONNX model: %s", onnxFile));
266 ONNXImporter::ONNXImporter(Net& net, const char* buffer, size_t sizeBuffer)
267 : layerHandler(DNN_DIAGNOSTICS_RUN ? new ONNXLayerHandler(this) : nullptr)
270 , useLegacyNames(getParamUseLegacyNames())
272 hasDynamicShapes = false;
273 CV_LOG_DEBUG(NULL, "DNN/ONNX: processing in-memory ONNX model (" << sizeBuffer << " bytes)");
275 struct _Buf : public std::streambuf
277 _Buf(const char* buffer, size_t sizeBuffer)
279 char* p = const_cast<char*>(buffer);
280 setg(p, p, p + sizeBuffer);
284 _Buf buf(buffer, sizeBuffer);
285 std::istream input(&buf);
287 if (!model_proto.ParseFromIstream(&input))
288 CV_Error(Error::StsUnsupportedFormat, "Failed to parse onnx model from in-memory byte array.");
294 inline void replaceLayerParam(LayerParams& layerParams, const String& oldKey, const String& newKey)
296 if (layerParams.has(oldKey)) {
297 layerParams.set(newKey, layerParams.get(oldKey));
298 layerParams.erase(oldKey);
303 void dumpValueInfoProto(int i, const opencv_onnx::ValueInfoProto& valueInfoProto, const std::string& prefix)
305 CV_Assert(valueInfoProto.has_name());
306 CV_Assert(valueInfoProto.has_type());
307 const opencv_onnx::TypeProto& typeProto = valueInfoProto.type();
308 CV_Assert(typeProto.has_tensor_type());
309 const opencv_onnx::TypeProto::Tensor& tensor = typeProto.tensor_type();
310 CV_Assert(tensor.has_shape());
311 const opencv_onnx::TensorShapeProto& tensorShape = tensor.shape();
313 int dim_size = tensorShape.dim_size();
314 CV_CheckGE(dim_size, 0, "");
315 MatShape shape(dim_size);
316 for (int j = 0; j < dim_size; ++j)
318 const opencv_onnx::TensorShapeProto_Dimension& dimension = tensorShape.dim(j);
319 if (dimension.has_dim_param())
321 CV_LOG_DEBUG(NULL, "DNN/ONNX: " << prefix << "[" << i << "] dim[" << j << "] = <" << dimension.dim_param() << "> (dynamic)");
323 // https://github.com/onnx/onnx/blob/master/docs/DimensionDenotation.md#denotation-definition
324 if (dimension.has_denotation())
326 CV_LOG_INFO(NULL, "DNN/ONNX: " << prefix << "[" << i << "] dim[" << j << "] denotation is '" << dimension.denotation() << "'");
328 shape[j] = dimension.dim_value();
330 CV_LOG_DEBUG(NULL, "DNN/ONNX: " << prefix << "[" << i << " as '" << valueInfoProto.name() << "'] shape=" << toString(shape));
334 void dumpTensorProto(int i, const opencv_onnx::TensorProto& tensorProto, const std::string& prefix)
336 if (utils::logging::getLogLevel() < utils::logging::LOG_LEVEL_VERBOSE)
338 int dim_size = tensorProto.dims_size();
339 CV_CheckGE(dim_size, 0, "");
340 MatShape shape(dim_size);
341 for (int j = 0; j < dim_size; ++j)
343 int sz = static_cast<int>(tensorProto.dims(j));
346 CV_LOG_VERBOSE(NULL, 0, "DNN/ONNX: " << prefix << "[" << i << " as '" << tensorProto.name() << "'] shape=" << toString(shape) << " data_type=" << (int)tensorProto.data_type());
349 void releaseONNXTensor(opencv_onnx::TensorProto& tensor_proto)
351 if (!tensor_proto.raw_data().empty()) {
352 delete tensor_proto.release_raw_data();
356 void runLayer(LayerParams& params, const std::vector<Mat>& inputs,
357 std::vector<Mat>& outputs)
359 Ptr<Layer> layer = LayerFactory::createLayerInstance(params.type, params);
360 CV_Assert((bool)layer);
362 std::vector<MatShape> inpShapes(inputs.size());
363 int ddepth = params.get<int>("depth", CV_32F);
364 for (size_t i = 0; i < inputs.size(); ++i)
366 inpShapes[i] = shape(inputs[i]);
367 if (i > 0 && ddepth != inputs[i].depth())
368 CV_Error(Error::StsNotImplemented, "Mixed input data types.");
369 ddepth = inputs[i].depth();
372 std::vector<MatShape> outShapes, internalShapes;
373 layer->getMemoryShapes(inpShapes, 0, outShapes, internalShapes);
375 std::vector<Mat> internals(internalShapes.size());
376 outputs.resize(outShapes.size());
377 for (size_t i = 0; i < outShapes.size(); ++i)
378 outputs[i].create(outShapes[i], ddepth);
379 for (size_t i = 0; i < internalShapes.size(); ++i)
380 internals[i].create(internalShapes[i], ddepth);
382 layer->finalize(inputs, outputs);
383 layer->forward(inputs, outputs, internals);
386 std::map<std::string, Mat> ONNXImporter::getGraphTensors(
387 const opencv_onnx::GraphProto& graph_proto)
389 std::map<std::string, Mat> layers_weights;
391 for (int i = 0; i < graph_proto.initializer_size(); i++)
393 const opencv_onnx::TensorProto& tensor_proto = graph_proto.initializer(i);
394 dumpTensorProto(i, tensor_proto, "initializer");
395 Mat mat = getMatFromTensor(tensor_proto);
396 releaseONNXTensor(const_cast<opencv_onnx::TensorProto&>(tensor_proto)); // drop already loaded data
398 if (DNN_DIAGNOSTICS_RUN && mat.empty())
401 layers_weights.insert(std::make_pair(tensor_proto.name(), mat));
403 return layers_weights;
406 static DictValue parse(const ::google::protobuf::RepeatedField< ::google::protobuf::int64>& src) {
407 std::vector<int32_t> dst(src.size());
408 convertInt64ToInt32(src, dst, src.size());
409 return DictValue::arrayInt(&dst[0], src.size());
412 static DictValue parseStr(const ::google::protobuf::RepeatedPtrField< ::std::string>& src) {
413 return DictValue::arrayString(src.begin(), static_cast<int>(src.size()));
416 LayerParams ONNXImporter::getLayerParams(const opencv_onnx::NodeProto& node_proto)
419 for(int i = 0; i < node_proto.attribute_size(); i++)
421 opencv_onnx::AttributeProto attribute_proto = node_proto.attribute(i);
422 std::string attribute_name = attribute_proto.name();
426 if(attribute_name == "kernel_shape")
428 CV_Assert(attribute_proto.ints_size() == 1 || attribute_proto.ints_size() == 2 || attribute_proto.ints_size() == 3);
429 lp.set("kernel_size", parse(attribute_proto.ints()));
431 else if(attribute_name == "strides")
433 CV_Assert(attribute_proto.ints_size() == 1 || attribute_proto.ints_size() == 2 || attribute_proto.ints_size() == 3);
434 lp.set("stride", parse(attribute_proto.ints()));
436 else if(attribute_name == "pads")
438 if (node_proto.op_type() == "Pad")
441 // Paddings are in order begin0, begin1, .. beginN, end0, end1, ..., endN.
442 // We need to shuffle it to begin0, end0, begin1, end1, ...
443 CV_Assert(attribute_proto.ints_size() % 2 == 0);
444 const int dims = attribute_proto.ints_size() / 2;
445 std::vector<int32_t> paddings;
446 paddings.reserve(attribute_proto.ints_size());
447 for (int i = 0; i < dims; ++i)
449 paddings.push_back(attribute_proto.ints(i));
450 paddings.push_back(attribute_proto.ints(dims + i));
452 lp.set("paddings", DictValue::arrayInt(&paddings[0], paddings.size()));
456 // Convolution or pooling.
457 CV_Assert(attribute_proto.ints_size() == 2 || attribute_proto.ints_size() == 4 || attribute_proto.ints_size() == 6);
458 lp.set("pad", parse(attribute_proto.ints()));
461 else if(attribute_name == "auto_pad")
463 if (attribute_proto.s() == "SAME_UPPER" || attribute_proto.s() == "SAME_LOWER") {
464 lp.set("pad_mode", "SAME");
466 else if (attribute_proto.s() == "VALID") {
467 lp.set("pad_mode", "VALID");
470 else if(attribute_name == "dilations")
472 CV_Assert(attribute_proto.ints_size() == 1 || attribute_proto.ints_size() == 2 || attribute_proto.ints_size() == 3);
473 lp.set("dilation", parse(attribute_proto.ints()));
475 else if(attribute_name == "activations" && node_proto.op_type() == "LSTM")
477 lp.set(attribute_name, parseStr(attribute_proto.strings()));
479 else if (attribute_proto.has_i())
481 ::google::protobuf::int64 src = attribute_proto.i();
482 if (src < std::numeric_limits<int32_t>::min() || src > std::numeric_limits<int32_t>::max())
483 CV_Error(Error::StsOutOfRange, "Input is out of OpenCV 32S range");
485 lp.set(attribute_name, saturate_cast<int32_t>(src));
487 else if (attribute_proto.has_f())
489 lp.set(attribute_name, attribute_proto.f());
491 else if (attribute_proto.has_s())
493 lp.set(attribute_name, attribute_proto.s());
495 else if (attribute_proto.floats_size() > 0)
497 lp.set(attribute_name, DictValue::arrayReal(
498 attribute_proto.floats().data(), attribute_proto.floats_size()));
500 else if (attribute_proto.ints_size() > 0)
502 lp.set(attribute_name, parse(attribute_proto.ints()));
504 else if (attribute_proto.has_t())
506 opencv_onnx::TensorProto tensor = attribute_proto.t();
507 Mat blob = getMatFromTensor(tensor);
508 lp.blobs.push_back(blob);
510 else if (attribute_proto.has_g())
512 CV_Error(Error::StsNotImplemented, cv::format("DNN/ONNX/Attribute[%s]: 'Graph' is not supported", attribute_name.c_str()));
514 else if (attribute_proto.graphs_size() > 0)
516 CV_Error(Error::StsNotImplemented,
517 cv::format("DNN/ONNX/Attribute[%s]: 'Graphs' (%d) in attributes is not supported",
518 attribute_name.c_str(), attribute_proto.graphs_size())
521 else if (attribute_proto.strings_size() > 0)
523 std::string msg = cv::format("DNN/ONNX/Attribute[%s]: 'Strings' (%d) are not supported",
524 attribute_name.c_str(), attribute_proto.strings_size());
525 CV_LOG_ERROR(NULL, msg);
526 for (int i = 0; i < attribute_proto.strings_size(); i++)
528 CV_LOG_ERROR(NULL, " Attribute[" << attribute_name << "].string(" << i << ") = '" << attribute_proto.strings(i) << "'");
530 CV_Error(Error::StsNotImplemented, msg);
532 else if (attribute_proto.tensors_size() > 0)
534 CV_Error(Error::StsNotImplemented,
535 cv::format("DNN/ONNX/Attribute[%s]: 'Tensors' (%d) in attributes are not supported",
536 attribute_name.c_str(), attribute_proto.tensors_size())
541 CV_Error(Error::StsNotImplemented, cv::format("DNN/ONNX/Attribute[%s]: unsupported attribute format", attribute_name.c_str()));
544 catch (const cv::Exception& e)
547 if (DNN_DIAGNOSTICS_RUN)
549 CV_LOG_ERROR(NULL, "DNN/ONNX: Potential problem with processing attributes for node " << node_proto.name() << " Attribute " << attribute_name.c_str()
559 Mat ONNXImporter::getBlob(const opencv_onnx::NodeProto& node_proto, int index)
561 CV_Assert(index < node_proto.input_size());
562 const std::string& input_name = node_proto.input(index);
563 return getBlob(input_name);
566 Mat ONNXImporter::getBlob(const std::string& input_name)
568 std::map<std::string, Mat>::const_iterator constBlob = constBlobs.find(input_name);
569 if (constBlob == constBlobs.end())
571 CV_Error(Error::StsBadArg, std::string("Blob ") + input_name + " not found in const blobs");
573 return constBlob->second;
576 void ONNXImporter::addLayer(LayerParams& layerParams,
577 const opencv_onnx::NodeProto& node_proto)
579 int depth = layerParams.get<int>("depth", CV_32F);
580 int id = dstNet.addLayer(layerParams.name, layerParams.type, depth, layerParams);
581 for (int i = 0; i < node_proto.output_size(); ++i)
583 const std::string& output_name = node_proto.output(i);
584 if (!output_name.empty())
586 layer_id.insert(std::make_pair(output_name, LayerInfo(id, i)));
590 std::vector<MatShape> layerInpShapes, layerOutShapes, layerInternalShapes;
592 for (int j = 0; j < node_proto.input_size(); j++)
594 const std::string& input_name = node_proto.input(j);
595 IterLayerId_t layerId = layer_id.find(input_name);
596 if (layerId != layer_id.end()) {
597 dstNet.connect(layerId->second.layerId, layerId->second.outputId, id, inpNum);
599 // Collect input shapes.
600 IterShape_t shapeIt = outShapes.find(input_name);
601 CV_Assert(shapeIt != outShapes.end());
602 layerInpShapes.push_back(shapeIt->second);
605 // Compute shape of output blob for this layer.
606 Ptr<Layer> layer = dstNet.getLayer(id); // FIXIT: avoid instantiation of layers during the import stage
607 layer->getMemoryShapes(layerInpShapes, 0, layerOutShapes, layerInternalShapes);
608 for (int i = 0; i < node_proto.output_size() && i < (int)layerOutShapes.size(); ++i)
610 const std::string& output_name = node_proto.output(i);
611 if (!output_name.empty())
613 outShapes[node_proto.output(i)] = layerOutShapes[i];
618 /** @brief Make N copies of input layer and set them as input to node_proto.
619 * @param prefix prefix of new layers' names
620 * @param node_proto node which will contain all copies as inputs
621 * @param input name of the node to copy
622 * @param n number of copies
624 void ONNXImporter::expandMid(const std::string& prefix, opencv_onnx::NodeProto& node_proto,
625 const std::string& input, size_t n)
627 std::vector<std::string> input_names;
628 input_names.reserve(n);
629 for (size_t j = 0; j < n; j++)
632 copyLP.name = format("%s/copy_%zu", prefix.c_str(), j);
633 copyLP.type = "Identity";
634 CV_Assert((layer_id.find(copyLP.name) == layer_id.end()) &&
635 "Couldn't copy the node: generated name already exists in the graph.");
636 input_names.push_back(copyLP.name);
638 node_proto.set_input(0, input);
639 node_proto.set_output(0, copyLP.name);
640 addLayer(copyLP, node_proto);
642 node_proto.clear_input();
643 for (size_t i = 0; i < input_names.size(); i++)
645 node_proto.add_input(input_names[i]);
649 /** @brief Multiply one of node_proto inputs by -1
650 * @param layerParams parameters of the node
651 * @param node_proto node which input will be replaced
652 * @param input_id id of input to be multiplied by -1
654 void ONNXImporter::addNegation(const LayerParams& layerParams, opencv_onnx::NodeProto& node_proto, int input_id)
656 LayerParams powerParams;
657 powerParams.name = layerParams.name + "/neg";
658 powerParams.type = "Power";
659 powerParams.set("scale", -1.f);
662 int id = dstNet.addLayer(powerParams.name, powerParams.type, powerParams);
664 IterLayerId_t layerId = layer_id.find(node_proto.input(input_id));
665 CV_Assert(layerId != layer_id.end());
666 dstNet.connect(layerId->second.layerId, layerId->second.outputId, id, 0);
668 layer_id.insert(std::make_pair(powerParams.name, LayerInfo(id, 0)));
669 outShapes[powerParams.name] = outShapes[node_proto.input(input_id)];
671 //Replace input to Power
672 node_proto.set_input(input_id, powerParams.name);
675 void ONNXImporter::addConstant(const std::string& name, const Mat& blob)
677 CV_LOG_DEBUG(NULL, "DNN/ONNX: add constant '" << name << "' shape=" << toString(shape(blob)) << ": " << toString(blob));
678 constBlobs.insert(std::make_pair(name, blob));
679 outShapes.insert(std::make_pair(name, shape(blob)));
682 void ONNXImporter::parseOperatorSet()
684 int ir_version = model_proto.has_ir_version() ? static_cast<int>(model_proto.ir_version()) : -1;
688 int opset_size = model_proto.opset_import_size();
691 CV_LOG_INFO(NULL, "DNN/ONNX: missing opset information")
695 for (int i = 0; i < opset_size; ++i)
697 const ::opencv_onnx::OperatorSetIdProto& opset_entry = model_proto.opset_import(i);
698 const std::string& domain = opset_entry.has_domain() ? opset_entry.domain() : std::string();
699 int version = opset_entry.has_version() ? opset_entry.version() : -1;
700 if (domain.empty() || domain == str_domain_ai_onnx)
702 // ONNX opset covered by specification: https://github.com/onnx/onnx/blob/master/docs/Operators.md
703 onnx_opset = std::max(onnx_opset, version);
704 onnx_opset_map[str_domain_ai_onnx] = onnx_opset;
708 CV_LOG_DEBUG(NULL, "DNN/ONNX: using non-standard ONNX opset[" << i << "]: domain='" << domain << "' version=" << version);
709 onnx_opset_map[domain] = onnx_opset;
713 CV_LOG_INFO(NULL, "DNN/ONNX: ONNX opset version = " << onnx_opset);
715 buildDispatchMap_ONNX_AI(onnx_opset);
716 for (const auto& pair : onnx_opset_map)
718 if (pair.first == str_domain_ai_onnx)
720 continue; // done above
722 else if (pair.first == "com.microsoft")
724 buildDispatchMap_COM_MICROSOFT(pair.second);
728 CV_LOG_INFO(NULL, "DNN/ONNX: unknown domain='" << pair.first << "' version=" << pair.second << ". No dispatch map, you may need to register 'custom' layers.");
733 void ONNXImporter::handleQuantizedNode(LayerParams& layerParams,
734 const opencv_onnx::NodeProto& node_proto)
736 // Quantized nodes have output names ending with 'quantized'
737 std::string outName = node_proto.output(0);
738 int len = outName.length();
742 if (outName.substr(len - 9) == "quantized")
744 outName = outName.substr(0, len - 9);
745 Mat scale, zeropoint;
747 if (constBlobs.find(outName + "scale") != constBlobs.end() &&
748 constBlobs.find(outName + "zero_point") != constBlobs.end())
750 scale = getBlob(outName + "scale");
751 zeropoint = getBlob(outName + "zero_point");
755 std::string inpName = node_proto.input(0);
756 inpName = inpName.substr(0, inpName.length() - 9);
757 scale = getBlob(inpName + "scale");
758 zeropoint = getBlob(inpName + "zero_point");
760 for (int i = 0; i < node_proto.output_size(); i++)
762 std::string out = node_proto.output(i);
763 out = out.substr(0, out.length() - 9);
764 addConstant(out + "scale", scale);
765 addConstant(out + "zero_point", zeropoint);
769 if (scale.total() != 1 || zeropoint.total() != 1)
770 CV_Error(Error::StsNotImplemented, "Per-channel scales/zeropoints are not supported");
772 layerParams.set("depth", CV_8S);
773 layerParams.set("scales", DictValue::arrayReal(scale.ptr<float>(), 1));
774 layerParams.set("zeropoints", DictValue::arrayInt(zeropoint.ptr<int8_t>(), 1));
778 void ONNXImporter::populateNet()
780 CV_Assert(model_proto.has_graph());
781 graph_proto = model_proto.graph();
783 std::string framework_version;
784 if (model_proto.has_producer_name())
785 framework_name = model_proto.producer_name();
786 if (model_proto.has_producer_version())
787 framework_version = model_proto.producer_version();
789 CV_LOG_INFO(NULL, "DNN/ONNX: loading ONNX"
790 << (model_proto.has_ir_version() ? cv::format(" v%d", (int)model_proto.ir_version()) : cv::String())
791 << " model produced by '" << framework_name << "'"
792 << (framework_version.empty() ? cv::String() : cv::format(":%s", framework_version.c_str()))
793 << ". Number of nodes = " << graph_proto.node_size()
794 << ", initializers = " << graph_proto.initializer_size()
795 << ", inputs = " << graph_proto.input_size()
796 << ", outputs = " << graph_proto.output_size()
801 simplifySubgraphs(graph_proto);
803 const int layersSize = graph_proto.node_size();
804 CV_LOG_DEBUG(NULL, "DNN/ONNX: graph simplified to " << layersSize << " nodes");
806 constBlobs = getGraphTensors(graph_proto); // scan GraphProto.initializer
807 std::vector<String> netInputs; // map with network inputs (without const blobs)
808 // Add all the inputs shapes. It includes as constant blobs as network's inputs shapes.
809 for (int i = 0; i < graph_proto.input_size(); ++i)
811 const opencv_onnx::ValueInfoProto& valueInfoProto = graph_proto.input(i);
812 CV_Assert(valueInfoProto.has_name());
813 const std::string& name = valueInfoProto.name();
814 CV_Assert(valueInfoProto.has_type());
815 const opencv_onnx::TypeProto& typeProto = valueInfoProto.type();
816 CV_Assert(typeProto.has_tensor_type());
817 const opencv_onnx::TypeProto::Tensor& tensor = typeProto.tensor_type();
818 CV_Assert(tensor.has_shape());
819 const opencv_onnx::TensorShapeProto& tensorShape = tensor.shape();
821 int dim_size = tensorShape.dim_size();
822 CV_CheckGE(dim_size, 0, ""); // some inputs are scalars (dims=0), e.g. in Test_ONNX_nets.Resnet34_kinetics test
823 MatShape inpShape(dim_size);
824 for (int j = 0; j < dim_size; ++j)
826 const opencv_onnx::TensorShapeProto_Dimension& dimension = tensorShape.dim(j);
827 if (dimension.has_dim_param())
829 CV_LOG_DEBUG(NULL, "DNN/ONNX: input[" << i << "] dim[" << j << "] = <" << dimension.dim_param() << "> (dynamic)");
831 // https://github.com/onnx/onnx/blob/master/docs/DimensionDenotation.md#denotation-definition
832 if (dimension.has_denotation())
834 CV_LOG_INFO(NULL, "DNN/ONNX: input[" << i << "] dim[" << j << "] denotation is '" << dimension.denotation() << "'");
836 inpShape[j] = dimension.dim_value();
837 // NHW, NCHW(NHWC), NCDHW(NDHWC); do not set this flag if only N is dynamic
838 if (dimension.has_dim_param() && !(j == 0 && inpShape.size() >= 3))
840 hasDynamicShapes = true;
843 bool isInitialized = ((constBlobs.find(name) != constBlobs.end()));
844 CV_LOG_IF_DEBUG(NULL, !isInitialized, "DNN/ONNX: input[" << i << " as '" << name << "'] shape=" << toString(inpShape));
845 CV_LOG_IF_VERBOSE(NULL, 0, isInitialized, "DNN/ONNX: pre-initialized input[" << i << " as '" << name << "'] shape=" << toString(inpShape));
846 if (dim_size > 0 && !hasDynamicShapes) // FIXIT result is not reliable for models with multiple inputs
848 inpShape[0] = std::max(inpShape[0], 1); // It's OK to have undetermined batch size
850 outShapes[valueInfoProto.name()] = inpShape;
851 // fill map: push layer name, layer id and output id
854 netInputs.push_back(name);
855 layer_id.insert(std::make_pair(name, LayerInfo(0, netInputs.size() - 1)));
859 dstNet.setInputsNames(netInputs);
862 for (int i = 0; i < graph_proto.output_size(); ++i)
864 dumpValueInfoProto(i, graph_proto.output(i), "output");
867 if (DNN_DIAGNOSTICS_RUN) {
868 CV_LOG_INFO(NULL, "DNN/ONNX: start diagnostic run!");
869 layerHandler->fillRegistry(graph_proto);
872 for(int li = 0; li < layersSize; li++)
874 const opencv_onnx::NodeProto& node_proto = graph_proto.node(li);
875 handleNode(node_proto);
879 for (int i = 0; i < graph_proto.output_size(); ++i)
881 const std::string& output_name = graph_proto.output(i).name();
882 if (output_name.empty())
884 CV_LOG_ERROR(NULL, "DNN/ONNX: can't register output without name: " << i);
887 ConstIterLayerId_t layerIt = layer_id.find(output_name);
888 if (layerIt == layer_id.end())
890 CV_LOG_ERROR(NULL, "DNN/ONNX: can't find layer for output name: '" << output_name << "'. Does model imported properly?");
894 const LayerInfo& li = layerIt->second;
895 int outputId = dstNet.registerOutput(output_name, li.layerId, li.outputId); CV_UNUSED(outputId);
896 // no need to duplicate message from engine: CV_LOG_DEBUG(NULL, "DNN/ONNX: registered output='" << output_name << "' with id=" << outputId);
899 CV_LOG_DEBUG(NULL, (DNN_DIAGNOSTICS_RUN ? "DNN/ONNX: diagnostic run completed!" : "DNN/ONNX: import completed!"));
902 std::string ONNXImporter::getLayerTypeDomain(const opencv_onnx::NodeProto& node_proto)
904 if (!node_proto.has_domain())
905 return str_domain_ai_onnx;
906 const std::string& domain = node_proto.domain();
908 return str_domain_ai_onnx;
912 const ONNXImporter::DispatchMap& ONNXImporter::getDispatchMap(const opencv_onnx::NodeProto& node_proto)
914 static DispatchMap empty_map;
915 const std::string& layer_type_domain = getLayerTypeDomain(node_proto);
916 auto it = domain_dispatch_map.find(layer_type_domain);
917 if (it == domain_dispatch_map.end())
925 std::string ONNXImporter::extractNodeName(const opencv_onnx::NodeProto& node_proto)
927 // We need to rework DNN outputs API, this is a workaround for #21698
928 if (node_proto.has_name() && !node_proto.name().empty())
931 return node_proto.name();
932 return cv::format("onnx_node!%s", node_proto.name().c_str());
934 for (int i = 0; i < node_proto.output_size(); ++i)
936 const std::string& name = node_proto.output(i);
937 // There are two ways to leave an optional input or output unspecified:
938 // the first, available only for trailing inputs and outputs, is to simply not provide that input;
939 // the second method is to use an empty string in place of an input or output name.
944 return cv::format("onnx_node_output_%d!%s", i, name.c_str());
947 CV_Error(Error::StsAssert, "Couldn't deduce Node name.");
950 void ONNXImporter::handleNode(const opencv_onnx::NodeProto& node_proto)
952 CV_Assert(node_proto.output_size() >= 1);
953 const std::string& name = extractNodeName(node_proto);
954 const std::string& layer_type = node_proto.op_type();
955 const std::string& layer_type_domain = getLayerTypeDomain(node_proto);
956 const auto& dispatch = getDispatchMap(node_proto);
958 CV_LOG_DEBUG(NULL, "DNN/ONNX: processing node with " << node_proto.input_size() << " inputs and "
959 << node_proto.output_size() << " outputs: "
960 << cv::format("[%s]:(%s)", layer_type.c_str(), name.c_str())
961 << cv::format(" from %sdomain='", onnx_opset_map.count(layer_type_domain) == 1 ? "" : "undeclared ")
962 << layer_type_domain << "'"
965 if (dispatch.empty())
967 CV_LOG_WARNING(NULL, "DNN/ONNX: missing dispatch map for domain='" << layer_type_domain << "'");
970 LayerParams layerParams;
973 // FIXIT not all cases can be repacked into "LayerParams". Importer should handle such cases directly for each "layer_type"
974 layerParams = getLayerParams(node_proto);
976 layerParams.name = name;
977 layerParams.type = layer_type;
978 layerParams.set("has_dynamic_shapes", hasDynamicShapes);
980 handleQuantizedNode(layerParams, node_proto);
982 DispatchMap::const_iterator iter = dispatch.find(layer_type);
983 if (iter != dispatch.end())
985 CALL_MEMBER_FN(*this, iter->second)(layerParams, node_proto);
989 parseCustomLayer(layerParams, node_proto);
992 catch (const cv::Exception& e)
994 if (DNN_DIAGNOSTICS_RUN)
996 CV_LOG_ERROR(NULL, "DNN/ONNX: Potential problem during processing node with " << node_proto.input_size() << " inputs and " << node_proto.output_size() << " outputs: "
997 << cv::format("[%s]:(%s)", layer_type.c_str(), name.c_str())
998 << " from domain='" << layer_type_domain << "'"
1001 cv::AutoLock lock(getLayerFactoryMutex());
1002 auto registeredLayers = getLayerFactoryImpl();
1003 if (registeredLayers.find(layerParams.type) != registeredLayers.end())
1007 Ptr<Layer> layer = LayerFactory::createLayerInstance(layerParams.type, layerParams);
1009 catch (const std::exception& e)
1011 CV_LOG_ERROR(NULL, "DNN/ONNX: Layer of type " << layerParams.type << "(" << layer_type << ") cannot be created with parameters " << layerParams << ". Error: " << e.what()
1018 CV_LOG_ERROR(NULL, "DNN/ONNX: ERROR during processing node with " << node_proto.input_size() << " inputs and " << node_proto.output_size() << " outputs: "
1019 << cv::format("[%s]:(%s)", layer_type.c_str(), name.c_str())
1020 << " from domain='" << layer_type_domain << "'"
1023 for (int i = 0; i < node_proto.input_size(); i++)
1025 CV_LOG_INFO(NULL, " Input[" << i << "] = '" << node_proto.input(i) << "'");
1027 for (int i = 0; i < node_proto.output_size(); i++)
1029 CV_LOG_INFO(NULL, " Output[" << i << "] = '" << node_proto.output(i) << "'");
1031 if (DNN_DIAGNOSTICS_RUN)
1033 for (int i = 0; i < node_proto.output_size(); ++i)
1035 layer_id.insert(std::make_pair(node_proto.output(i), LayerInfo(0, i)));
1036 outShapes[node_proto.output(i)] = outShapes[node_proto.input(0)];
1040 CV_Error(Error::StsError, cv::format("Node [%s@%s]:(%s) parse error: %s", layer_type.c_str(), layer_type_domain.c_str(), name.c_str(), e.what()));
1044 void ONNXImporter::parseArg(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1046 const std::string& layer_type = node_proto.op_type();
1047 layerParams.type = "Arg";
1048 layerParams.set("op", layer_type == "ArgMax" ? "max" : "min");
1049 addLayer(layerParams, node_proto);
1052 void setCeilMode(LayerParams& layerParams)
1054 // auto_pad attribute is deprecated and uses ceil
1055 if (layerParams.has("pad_mode"))
1057 layerParams.set("ceil_mode", true);
1059 else if (!layerParams.has("ceil_mode"))
1061 layerParams.set("ceil_mode", false);
1065 void ONNXImporter::parseMaxUnpool(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1067 layerParams.type = "MaxUnpool";
1069 DictValue kernel_shape = layerParams.get("kernel_size");
1070 CV_Assert(kernel_shape.size() == 2);
1071 layerParams.set("pool_k_w", kernel_shape.get<int>(0));
1072 layerParams.set("pool_k_h", kernel_shape.get<int>(1));
1074 int pool_pad_w = 0, pool_pad_h = 0;
1075 if (layerParams.has("pad"))
1077 DictValue pads = layerParams.get("pad");
1078 CV_CheckEQ(pads.size(), 2, "");
1079 pool_pad_w = pads.get<int>(0);
1080 pool_pad_h = pads.get<int>(1);
1082 layerParams.set("pool_pad_w", pool_pad_w);
1083 layerParams.set("pool_pad_h", pool_pad_h);
1086 int pool_stride_w = 1, pool_stride_h = 1;
1087 if (layerParams.has("stride"))
1089 DictValue strides = layerParams.get("stride");
1090 CV_CheckEQ(strides.size(), 2, "");
1091 pool_stride_w = strides.get<int>(0);
1092 pool_stride_h = strides.get<int>(1);
1094 layerParams.set("pool_stride_w", pool_stride_w);
1095 layerParams.set("pool_stride_h", pool_stride_h);
1097 addLayer(layerParams, node_proto);
1100 void ONNXImporter::parseMaxPool(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1102 int depth = layerParams.get<int>("depth", CV_32F);
1103 layerParams.type = (depth == CV_8S) ? "PoolingInt8" : "Pooling";
1104 layerParams.set("pool", "MAX");
1105 setCeilMode(layerParams);
1106 addLayer(layerParams, node_proto);
1109 void ONNXImporter::parseAveragePool(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1111 layerParams.type = "Pooling";
1112 layerParams.set("pool", "AVE");
1113 setCeilMode(layerParams);
1114 layerParams.set("ave_pool_padded_area", framework_name == "pytorch");
1115 addLayer(layerParams, node_proto);
1118 void ONNXImporter::parseGlobalPool(LayerParams &layerParams, const opencv_onnx::NodeProto &node_proto_)
1120 opencv_onnx::NodeProto node_proto = node_proto_;
1121 const std::string& layer_type = node_proto.op_type();
1122 const std::string output_name = node_proto.output(0);
1124 CV_Assert(node_proto.input_size() == 1);
1125 layerParams.type = "Pooling";
1127 if (layer_type == "GlobalMaxPool")
1129 else if (layer_type == "GlobalAveragePool")
1132 CV_Error(Error::StsNotImplemented, "Unsupported Pooling type of " + layer_type + " operation.");
1134 CV_Assert(!layerParams.has("axes"));
1135 layerParams.set("global_pooling", true);
1136 layerParams.set("pool", pool);
1137 addLayer(layerParams, node_proto);
1140 void ONNXImporter::parseReduce(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
1142 opencv_onnx::NodeProto node_proto = node_proto_;
1143 const std::string& layer_type = node_proto.op_type();
1144 const std::string output_name = node_proto.output(0);
1145 int depth = layerParams.get<int>("depth", CV_32F);
1147 CV_Assert(node_proto.input_size() <= 2);
1150 if (layer_type == "ReduceMax")
1152 else if (layer_type == "ReduceMin")
1154 else if (layer_type == "ReduceSum")
1156 else if (layer_type == "ReduceSumSquare")
1157 reduceType = "SUM_SQUARE";
1158 else if (layer_type == "ReduceProd")
1159 reduceType = "PROD";
1160 else if (layer_type == "ReduceL1")
1162 else if (layer_type == "ReduceL2")
1164 else if (layer_type == "ReduceLogSum")
1165 reduceType = "LOG_SUM";
1166 else if (layer_type == "ReduceLogSumExp")
1167 reduceType = "LOG_SUM_EXP";
1168 else if (layer_type == "ReduceMean")
1171 CV_Error(Error::StsNotImplemented, "Unsupported Pooling type of " + layer_type + " operation.");
1173 // The ReduceInt8 can only support "MAX" and "MIN".
1176 CV_Assert(reduceType == "MAX" || reduceType == "MIN");
1179 layerParams.type = (depth == CV_8S) ? "ReduceInt8" : "Reduce";
1180 layerParams.set("reduce", reduceType);
1181 bool keepdims = layerParams.get<int>("keepdims", 1) == 1;
1183 MatShape inpShape = outShapes[node_proto.input(0)];
1184 std::vector<bool> shouldDelete(inpShape.size(), false);
1186 if (layer_type == "ReduceSum" && node_proto.input_size() == 2)
1188 if (constBlobs.find(node_proto.input(1)) != constBlobs.end())
1190 Mat axesMat = getBlob(node_proto, 1);
1191 int axesNum = axesMat.total();
1192 for (int i = 0; i < axesNum; i++)
1194 int axis = normalize_axis(axesMat.at<int>(i), inpShape.size());
1195 shouldDelete[axis] = true;
1199 // in opset 13, the ReduceSum has two input, it takes axes as input instead of attribute
1200 // details:https://github.com/onnx/onnx/issues/3420#issuecomment-844295687
1201 CV_Error(Error::StsNotImplemented, "Non-constant axis values in ReduceSum are not supported.");
1205 if (layerParams.has("axes"))
1207 DictValue axes = layerParams.get("axes");
1208 for (int i = 0; i < axes.size(); i++)
1210 int axis = normalize_axis(axes.get<int>(i), inpShape.size());
1211 shouldDelete[axis] = true;
1216 for (int i = 0; i < inpShape.size(); i++)
1218 shouldDelete[i] = true;
1223 std::vector<int> targetShape;
1224 for (int i = 0; i < inpShape.size(); ++i)
1226 if (!shouldDelete[i])
1228 targetShape.push_back(inpShape[i]);
1232 targetShape.push_back(1);
1236 if (targetShape.empty())
1237 targetShape.push_back(1);
1239 // Using PermuteLayer to move the deleted axis to the last.
1240 std::vector<int> perm(inpShape.size(), 0);
1241 for (int i = 0; i < inpShape.size(); i++)
1244 bool needPermuet = false;
1245 for (int i = 0; i < inpShape.size(); i++)
1247 if (shouldDelete[i])
1249 // find the first not deleted element.
1250 std::vector<bool>::iterator iter = std::find(shouldDelete.begin() + i, shouldDelete.end(), false);
1252 if (iter != shouldDelete.end())
1254 int index = iter - shouldDelete.begin();
1256 bool temp = shouldDelete[index];
1257 shouldDelete[index] = shouldDelete[i];
1258 shouldDelete[i] = temp;
1260 std::swap(perm[index], perm[i]);
1261 std::swap(inpShape[index], inpShape[i]);
1269 auto inputString= node_proto.input(0);
1272 LayerParams permuteLp;
1273 permuteLp.name = layerParams.name + "/permute";
1274 permuteLp.type = (depth == CV_8S) ? "PermuteInt8" : "Permute";
1275 permuteLp.set("order", DictValue::arrayInt(perm.data(), perm.size()));
1277 opencv_onnx::NodeProto protoPermute;
1278 protoPermute.add_input(inputString);
1279 protoPermute.add_output(permuteLp.name);
1280 addLayer(permuteLp, protoPermute);
1281 inputString = permuteLp.name;
1284 std::vector<int> deletedDims;
1285 for (int axis_i = 0; axis_i < inpShape.size(); ++axis_i)
1287 if (shouldDelete[axis_i])
1289 deletedDims.push_back(inpShape[axis_i]);
1293 layerParams.set("deleted_dims", DictValue::arrayInt(&deletedDims[0], deletedDims.size()));
1294 layerParams.set("target_dims", DictValue::arrayInt(&targetShape[0], targetShape.size()));
1296 node_proto.set_input(0, inputString);
1297 node_proto.set_output(0, output_name);
1299 addLayer(layerParams, node_proto);
1302 void ONNXImporter::parseSlice(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1305 std::vector<int> begin;
1306 std::vector<int> end;
1307 std::vector<int> steps;
1308 int inp_size = node_proto.input_size();
1312 if (layerParams.has("axes")) {
1313 DictValue axes = layerParams.get("axes");
1314 for (int i = 1; i < axes.size(); ++i) {
1315 CV_Assert(axes.get<int>(i - 1) == axes.get<int>(i) - 1);
1317 axis = axes.get<int>(0);
1320 DictValue starts = layerParams.get("starts");
1321 DictValue ends = layerParams.get("ends");
1322 CV_Assert(starts.size() == ends.size());
1325 CV_CheckLE(axis, 1024, "Slice layer can't have more than 1024 axes"); // arbitrary limit
1326 begin.resize(axis, 0);
1327 end.resize(axis, INT_MAX);
1329 for (int i = 0; i < starts.size(); ++i)
1331 begin.push_back(starts.get<int>(i));
1332 end.push_back(ends.get<int>(i));
1334 } else { // inp_size > 1
1335 CV_Assert(inp_size >= 3);
1336 for (int i = 1; i < inp_size; i++) {
1337 CV_Assert(constBlobs.find(node_proto.input(i)) != constBlobs.end());
1339 Mat start_blob = getBlob(node_proto, 1);
1340 Mat end_blob = getBlob(node_proto, 2);
1341 CV_Assert(start_blob.total() == end_blob.total());
1344 Mat axes_blob = getBlob(node_proto, 3);
1345 const int* axes = (int*)axes_blob.data;
1346 for (int i = 1; i < axes_blob.total(); ++i) {
1347 CV_Assert(axes[i - 1] == axes[i] - 1);
1352 const int* starts = start_blob.ptr<int>();
1353 const int* ends = end_blob.ptr<int>();
1355 begin.resize(axis, 0);
1356 end.resize(axis, INT_MAX);
1358 std::copy(starts, starts + start_blob.total(), std::back_inserter(begin));
1359 std::copy(ends, ends + end_blob.total(), std::back_inserter(end));
1361 if (inp_size == 5) {
1362 CV_Assert(constBlobs.find(node_proto.input(4)) != constBlobs.end());
1363 Mat step_blob = getBlob(node_proto, 4);
1364 const int* steps_ptr = step_blob.ptr<int>();
1367 steps.resize(axis, 1);
1369 std::copy(steps_ptr, steps_ptr + step_blob.total(), std::back_inserter(steps));
1371 // Very strange application for Slice op with tensor reversing.
1372 // We just workaround it for 2d constants.
1373 if (constBlobs.find(node_proto.input(0)) != constBlobs.end() &&
1375 start_blob.at<int>(0) == -1 && step_blob.at<int>(0) == -1 &&
1376 end_blob.at<int>(0) == std::numeric_limits<int32_t>::min())
1378 Mat inp = getBlob(node_proto, 0);
1382 flip(inp, flipped, 0);
1383 addConstant(node_proto.output(0), flipped);
1389 layerParams.set("begin", DictValue::arrayInt(&begin[0], begin.size()));
1390 layerParams.set("end", DictValue::arrayInt(&end[0], end.size()));
1391 layerParams.set("axis", axis);
1394 layerParams.set("steps", DictValue::arrayInt(&steps[0], steps.size()));
1396 if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
1398 Mat inp = getBlob(node_proto, 0);
1399 std::vector<Mat> inputs, sliced;
1400 inputs.push_back(inp);
1401 runLayer(layerParams, inputs, sliced);
1402 CV_Assert(sliced.size() == 1);
1403 addConstant(node_proto.output(0), sliced[0]);
1406 addLayer(layerParams, node_proto);
1409 void ONNXImporter::parseSplit(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1411 if (layerParams.has("split"))
1413 DictValue splits = layerParams.get("split");
1414 const int numSplits = splits.size();
1415 CV_Assert(numSplits > 1);
1417 std::vector<int> slicePoints(numSplits - 1, splits.get<int>(0));
1418 for (int i = 1; i < splits.size() - 1; ++i)
1420 slicePoints[i] = slicePoints[i - 1] + splits.get<int>(i);
1422 layerParams.set("slice_point", DictValue::arrayInt(&slicePoints[0], slicePoints.size()));
1426 layerParams.set("num_split", node_proto.output_size());
1428 int depth = layerParams.get<int>("depth", CV_32F);
1429 layerParams.type = (depth == CV_8S) ? "SliceInt8" : "Slice";
1430 layerParams.set("axis", layerParams.get<float>("axis", 0));
1431 addLayer(layerParams, node_proto);
1434 void ONNXImporter::parseBias(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
1436 opencv_onnx::NodeProto node_proto = node_proto_;
1437 const std::string& layer_type = node_proto.op_type();
1438 bool isSub = layer_type == "Sub";
1440 if (layer_type == "Sum" && node_proto.input_size() == 1)
1442 layerParams.type = "Identity";
1443 addLayer(layerParams, node_proto);
1447 CV_Assert((node_proto.input_size() == 2) || (layer_type == "Sum" && node_proto.input_size() > 2));
1449 if (layer_type == "Sum" && node_proto.input_size() > 2)
1451 for (int i = 0; i < node_proto.input_size(); ++i)
1453 if (layer_id.find(node_proto.input(i)) == layer_id.end())
1455 CV_Error(Error::StsNotImplemented, "Sum of constants is not implemented for inputs > 2");
1460 bool is_const_0 = layer_id.find(node_proto.input(0)) == layer_id.end();
1461 bool is_const_1 = layer_id.find(node_proto.input(1)) == layer_id.end();
1462 if (is_const_0 && is_const_1)
1464 Mat blob_0 = getBlob(node_proto, 0);
1465 Mat blob_1 = getBlob(node_proto, 1);
1466 CV_Assert(blob_0.size == blob_1.size);
1467 Mat output = isSub ? (blob_0 - blob_1) : (blob_0 + blob_1);
1468 addConstant(node_proto.output(0), output);
1471 else if (is_const_0 || is_const_1)
1473 int const_blob_id = is_const_0 ? 0 : 1;
1474 int input_id = 1 - const_blob_id;
1475 Mat blob = getBlob(node_proto, const_blob_id);
1476 int blob_total = blob.total();
1478 const float inputScale = isSub && is_const_0 ? -1.f : 1.f;
1479 const float constScale = isSub && is_const_1 ? -1.f : 1.f;
1481 if (blob_total == 1) {
1482 layerParams.type = "Power";
1483 layerParams.set("scale", inputScale);
1484 layerParams.set("shift", constScale * blob.ptr<float>()[0]);
1487 MatShape inpShape = outShapes[node_proto.input(input_id)];
1488 if (shape(blob) == inpShape)
1490 LayerParams constParams;
1491 constParams.name = layerParams.name + "/const";
1492 constParams.type = "Const";
1493 constParams.blobs.push_back(blob);
1494 int id = dstNet.addLayer(constParams.name, constParams.type, constParams);
1495 layer_id.insert(std::make_pair(constParams.name, LayerInfo(id, 0)));
1496 outShapes[constParams.name] = shape(blob);
1498 layerParams.type = "Eltwise";
1499 float coeffs[] = {1., isSub ? -1.f : 1.f};
1500 layerParams.set("coeff", DictValue::arrayReal<float*>(coeffs, 2));
1501 node_proto.set_input(const_blob_id, constParams.name);
1505 if (inputScale < 0.f)
1507 addNegation(layerParams, node_proto, input_id);
1510 layerParams.type = "Scale";
1511 layerParams.set("bias_term", true);
1513 for (int i = 0; i < graph_proto.initializer_size(); i++)
1515 opencv_onnx::TensorProto tensor_proto = graph_proto.initializer(i);
1516 if (tensor_proto.name() == node_proto.input(const_blob_id))
1518 axis = inpShape.size() - tensor_proto.dims_size();
1522 layerParams.set("axis", axis);
1523 blob = blob.reshape(1, 1);
1524 layerParams.blobs.push_back(constScale * blob);
1528 else if (outShapes[node_proto.input(0)] == outShapes[node_proto.input(1)])
1530 layerParams.type = "Eltwise";
1533 static float subCoeffs[] = {1.f, -1.f};
1534 layerParams.set("coeff", DictValue::arrayReal<float*>(subCoeffs, 2));
1541 addNegation(layerParams, node_proto, 1);
1543 layerParams.type = "Scale";
1544 layerParams.set("bias_term", true);
1546 addLayer(layerParams, node_proto);
1549 void ONNXImporter::parsePow(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1551 if (layer_id.find(node_proto.input(1)) != layer_id.end())
1552 CV_Error(Error::StsNotImplemented, "Unsupported Pow op with variable power");
1554 Mat blob = getBlob(node_proto, 1);
1555 if (blob.total() != 1)
1556 CV_Error(Error::StsNotImplemented, "Pow op supports only scalar power");
1558 blob.convertTo(blob, CV_32F);
1559 layerParams.type = "Power";
1560 layerParams.set("power", blob.ptr<float>()[0]);
1561 addLayer(layerParams, node_proto);
1565 void ONNXImporter::parseMinMax(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1567 const std::string& layer_type = node_proto.op_type();
1568 layerParams.type = "Eltwise";
1569 layerParams.set("operation", layer_type == "Max" ? "max" : "min");
1570 addLayer(layerParams, node_proto);
1573 void ONNXImporter::parseNeg(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1575 layerParams.type = "Power";
1576 layerParams.set("scale", -1);
1577 addLayer(layerParams, node_proto);
1580 void ONNXImporter::parseConstant(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1582 CV_Assert(node_proto.input_size() == 0);
1583 CV_Assert(layerParams.blobs.size() == 1);
1584 addConstant(node_proto.output(0), layerParams.blobs[0]);
1587 void transformBlobs(std::vector<Mat>& blobs)
1592 std::vector<Mat> cudaWorkaround;
1593 cudaWorkaround.push_back(Wx.clone());
1594 cudaWorkaround.push_back(Wh.clone());
1595 cudaWorkaround.push_back(b.clone());
1597 const int numHidden = Wh.size[2];
1600 h0 = h0.reshape(1, h0.size[0] * h0.size[1]);
1602 c0 = c0.reshape(1, c0.size[0] * c0.size[1]);
1604 b = b.reshape(1, b.size[0]);
1605 Mat bx = b.colRange(0, b.cols / 2);
1606 Mat bh = b.colRange(b.cols / 2, b.cols);
1609 auto toIFOC = [] (Mat& in) {
1610 int first = in.size[0];
1611 int rest = in.total() / first / 4;
1612 // every weight blob contains weights for Input, Output, Forget and Cell gates
1613 Mat m = in.reshape(1, {first, 4, rest});
1614 Mat outputGate = m.col(1);
1615 Mat forgetGate = m.col(2);
1616 std::swap_ranges(outputGate.begin<float>(), outputGate.end<float>(), forgetGate.begin<float>());
1623 Wx = Wx.reshape(1, Wx.size[0] * Wx.size[1]);
1624 Wh = Wh.reshape(1, Wh.size[0] * Wh.size[1]);
1628 blobs[2] = b.reshape(1, 1);
1632 if (blobs.size() == 5) {
1633 // so that future patch removing copies can leave all indexing as is
1634 blobs.insert(blobs.begin(), cudaWorkaround.begin(), cudaWorkaround.end());
1639 blobs[5] = P.colRange(0, numHidden);
1640 blobs[5] = blobs[5].clone().reshape(1, blobs[5].total()); // Single column.
1641 blobs[5] = Mat::diag(blobs[5]);
1643 blobs.push_back(P.colRange(numHidden, 2 * numHidden));
1644 blobs[6] = blobs[6].clone().reshape(1, blobs[6].total()); // Single column.
1645 blobs[6] = Mat::diag(blobs[6]);
1647 blobs.push_back(P.colRange(2 * numHidden, 3 * numHidden));
1648 blobs[7] = blobs[7].clone().reshape(1, blobs[7].total()); // Single column.
1649 blobs[7] = Mat::diag(blobs[7]);
1651 // so that future patch removing copies can leave all indexing as is
1652 blobs.insert(blobs.begin(), cudaWorkaround.begin(), cudaWorkaround.end());
1655 void ONNXImporter::lstm_extractConsts(LayerParams& layerParams, const opencv_onnx::NodeProto& lstm_proto, size_t idx, int* blobShape_, int size)
1657 MatShape blobShape(blobShape_, blobShape_ + size);
1659 if (idx < lstm_proto.input_size() && !lstm_proto.input(idx).empty())
1661 blob = getBlob(lstm_proto, idx);
1662 CV_Assert(shape(blob) == blobShape);
1666 blob = Mat(blobShape, CV_32FC1, 0.);
1668 layerParams.blobs.push_back(blob);
1671 void ONNXImporter::lstm_add_reshape(const std::string& input_name, const std::string& output_name, int* layerShape, size_t n)
1673 LayerParams reshapeLp;
1674 reshapeLp.name = cv::format("%s/reshape", input_name.c_str());
1675 reshapeLp.type = "Reshape";
1676 CV_Assert(layer_id.find(reshapeLp.name) == layer_id.end());
1678 reshapeLp.set("dim", DictValue::arrayInt(layerShape, n));
1680 opencv_onnx::NodeProto reshape_proto;
1681 reshape_proto.add_input(input_name);
1682 reshape_proto.add_output(output_name);
1683 addLayer(reshapeLp, reshape_proto);
1686 std::string ONNXImporter::lstm_add_slice(int index, const std::string& input_name, int* begin, int* end, size_t n)
1688 LayerParams sliceLP;
1689 sliceLP.name = cv::format("%s/slice_%d", input_name.c_str(), index);
1690 sliceLP.type = "Slice";
1691 CV_Assert(layer_id.find(sliceLP.name) == layer_id.end());
1693 sliceLP.set("begin", DictValue::arrayInt(begin, n));
1694 sliceLP.set("end", DictValue::arrayInt(end, n));
1695 sliceLP.set("axis", 0);
1697 opencv_onnx::NodeProto slice_proto;
1698 slice_proto.add_input(input_name);
1699 slice_proto.add_output(sliceLP.name);
1700 addLayer(sliceLP, slice_proto);
1702 return slice_proto.output(0);
1705 std::string ONNXImporter::lstm_fix_dims(LayerParams& layerParams, const opencv_onnx::NodeProto& lstm_proto,
1706 int batch_size, int num_directions, int hidden_size, bool need_y, const std::string& y_name,
1709 std::string reshape_output = cv::format("%s/reshape_%d", layerParams.name.c_str(), index);
1711 // reshape from Seq, Batch, Dirs*Hidden to Seq, Batch, Dirs, Hidden
1712 // to not confuse reshape with dynamic first dimension, zero means 'leave unchanged'
1713 int layerShape[] = {0, batch_size, num_directions, hidden_size};
1714 lstm_add_reshape(lstm_proto.output(index), reshape_output, layerShape, sizeof(layerShape) / sizeof(layerShape[0]));
1716 // permute from Seq, Batch, Dirs, Hidden to Seq, Dirs, Batch, Hidden
1717 LayerParams permuteLP;
1718 permuteLP.name = reshape_output + "/permute";
1719 permuteLP.type = "Permute";
1720 CV_Assert(layer_id.find(permuteLP.name) == layer_id.end());
1722 int order[] = {0, 2, 1, 3};
1723 permuteLP.set("order", DictValue::arrayInt(order, 4));
1725 opencv_onnx::NodeProto permute_proto;
1726 permute_proto.add_input(reshape_output);
1727 permute_proto.add_output((need_y && index == 0) ? y_name : static_cast<std::string>(permuteLP.name));
1728 addLayer(permuteLP, permute_proto);
1730 return permute_proto.output(0);
1733 void ONNXImporter::lstm_add_transform(int num_directions, int batch_size, int hidden_size,
1734 int index, const std::string& input_name, const std::string& output_name)
1736 if (num_directions == 1)
1738 // Slice: Yh = Y[-1, :, :, :]
1739 int begin[] = {-1}, end[] = {INT_MAX};
1740 std::string slice_output = lstm_add_slice(index, input_name, begin, end, sizeof(begin) / sizeof(begin[0]));
1742 // Reshape: 1x1xBxH -> 1xBxH
1743 int layerShape[] = {1, batch_size, hidden_size};
1744 lstm_add_reshape(slice_output, output_name, layerShape, sizeof(layerShape) / sizeof(layerShape[0]));
1748 // Slice: SxDxBxH -> last sequence, first direction
1749 int begin0[] = {-1, 0}, end0[] = {INT_MAX, 1};
1750 std::string slice_0 = lstm_add_slice(0, input_name, begin0, end0, sizeof(begin0) / sizeof(begin0[0]));
1752 // Slice: SxDxBxH -> first sequence, last direction
1753 int begin1[] = {0, -1}, end1[] = {1, INT_MAX};
1754 std::string slice_1 = lstm_add_slice(1, input_name, begin1, end1, sizeof(begin1) / sizeof(begin1[0]));
1756 LayerParams concatLP;
1757 concatLP.name = cv::format("%s/concat", input_name.c_str());
1758 concatLP.type = "Concat";
1759 CV_Assert(layer_id.find(concatLP.name) == layer_id.end());
1761 concatLP.set("axis", 1); // 1x1xBxH -> 1x2xBxH
1763 opencv_onnx::NodeProto concat_proto;
1764 concat_proto.add_input(slice_0);
1765 concat_proto.add_input(slice_1);
1766 concat_proto.add_output(concatLP.name);
1767 addLayer(concatLP, concat_proto);
1769 // Reshape: 1x2xBxH -> 2xBxH
1770 int layerShape[] = {2, batch_size, hidden_size};
1771 lstm_add_reshape(concat_proto.output(0), output_name, layerShape, sizeof(layerShape) / sizeof(layerShape[0]));
1775 void ONNXImporter::parseLSTM(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
1777 opencv_onnx::NodeProto lstm_proto = node_proto_;
1778 layerParams.name += "/lstm";
1780 // https://github.com/onnx/onnx/blob/main/docs/Operators.md#LSTM
1781 CV_Assert(lstm_proto.input_size() >= 3);
1782 for (size_t i = 1; i < 3; ++i)
1784 const std::string& name = lstm_proto.input(i);
1785 CV_Assert(!name.empty() && constBlobs.count(name) == 1);
1788 IterShape_t shapeIt = outShapes.find(lstm_proto.input(0));
1789 CV_Assert(shapeIt != outShapes.end());
1790 const MatShape x_shape = shapeIt->second;
1792 const int seq_length = x_shape[0];
1793 const int batch_size = x_shape[1];
1794 const int input_size = x_shape[2];
1795 const int hidden_size = layerParams.get<int>("hidden_size");
1796 const int num_directions = constBlobs[lstm_proto.input(1)].size[0];
1798 int w_size[] = {num_directions, 4*hidden_size, input_size};
1799 lstm_extractConsts(layerParams, lstm_proto, 1, w_size, sizeof(w_size) / sizeof(w_size[0])); // W
1801 int r_size[] = {num_directions, 4*hidden_size, hidden_size};
1802 lstm_extractConsts(layerParams, lstm_proto, 2, r_size, sizeof(r_size) / sizeof(r_size[0])); // R
1804 int b_size[] = {num_directions, 8*hidden_size};
1805 lstm_extractConsts(layerParams, lstm_proto, 3, b_size, sizeof(b_size) / sizeof(b_size[0])); // B
1807 if (4 < lstm_proto.input_size() && !lstm_proto.input(4).empty())
1809 Mat blob = getBlob(lstm_proto, 4);
1810 CV_Assert(blob.total() == batch_size);
1811 for (MatIterator_<int32_t> it = blob.begin<int32_t>(); it != blob.end<int32_t>(); ++it)
1813 CV_Assert(*it == seq_length);
1817 int h_size[] = {num_directions, batch_size, hidden_size};
1818 lstm_extractConsts(layerParams, lstm_proto, 5, h_size, sizeof(h_size) / sizeof(h_size[0])); // initial_h
1820 int c_size[] = {num_directions, batch_size, hidden_size};
1821 lstm_extractConsts(layerParams, lstm_proto, 6, c_size, sizeof(c_size) / sizeof(c_size[0])); // initial_c
1823 if (lstm_proto.input_size() > 7 && !lstm_proto.input(7).empty())
1825 layerParams.set("use_peephole", true);
1826 int p_size[] = {num_directions, 3 * hidden_size};
1827 lstm_extractConsts(layerParams, lstm_proto, 7, p_size, sizeof(p_size) / sizeof(p_size[0])); // P
1830 transformBlobs(layerParams.blobs);
1832 layerParams.set("is_onnx", true);
1833 layerParams.set("reverse", layerParams.get<String>("direction", "") == "reverse");
1834 layerParams.set("bidirectional", layerParams.get<String>("direction", "") == "bidirectional");
1836 bool need_yc = lstm_proto.output_size() > 2 && !lstm_proto.output(2).empty();
1837 bool need_yh = lstm_proto.output_size() > 1 && !lstm_proto.output(1).empty();
1838 bool need_y = lstm_proto.output_size() > 0 && !lstm_proto.output(0).empty();
1840 const std::string y_name = need_y ? lstm_proto.output(0) : "";
1841 const std::string yh_name = need_yh ? lstm_proto.output(1) : "";
1842 const std::string yc_name = need_yc ? lstm_proto.output(2) : "";
1844 layerParams.set("produce_cell_output", need_yc);
1846 lstm_proto.clear_output();
1847 if (need_y || need_yh)
1849 // give random names to LSTMLayer's outputs because every output needs postprocessing
1850 lstm_proto.add_output(cv::format("%s_y", layerParams.name.c_str()));
1854 lstm_proto.add_output(yc_name);
1857 addLayer(layerParams, lstm_proto);
1859 std::string y_output = lstm_fix_dims(layerParams, lstm_proto, batch_size, num_directions, hidden_size, need_y,
1863 lstm_add_transform(num_directions, batch_size, hidden_size, 0, y_output, yh_name);
1867 void ONNXImporter::parseGRU(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
1869 opencv_onnx::NodeProto node_proto = node_proto_;
1870 const std::string output_name = node_proto.output(0);
1871 LayerParams gruParams = layerParams;
1872 gruParams.name += "/gru";
1874 // https://pytorch.org/docs/stable/generated/torch.nn.GRU.html?highlight=gru#
1875 CV_Assert(node_proto.input_size() == 6);
1876 Mat Wx = getBlob(node_proto, 1);
1877 Mat Wh = getBlob(node_proto, 2);
1878 Mat b = getBlob(node_proto, 3);
1879 Mat h0 = getBlob(node_proto, 5);
1881 Wx = Wx.reshape(1, Wx.size[0] * Wx.size[1]);
1882 Wh = Wh.reshape(1, Wh.size[0] * Wh.size[1]);
1883 h0 = h0.reshape(1, h0.size[0] * h0.size[1]);
1884 b = b.reshape(1, b.size[0]);
1886 gruParams.blobs.resize(4);
1887 gruParams.blobs[0] = Wh;
1888 gruParams.blobs[1] = Wx;
1889 gruParams.blobs[2] = b;
1890 gruParams.blobs[3] = h0;
1891 gruParams.set("bidirectional", gruParams.get<String>("direction", "") == "bidirectional");
1893 node_proto.set_output(0, gruParams.name); // set different name so output shapes will be registered on that name
1894 addLayer(gruParams, node_proto);
1896 MatShape gruShape = outShapes[node_proto.output(0)];
1898 // Add fake 1 as it is done in ONNX
1899 gruShape.insert(gruShape.begin() + 1, 1);
1901 layerParams.type = "Reshape";
1902 layerParams.set("dim", DictValue::arrayInt(&gruShape[0], gruShape.size()));
1903 node_proto.set_input(0, gruParams.name); // redirect input to GRU
1904 node_proto.set_output(0, output_name); // keep origin GRU's name
1905 addLayer(layerParams, node_proto);
1908 void ONNXImporter::parseImageScaler(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1910 const float scale = layerParams.has("scale") ? layerParams.get<float>("scale") : 1.0f;
1911 layerParams.erase("scale");
1913 if (layerParams.has("bias"))
1915 layerParams.type = "Scale";
1916 layerParams.blobs.push_back(
1917 Mat(Size(1, layerParams.get("bias").size()), CV_32FC1, scale));
1919 layerParams.set("bias_term", true);
1920 Mat bias(1, layerParams.get("bias").size(), CV_32FC1);
1921 for (int j = 0; j < bias.total(); j++) {
1922 bias.at<float>(0, j) = layerParams.get("bias").getRealValue(j);
1924 layerParams.blobs.push_back(bias);
1925 layerParams.erase("bias");
1928 layerParams.set("scale", scale);
1929 layerParams.type = "Power";
1931 addLayer(layerParams, node_proto);
1934 void ONNXImporter::parseClip(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1936 layerParams.type = "ReLU6";
1937 float min_value = -FLT_MAX, max_value = FLT_MAX;
1938 int input_size = node_proto.input_size();
1939 CV_Check(input_size, 1 <= input_size && input_size <= 3, "");
1941 if (input_size >= 2 && !node_proto.input(1).empty())
1943 if (constBlobs.find(node_proto.input(1)) != constBlobs.end())
1944 min_value = getBlob(node_proto, 1).at<float>(0);
1946 CV_Error(Error::StsNotImplemented, "Non-constant min values in Clip are not supported");
1949 if (input_size == 3 && !node_proto.input(2).empty())
1951 if (constBlobs.find(node_proto.input(2)) != constBlobs.end())
1952 max_value = getBlob(node_proto, 2).at<float>(0);
1954 CV_Error(Error::StsNotImplemented, "Non-constant max values in Clip are not supported");
1957 layerParams.set("min_value", layerParams.get<float>("min", min_value));
1958 layerParams.set("max_value", layerParams.get<float>("max", max_value));
1959 addLayer(layerParams, node_proto);
1962 void ONNXImporter::parseLeakyRelu(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1964 layerParams.type = "ReLU";
1965 layerParams.set("negative_slope", layerParams.get<float>("alpha", 0.01));
1966 addLayer(layerParams, node_proto);
1969 void ONNXImporter::parseRelu(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1971 layerParams.type = "ReLU";
1972 addLayer(layerParams, node_proto);
1975 void ONNXImporter::parseElu(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1977 layerParams.type = "ELU";
1978 addLayer(layerParams, node_proto);
1981 void ONNXImporter::parseTanh(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1983 layerParams.type = "TanH";
1984 addLayer(layerParams, node_proto);
1987 void ONNXImporter::parseAbs(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1989 layerParams.type = "AbsVal";
1990 addLayer(layerParams, node_proto);
1993 void ONNXImporter::parseCompare(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1995 CV_Assert(node_proto.input_size() == 2);
1996 const std::string& layer_type = node_proto.op_type();
1998 bool is_const_0 = layer_id.find(node_proto.input(0)) == layer_id.end();
1999 bool is_const_1 = layer_id.find(node_proto.input(1)) == layer_id.end();
2001 if (is_const_0 || is_const_1)
2003 Mat blob = getBlob(node_proto, static_cast<int>(is_const_1));
2004 blob = blob.reshape(1, 1);
2005 layerParams.blobs.push_back(blob);
2008 layerParams.type = "Compare";
2010 if (layer_type == "Equal")
2011 layerParams.set("mode", "equal");
2012 else if (layer_type == "Greater")
2013 layerParams.set("mode", "greater");
2015 layerParams.set("mode", "less");
2016 addLayer(layerParams, node_proto);
2019 void ONNXImporter::parsePRelu(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2021 layerParams.type = "PReLU";
2022 layerParams.blobs.push_back(getBlob(node_proto, 1));
2023 addLayer(layerParams, node_proto);
2026 void ONNXImporter::parseLRN(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2028 replaceLayerParam(layerParams, "size", "local_size");
2029 addLayer(layerParams, node_proto);
2032 void ONNXImporter::parseInstanceNormalization(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
2034 opencv_onnx::NodeProto node_proto = node_proto_;
2035 if (node_proto.input_size() != 3)
2036 CV_Error(Error::StsNotImplemented,
2037 "Expected input, scale, bias");
2039 layerParams.blobs.resize(4);
2040 layerParams.blobs[2] = getBlob(node_proto, 1); // weightData
2041 layerParams.blobs[3] = getBlob(node_proto, 2); // biasData
2042 layerParams.set("has_bias", true);
2043 layerParams.set("has_weight", true);
2045 // Get number of channels in input
2046 int size = layerParams.blobs[2].total();
2047 layerParams.blobs[0] = Mat::zeros(size, 1, CV_32F); // mean
2048 layerParams.blobs[1] = Mat::ones(size, 1, CV_32F); // std
2050 LayerParams mvnParams;
2051 mvnParams.name = layerParams.name + "/MVN";
2052 mvnParams.type = "MVN";
2053 mvnParams.set("eps", layerParams.get<float>("epsilon"));
2054 layerParams.erase("epsilon");
2057 int id = dstNet.addLayer(mvnParams.name, mvnParams.type, mvnParams);
2059 IterLayerId_t layerId = layer_id.find(node_proto.input(0));
2060 CV_Assert(layerId != layer_id.end());
2061 dstNet.connect(layerId->second.layerId, layerId->second.outputId, id, 0);
2063 layer_id.insert(std::make_pair(mvnParams.name, LayerInfo(id, 0)));
2064 outShapes[mvnParams.name] = outShapes[node_proto.input(0)];
2066 //Replace Batch Norm's input to MVN
2067 node_proto.set_input(0, mvnParams.name);
2068 layerParams.type = "BatchNorm";
2069 addLayer(layerParams, node_proto);
2072 void ONNXImporter::parseBatchNormalization(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2074 if (node_proto.input_size() != 5)
2075 CV_Error(Error::StsNotImplemented,
2076 "Expected input, scale, bias, mean and var");
2078 layerParams.type = "BatchNorm";
2079 replaceLayerParam(layerParams, "epsilon", "eps");
2080 replaceLayerParam(layerParams, "spatial", "use_global_stats");
2082 Mat meanData = getBlob(node_proto, 3);
2083 Mat stdData = getBlob(node_proto, 4);
2085 layerParams.blobs.push_back(meanData);
2086 layerParams.blobs.push_back(stdData);
2088 if (!node_proto.input(1).empty()) {
2089 layerParams.set("has_weight", true);
2090 layerParams.blobs.push_back(getBlob(node_proto, 1)); // weightData
2092 layerParams.set("has_weight", false);
2095 if (!node_proto.input(2).empty()) {
2096 layerParams.set("has_bias", true);
2097 layerParams.blobs.push_back(getBlob(node_proto, 2)); // biasData
2099 layerParams.set("has_bias", false);
2101 addLayer(layerParams, node_proto);
2104 // A * B + C = Y, we require that the dimension of A is [m, k], and the dimension of B is [n, k].
2105 // And the dim of output Y is [m, n]
2106 void ONNXImporter::parseGemm(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2108 CV_Assert(node_proto.input_size() >= 2);
2109 layerParams.type = "InnerProduct";
2110 Mat weights = getBlob(node_proto, 1);
2112 if (!layerParams.get<int>("transB", 0))
2114 transpose(weights, weights);
2116 layerParams.blobs.push_back(weights);
2118 if (node_proto.input_size() == 3) {
2119 Mat bias = getBlob(node_proto, 2);
2120 layerParams.blobs.push_back(bias);
2122 if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
2124 Mat inputBuf = getBlob(node_proto, 0);
2126 LayerParams constParams;
2127 constParams.name = node_proto.input(0);
2128 constParams.type = "Const";
2129 constParams.blobs.push_back(inputBuf);
2131 opencv_onnx::NodeProto proto;
2132 proto.add_output(constParams.name);
2133 addLayer(constParams, proto);
2136 layerParams.set("num_output", layerParams.blobs[0].size[0]);
2137 layerParams.set("bias_term", node_proto.input_size() == 3);
2138 addLayer(layerParams, node_proto);
2141 void ONNXImporter::parseMatMul(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2143 CV_Assert(node_proto.input_size() == 2);
2144 layerParams.type = "InnerProduct";
2145 layerParams.set("bias_term", false);
2146 CV_Assert(constBlobs.find(node_proto.input(0)) == constBlobs.end());
2147 int firstInpDims = outShapes[node_proto.input(0)].size();
2150 if (constBlobs.find(node_proto.input(1)) != constBlobs.end())
2152 Mat blob = getBlob(node_proto, 1);
2153 secondInpDims = blob.dims;
2154 layerParams.blobs.push_back(blob.t());
2155 layerParams.set("num_output", layerParams.blobs[0].size[0]);
2157 secondInpDims = outShapes[node_proto.input(1)].size();
2159 layerParams.set("axis", firstInpDims - secondInpDims + 1);
2160 addLayer(layerParams, node_proto);
2163 void findBroadAxis(const MatShape& broadShape, const MatShape& outShape, size_t& axis, int& broadAxis)
2165 const size_t diff = outShape.size() - broadShape.size();
2167 // find the first non-one element of the broadcasting shape
2169 for (; axis < broadShape.size() && broadShape[axis] == 1; ++axis) {}
2171 // find the last non-one element of the broadcasting shape
2172 size_t endAxis = broadShape.size();
2173 for (; endAxis > axis && broadShape[endAxis - 1] == 1; --endAxis) {}
2175 // find one between axis and endAxis - as it needs to be broadcasted,
2176 // dimensions from the left of axis and from the right of endAxis will be handled by Scale layer
2178 for (size_t i = axis; i < endAxis; ++i)
2180 size_t outAxis = i + diff;
2181 if (outShape[outAxis] == broadShape[i])
2186 // ensure we need to broadcast only 1 dimension in the middle
2187 CV_Assert(broadShape[i] == 1 && broadAxis == -1);
2188 broadAxis = static_cast<int>(outAxis);
2195 void ONNXImporter::parseMul(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
2197 opencv_onnx::NodeProto node_proto = node_proto_;
2198 const std::string& layer_type = node_proto.op_type();
2199 const std::string output_name = node_proto.output(0);
2200 CV_Assert(node_proto.input_size() == 2);
2202 bool isDiv = layer_type == "Div";
2204 bool haveVariables = false;
2205 for (int i = 0; i < 2; ++i)
2207 if (constBlobs.find(node_proto.input(i)) != constBlobs.end())
2210 haveVariables = true;
2212 if (constId != -1 && haveVariables)
2214 Mat blob = getBlob(node_proto, constId);
2215 blob = blob.reshape(1, 1);
2216 if (blob.total() == 1) {
2217 float blob_value = blob.ptr<float>()[0];
2218 float coeff = blob_value;
2221 coeff = 1.f / blob_value;
2224 // Power layer calculates (x*scale + shift)^power, so const/x -> (x * (1/const) + 0)^(-1)
2225 layerParams.set("power", -1.f);
2228 layerParams.set("scale", coeff);
2229 layerParams.type = "Power";
2233 divide(1.0, blob, blob);
2234 layerParams.blobs.push_back(blob);
2235 layerParams.type = "Scale";
2238 else if (!haveVariables)
2240 Mat inp0 = getBlob(node_proto, 0);
2241 Mat inp1 = getBlob(node_proto, 1);
2243 if (inp0.size != inp1.size && (inp0.total() != 1 || inp1.total() != 1))
2244 CV_Error_(Error::StsNotImplemented, ("Different shapes case is not supported with constant inputs: %s", layer_type.c_str()));
2246 if (inp0.total() == 1 && inp1.total() == 1 && inp0.dims != inp1.dims)
2248 if (inp0.dims < inp1.dims)
2250 inp0 = inp0.reshape(1, inp1.dims, inp1.size);
2251 inp0.dims = inp1.dims;
2255 inp1 = inp1.reshape(1, inp0.dims, inp0.size);
2256 inp1.dims = inp0.dims;
2261 if (inp0.total() != inp1.total())
2263 if (inp0.total() == 1)
2265 float inp0_value = inp0.ptr<float>()[0];
2266 float coeff = isDiv ? 1.0 / inp0_value : inp0_value;
2267 multiply(inp1, coeff, out);
2271 float inp1_value = inp1.ptr<float>()[0];
2272 float coeff = isDiv ? 1.0 / inp1_value : inp1_value;
2273 multiply(inp0, coeff, out);
2279 out = isDiv ? inp0 / inp1 : inp0.mul(inp1);
2282 if (inp0.dims == 1 && inp1.dims == 1)
2283 out.dims = 1; // to workaround dims == 1
2284 addConstant(output_name, out);
2287 else if (outShapes[node_proto.input(0)] == outShapes[node_proto.input(1)])
2289 layerParams.type = "Eltwise";
2290 layerParams.set("operation", isDiv ? "div" : "prod");
2294 // Scale layer allocate output with the first input shape
2295 if (total(outShapes[node_proto.input(0)]) < total(outShapes[node_proto.input(1)]))
2297 opencv_onnx::NodeProto proto;
2298 proto.add_input(node_proto.input(1));
2299 proto.add_input(node_proto.input(0));
2300 proto.add_output(output_name);
2306 LayerParams powerParams;
2307 powerParams.name = layerParams.name + "/inv";
2308 powerParams.type = "Power";
2309 powerParams.set("power", -1);
2311 //Create Power layer
2312 int id = dstNet.addLayer(powerParams.name, powerParams.type, powerParams);
2314 IterLayerId_t layerId = layer_id.find(node_proto.input(1));
2315 CV_Assert(layerId != layer_id.end());
2316 dstNet.connect(layerId->second.layerId, layerId->second.outputId, id, 0);
2318 layer_id.insert(std::make_pair(powerParams.name, LayerInfo(id, 0)));
2319 outShapes[powerParams.name] = outShapes[node_proto.input(1)];
2321 //Replace input to Power
2322 node_proto.set_input(1, powerParams.name);
2325 const MatShape& broadShape = outShapes[node_proto.input(1)];
2326 const MatShape& outShape = outShapes[node_proto.input(0)];
2330 findBroadAxis(broadShape, outShape, axis, broadAxis);
2332 // if there is a one dimension in the middle that should be broadcasted, broadcast it
2333 if (broadAxis != -1)
2335 opencv_onnx::NodeProto concat_node_proto = node_proto;
2336 const std::string& input1 = concat_node_proto.input(1);
2338 expandMid(layerParams.name, concat_node_proto, input1, outShape[broadAxis]);
2340 LayerParams concatLP;
2341 concatLP.name = layerParams.name + "/concat";
2342 concatLP.set("axis", broadAxis);
2343 concatLP.type = "Concat";
2344 concat_node_proto.set_output(0, concatLP.name);
2346 addLayer(concatLP, concat_node_proto);
2347 node_proto.set_input(1, concatLP.name);
2350 CV_Assert(axis != outShape.size());
2351 layerParams.set("axis", static_cast<int>(axis));
2352 layerParams.type = "Scale";
2354 addLayer(layerParams, node_proto);
2357 void ONNXImporter::parseConv(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
2359 opencv_onnx::NodeProto node_proto = node_proto_;
2360 CV_Assert(node_proto.input_size() >= 2);
2361 layerParams.type = "Convolution";
2362 for (int j = 1; j < node_proto.input_size(); j++) {
2363 if (constBlobs.find(node_proto.input(j)) != constBlobs.end())
2365 layerParams.blobs.push_back(getBlob(node_proto, j));
2368 int outCn = layerParams.blobs.empty() ? outShapes[node_proto.input(1)][0] : layerParams.blobs[0].size[0];
2369 layerParams.set("num_output", outCn);
2371 // Check for asymmetric padding in Conv2D
2372 if (layerParams.has("pad"))
2374 bool asymmetricPadding = false;
2375 DictValue pads = layerParams.get("pad");
2376 const int dims = pads.size() / 2;
2377 for (int i = 0; i < dims; ++i)
2379 if (pads.get<int>(i) != pads.get<int>(i + dims))
2381 asymmetricPadding = true;
2385 if (asymmetricPadding && pads.size() == 4) // [pad_t, pad_l, pad_b, pad_r]
2387 layerParams.erase("pad");
2388 // No paddings required for N, C axis
2389 std::vector<int> paddings(4, 0);
2390 // Add paddings for H, W axis
2391 for (int i = 0; i < dims; ++i)
2393 paddings.push_back(pads.get<int>(i));
2394 paddings.push_back(pads.get<int>(dims + i));
2397 padLp.name = layerParams.name + "/pad";
2398 padLp.type = "Padding";
2399 padLp.set("paddings", DictValue::arrayInt(&paddings[0], paddings.size()));
2401 opencv_onnx::NodeProto proto;
2402 proto.add_input(node_proto.input(0));
2403 proto.add_output(padLp.name);
2405 addLayer(padLp, proto);
2406 node_proto.set_input(0, padLp.name);
2409 addLayer(layerParams, node_proto);
2412 void ONNXImporter::parseConvTranspose(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2414 CV_Assert(node_proto.input_size() >= 2);
2415 layerParams.type = "Deconvolution";
2416 for (int j = 1; j < node_proto.input_size(); j++) {
2417 layerParams.blobs.push_back(getBlob(node_proto, j));
2419 layerParams.set("num_output", layerParams.blobs[0].size[1] * layerParams.get<int>("group", 1));
2420 layerParams.set("bias_term", node_proto.input_size() == 3);
2422 if (!layerParams.has("kernel_size"))
2423 CV_Error(Error::StsNotImplemented,
2424 "Required attribute 'kernel_size' is not present.");
2426 if (layerParams.has("output_shape"))
2428 const DictValue& outShape = layerParams.get("output_shape");
2429 DictValue strides = layerParams.get("stride");
2430 DictValue kernel = layerParams.get("kernel_size");
2433 std::vector<int> adjust_pads;
2434 if (layerParams.has("pad_mode"))
2436 padMode = toUpperCase(layerParams.get<String>("pad_mode"));
2437 if (padMode != "SAME" && padMode != "VALID")
2438 CV_Error(Error::StsError, "Unsupported padding mode " + padMode);
2440 for (int i = 0; i < strides.size(); i++)
2442 int sz = outShape.get<int>(2 + i);
2443 int stride = strides.get<int>(i);
2444 adjust_pads.push_back(padMode == "SAME"? (sz - 1) % stride :
2445 (sz - kernel.get<int>(i)) % stride);
2447 layerParams.set("adj", DictValue::arrayInt(&adjust_pads[0], adjust_pads.size()));
2450 else if (layerParams.has("output_padding"))
2452 replaceLayerParam(layerParams, "output_padding", "adj");
2454 addLayer(layerParams, node_proto);
2457 void ONNXImporter::parseTranspose(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2459 int depth = layerParams.get<int>("depth", CV_32F);
2460 layerParams.type = (depth == CV_8S) ? "PermuteInt8" : "Permute";
2461 replaceLayerParam(layerParams, "perm", "order");
2462 if (!layerParams.has("order")) {
2463 MatShape inpShape = outShapes[node_proto.input(0)];
2464 size_t dims = inpShape.size();
2465 std::vector<int> perm(dims);
2466 for (size_t d = 0; d < dims; ++d)
2468 perm[d] = static_cast<int>(dims - 1 - d);
2470 layerParams.set("order", DictValue::arrayInt(perm.data(), perm.size()));
2473 CV_Assert(node_proto.input_size() == 1);
2474 if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
2476 std::vector<Mat> inputs(1, getBlob(node_proto, 0)), transposed;
2477 runLayer(layerParams, inputs, transposed);
2478 CV_Assert(transposed.size() == 1);
2479 addConstant(node_proto.output(0), transposed[0]);
2482 addLayer(layerParams, node_proto);
2485 void ONNXImporter::parseSqueeze(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2487 CV_Assert_N(node_proto.input_size() == 1, layerParams.has("axes"));
2488 DictValue axes_dict = layerParams.get("axes");
2489 MatShape inpShape = outShapes[node_proto.input(0)];
2491 std::vector<bool> maskedAxes(inpShape.size(), false);
2492 for (int i = 0; i < axes_dict.size(); ++i)
2494 int axis = axes_dict.getIntValue(i);
2495 CV_CheckLE(axis, static_cast<int>(inpShape.size()), "Squeeze axis");
2496 maskedAxes[axis] = inpShape[axis] == 1;
2499 for (int i = 0; i < inpShape.size(); ++i)
2502 outShape.push_back(inpShape[i]);
2504 if (outShape.size() != inpShape.size())
2506 layerParams.type = "Reshape";
2507 layerParams.set("dim", DictValue::arrayInt(&outShape[0], outShape.size()));
2508 if (hasDynamicShapes)
2510 std::vector<int> dynamicAxes;
2511 std::vector<int> inputIndices;
2512 for (int index = 0; index < inpShape.size(); ++index)
2514 if (!maskedAxes[index])
2515 inputIndices.push_back(index);
2517 for (int index = 0; index < outShape.size(); ++index)
2518 dynamicAxes.push_back(index);
2519 layerParams.set("dynamic_axes", DictValue::arrayInt(dynamicAxes.data(), dynamicAxes.size()));
2520 layerParams.set("input_indices", DictValue::arrayInt(inputIndices.data(), inputIndices.size()));
2524 layerParams.type = "Identity";
2526 if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
2528 Mat inp = getBlob(node_proto, 0);
2529 Mat out = inp.reshape(1, outShape);
2530 out.dims = outShape.size(); // to workaround dims == 1
2531 addConstant(node_proto.output(0), out);
2534 int depth = layerParams.get<int>("depth", CV_32F);
2535 layerParams.type += (depth == CV_8S) ? "Int8" : "";
2536 addLayer(layerParams, node_proto);
2539 void ONNXImporter::parseFlatten(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
2541 opencv_onnx::NodeProto node_proto = node_proto_;
2542 CV_CheckEQ(node_proto.input_size(), 1, "");
2543 int axis_ = layerParams.get<int>("axis", 1);
2544 if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
2546 Mat input = getBlob(node_proto, 0);
2547 int axis = normalize_axis(axis_, input.dims);
2549 int out_size[2] = {1, 1};
2550 for (int i = 0; i < axis; ++i)
2552 out_size[0] *= input.size[i];
2554 for (int i = axis; i < input.dims; ++i)
2556 out_size[1] *= input.size[i];
2559 Mat output = input.reshape(1, 2, out_size);
2560 addConstant(node_proto.output(0), output);
2563 IterShape_t shapeIt = outShapes.find(node_proto.input(0));
2564 CV_Assert(shapeIt != outShapes.end());
2565 MatShape inpShape = shapeIt->second;
2566 int axis = normalize_axis(axis_, inpShape.size());
2568 if (axis == 0 || axis == inpShape.size())
2570 LayerParams reshapeLp;
2571 reshapeLp.name = layerParams.name + "/reshape";
2572 reshapeLp.type = "Reshape";
2573 CV_Assert(layer_id.find(reshapeLp.name) == layer_id.end());
2575 inpShape.insert(axis == 0 ? inpShape.begin() : inpShape.end(), 1);
2576 reshapeLp.set("dim", DictValue::arrayInt(&inpShape[0], inpShape.size()));
2578 opencv_onnx::NodeProto proto;
2579 proto.add_input(node_proto.input(0));
2580 proto.add_output(reshapeLp.name);
2581 addLayer(reshapeLp, proto);
2582 node_proto.set_input(0, reshapeLp.name);
2586 LayerParams first_pass;
2587 first_pass.name = layerParams.name + "/flatten";
2588 CV_Assert(layer_id.find(first_pass.name) == layer_id.end());
2589 first_pass.type = "Flatten";
2590 first_pass.set("axis", 0);
2591 first_pass.set("end_axis", axis - 1);
2593 opencv_onnx::NodeProto proto;
2594 proto.add_input(node_proto.input(0));
2595 proto.add_output(first_pass.name);
2596 addLayer(first_pass, proto);
2598 layerParams.set("axis", 1);
2599 node_proto.set_input(0, first_pass.name);
2600 addLayer(layerParams, node_proto);
2603 void ONNXImporter::parseUnsqueeze(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2605 CV_Assert(node_proto.input_size() == 1 || node_proto.input_size() == 2);
2607 if (node_proto.input_size() == 2)
2609 Mat blob = getBlob(node_proto, 1);
2610 axes = DictValue::arrayInt(blob.ptr<int>(), blob.total());
2613 axes = layerParams.get("axes");
2615 if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
2618 Mat input = getBlob(node_proto, 0);
2620 std::vector<int> dims;
2621 for (int j = 0; j < input.dims; j++) {
2622 dims.push_back(input.size[j]);
2624 CV_Assert(axes.getIntValue(axes.size()-1) <= dims.size());
2625 for (int j = 0; j < axes.size(); j++) {
2626 const int idx = axes.getIntValue(j);
2627 CV_Assert(idx <= dims.size());
2628 dims.insert(dims.begin() + idx, 1);
2631 Mat out = input.reshape(0, dims);
2632 addConstant(node_proto.output(0), out);
2637 if (axes.size() != 1)
2638 CV_Error(Error::StsNotImplemented, "Multidimensional unsqueeze");
2640 int depth = layerParams.get<int>("depth", CV_32F);
2642 MatShape inpShape = outShapes[node_proto.input(0)];
2643 int axis = axes.getIntValue(0);
2644 CV_Assert(0 <= axis && axis <= inpShape.size());
2645 std::vector<int> outShape = inpShape;
2646 outShape.insert(outShape.begin() + axis, 1);
2647 layerParams.type = (depth == CV_8S) ? "ReshapeInt8" : "Reshape";
2648 layerParams.set("dim", DictValue::arrayInt(&outShape[0], outShape.size()));
2649 if (hasDynamicShapes)
2651 std::vector<int> dynamicAxes;
2652 std::vector<int> inputIndices;
2653 for (int index = 0; index < outShape.size(); ++index) {
2655 dynamicAxes.push_back(index);
2657 for (int index = 0; index < inpShape.size(); ++index)
2658 inputIndices.push_back(index);
2659 layerParams.set("dynamic_axes", DictValue::arrayInt(dynamicAxes.data(), dynamicAxes.size()));
2660 layerParams.set("input_indices", DictValue::arrayInt(inputIndices.data(), inputIndices.size()));
2662 addLayer(layerParams, node_proto);
2665 void ONNXImporter::parseExpand(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
2667 opencv_onnx::NodeProto node_proto = node_proto_;
2668 CV_CheckEQ(node_proto.input_size(), 2, "");
2669 const std::string& input0 = node_proto.input(0);
2670 const std::string& input1 = node_proto.input(1);
2671 const std::string output_name = node_proto.output(0);
2672 Mat newShapeMat = getBlob(input1);
2673 MatShape targetShape(newShapeMat.ptr<int>(), newShapeMat.ptr<int>() + newShapeMat.total());
2676 bool haveVariables = constBlobs.find(input0) == constBlobs.end();
2679 IterShape_t shapeIt = outShapes.find(input0);
2680 CV_Assert(shapeIt != outShapes.end());
2681 inpShape = shapeIt->second;
2685 inpShape = shape(getBlob(input0));
2688 String srcName = input0;
2689 // Unsqueeze and repeat along new axis
2690 if (targetShape.size() == inpShape.size() + 1)
2692 inpShape.insert(inpShape.begin(), targetShape.size() - inpShape.size(), 1);
2693 for (int i = 0; i < targetShape.size(); i++)
2695 if (abs(targetShape[i]) == 1)
2696 targetShape[i] = inpShape[i];
2700 LayerParams reshapeLp;
2701 reshapeLp.name = layerParams.name + "/reshape";
2702 reshapeLp.type = "Reshape";
2703 CV_Assert(layer_id.find(reshapeLp.name) == layer_id.end());
2704 reshapeLp.set("dim", DictValue::arrayInt(&inpShape[0], inpShape.size()));
2706 opencv_onnx::NodeProto proto;
2707 proto.add_input(node_proto.input(0));
2708 proto.add_output(reshapeLp.name);
2709 addLayer(reshapeLp, proto);
2710 srcName = reshapeLp.name;
2713 CV_CheckEQ(inpShape.size(), targetShape.size(), "Unsupported Expand op with different dims");
2715 std::vector<int> broadcast_axes;
2716 // shapes aren't right-aligned here because targetShape.size() == inpShape.size()
2717 for (int i = 0; i < targetShape.size(); i++)
2719 if (targetShape[i] != inpShape[i])
2721 if (inpShape[i] == 1)
2723 broadcast_axes.push_back(i);
2725 else if (targetShape[i] != 1)
2727 CV_Error(Error::StsError, format("Could not be broadcast by axis: %d", i));
2734 if (broadcast_axes.size() > 1)
2735 CV_Error(Error::StsNotImplemented, "Expand op doesn't support multiple axes for constant input");
2737 if (broadcast_axes.empty())
2739 addConstant(output_name, getBlob(node_proto, 0));
2743 Mat input = getBlob(node_proto, 0);
2744 input = input.reshape(0, total(inpShape, 0, broadcast_axes[0]));
2745 Mat output = cv::repeat(input, 1, targetShape[broadcast_axes[0]]);
2746 output = output.reshape(0, targetShape);
2747 addConstant(output_name, output);
2751 if (broadcast_axes.size() == 2 &&
2752 broadcast_axes[0] == broadcast_axes[1] - 1 && broadcast_axes[1] == inpShape.size() - 1)
2754 LayerParams constParams;
2755 constParams.name = layerParams.name + "/const";
2756 CV_Assert(layer_id.find(constParams.name) == layer_id.end());
2757 constParams.type = "Const";
2759 Mat inp = Mat::ones(newShapeMat.total(), newShapeMat.ptr<int>(), CV_32F);
2760 constParams.blobs.push_back(inp);
2762 opencv_onnx::NodeProto proto;
2763 proto.add_output(constParams.name);
2764 addLayer(constParams, proto);
2766 layerParams.type = "Scale";
2767 layerParams.set("bias_term", false);
2768 node_proto.set_input(0, constParams.name);
2769 node_proto.set_input(1, srcName);
2771 else if (broadcast_axes.size() == 1 && broadcast_axes[0] <= 1)
2773 expandMid(layerParams.name, node_proto, srcName, targetShape[broadcast_axes[0]]);
2775 layerParams.set("axis", broadcast_axes[0]);
2776 layerParams.type = "Concat";
2777 node_proto.set_output(0, output_name);
2779 else if (broadcast_axes.empty())
2781 layerParams.type = "Identity";
2784 CV_Error(Error::StsNotImplemented, "Unsupported Expand op");
2785 addLayer(layerParams, node_proto);
2788 void ONNXImporter::parseReshape(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2790 CV_Assert(node_proto.input_size() == 2 || layerParams.has("shape"));
2791 int depth = layerParams.get<int>("depth", CV_32F);
2792 layerParams.type += (depth == CV_8S) ? "Int8" : "";
2794 if (node_proto.input_size() == 2) {
2795 Mat blob = getBlob(node_proto, 1);
2796 CV_Assert(blob.type() == CV_32SC1);
2798 layerParams.set("dim", DictValue::arrayInt<int*>(blob.ptr<int>(), blob.total()));
2800 if (layer_id.find(node_proto.input(0)) == layer_id.end()) {
2801 std::vector<Mat> inputs(1, getBlob(node_proto, 0)), outputs;
2802 runLayer(layerParams, inputs, outputs);
2803 addConstant(node_proto.output(0), outputs[0]);
2808 DictValue shape = layerParams.get("shape");
2809 std::vector<int> dim;
2810 for (int j = 0; j < shape.size(); j++) {
2811 dim.push_back(shape.getIntValue(j));
2814 if (layer_id.find(node_proto.input(0)) == layer_id.end()) {
2815 Mat input = getBlob(node_proto, 0);
2816 Mat out = input.reshape(0, dim);
2817 addConstant(node_proto.output(0), out);
2820 replaceLayerParam(layerParams, "shape", "dim");
2822 addLayer(layerParams, node_proto);
2825 void ONNXImporter::parsePad(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2827 int depth = layerParams.get<int>("depth", CV_32F);
2828 layerParams.type = (depth == CV_8S) ? "PaddingInt8" : "Padding";
2829 replaceLayerParam(layerParams, "mode", "type");
2830 if (node_proto.input_size() == 3 || node_proto.input_size() == 2)
2832 // Paddings are in order begin0, begin1, .. beginN, end0, end1, ..., endN.
2833 // We need to shuffle it to begin0, end0, begin1, end1, ...
2834 Mat paddings = getBlob(node_proto, 1).reshape(1, 2);
2835 paddings = paddings.t();
2836 layerParams.set("paddings", DictValue::arrayInt(paddings.ptr<int>(), paddings.total()));
2838 if (node_proto.input_size() == 3)
2840 Mat value = getBlob(node_proto, 2);
2841 float padValue = (depth == CV_8S) ? (float)value.ptr<int8_t>()[0] : value.ptr<float>()[0];
2842 layerParams.set("value", padValue);
2845 addLayer(layerParams, node_proto);
2848 void ONNXImporter::parseShape(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2850 CV_Assert(node_proto.input_size() == 1);
2851 IterShape_t shapeIt = outShapes.find(node_proto.input(0));
2852 CV_Assert(shapeIt != outShapes.end());
2853 const MatShape& inpShape = shapeIt->second;
2855 int dims = static_cast<int>(inpShape.size());
2856 Mat shapeMat(dims, 1, CV_32S);
2857 bool isDynamicShape = false;
2858 for (int j = 0; j < dims; ++j)
2860 int sz = inpShape[j];
2861 isDynamicShape |= (sz == 0);
2862 shapeMat.at<int>(j) = sz;
2864 shapeMat.dims = 1; // FIXIT Mat 1D
2868 CV_LOG_ERROR(NULL, "DNN/ONNX(Shape): dynamic 'zero' shapes are not supported, input " << toString(inpShape, node_proto.input(0)));
2869 CV_Assert(!isDynamicShape); // not supported
2871 addConstant(node_proto.output(0), shapeMat);
2874 void ONNXImporter::parseCast(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2876 if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
2878 Mat blob = getBlob(node_proto, 0);
2880 switch (layerParams.get<int>("to"))
2882 case opencv_onnx::TensorProto_DataType_FLOAT: type = CV_32F; break;
2883 case opencv_onnx::TensorProto_DataType_UINT8: type = CV_8U; break;
2884 case opencv_onnx::TensorProto_DataType_UINT16: type = CV_16U; break;
2885 case opencv_onnx::TensorProto_DataType_FLOAT16: type = CV_16S; break;
2886 case opencv_onnx::TensorProto_DataType_INT8:
2887 case opencv_onnx::TensorProto_DataType_INT16:
2888 case opencv_onnx::TensorProto_DataType_INT32:
2889 case opencv_onnx::TensorProto_DataType_INT64: type = CV_32S; break;
2890 default: type = blob.type();
2893 blob.convertTo(dst, type);
2894 dst.dims = blob.dims;
2895 addConstant(node_proto.output(0), dst);
2899 layerParams.type = "Identity";
2900 addLayer(layerParams, node_proto);
2903 void ONNXImporter::parseConstantFill(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2907 if (!layerParams.blobs.empty())
2909 CV_Assert(!layerParams.has("value"));
2910 depth = layerParams.blobs[0].depth();
2912 layerParams.blobs[0].convertTo(floats, CV_32F);
2913 fill_value = floats.at<float>(0, 0);
2916 fill_value = layerParams.get("value", 0);
2918 MatShape inpShape = getBlob(node_proto, 0);
2919 for (int i = 0; i < inpShape.size(); i++)
2920 CV_CheckGT(inpShape[i], 0, "");
2921 Mat tensor(inpShape.size(), &inpShape[0], depth, Scalar(fill_value));
2922 addConstant(node_proto.output(0), tensor);
2925 void ONNXImporter::parseGather(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
2927 opencv_onnx::NodeProto node_proto = node_proto_;
2928 CV_Assert(node_proto.input_size() == 2);
2929 Mat indexMat = getBlob(node_proto, 1);
2930 CV_Assert_N(indexMat.type() == CV_32S, indexMat.total() == 1);
2931 int index = indexMat.at<int>(0);
2932 int axis = layerParams.get<int>("axis", 0);
2934 if ((constBlobs.find(node_proto.input(0)) != constBlobs.end()))
2936 Mat input = getBlob(node_proto, 0);
2938 std::vector<cv::Range> ranges(input.dims, Range::all());
2939 ranges[axis] = Range(index, index + 1);
2941 out = input(ranges);
2942 MatShape outShape = shape(out);
2943 if (outShape.size() > 1)
2945 outShape.erase(outShape.begin() + axis);
2946 out.reshape(0, outShape);
2950 addConstant(node_proto.output(0), out);
2955 IterShape_t shapeIt = outShapes.find(node_proto.input(0));
2956 CV_Assert(shapeIt != outShapes.end());
2957 MatShape inpShape = shapeIt->second;
2959 LayerParams sliceLp;
2960 sliceLp.type = "Slice";
2961 sliceLp.name = inpShape.size() > 1 ? layerParams.name + "/slice" : layerParams.name;
2962 std::vector<int> begin(inpShape.size(), 0);
2963 std::vector<int> end(inpShape.size(), INT_MAX);
2964 begin[axis] = index;
2965 end[axis] = index + 1;
2967 cv::dnn::DictValue paramBegin = cv::dnn::DictValue::arrayInt(begin.data(), begin.size());
2968 cv::dnn::DictValue paramEnd = cv::dnn::DictValue::arrayInt(end.data(), end.size());
2969 sliceLp.set("begin", paramBegin);
2970 sliceLp.set("end", paramEnd);
2971 sliceLp.set("has_dynamic_shapes", hasDynamicShapes);
2973 if (inpShape.size() > 1)
2975 opencv_onnx::NodeProto proto;
2976 proto.add_input(node_proto.input(0));
2977 proto.add_output(sliceLp.name);
2978 addLayer(sliceLp, proto);
2980 inpShape.erase(inpShape.begin() + axis);
2981 layerParams.type = "Reshape";
2982 layerParams.set("axis", 0);
2983 layerParams.set("dim", DictValue::arrayInt(&inpShape[0], inpShape.size()));
2984 if (hasDynamicShapes)
2986 std::vector<int> dynamicAxes;
2987 std::vector<int> inputIndices;
2988 for (int index = 0; index < inpShape.size(); ++index)
2989 dynamicAxes.push_back(index);
2990 for (int index = 0; index < inpShape.size(); ++index)
2991 inputIndices.push_back(index);
2992 layerParams.set("dynamic_axes", DictValue::arrayInt(dynamicAxes.data(), dynamicAxes.size()));
2993 layerParams.set("input_indices", DictValue::arrayInt(inputIndices.data(), inputIndices.size()));
2995 node_proto.set_input(0, sliceLp.name);
2999 layerParams = sliceLp;
3002 addLayer(layerParams, node_proto);
3005 void ONNXImporter::parseConcat(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3007 bool hasVariableInps = false;
3008 for (int i = 0; i < node_proto.input_size(); ++i)
3010 if (layer_id.find(node_proto.input(i)) != layer_id.end())
3012 hasVariableInps = true;
3017 if (!hasVariableInps)
3019 std::vector<Mat> inputs(node_proto.input_size()), concatenated;
3020 // Due constant folding we can get inputs with different number of dimensions
3021 // Insert the missing dimension to inputs
3022 MatShape inputShape;
3023 for (size_t i = 0; i < inputs.size(); ++i)
3025 inputs[i] = getBlob(node_proto, i);
3026 if (inputs[i].size.dims() > inputShape.size())
3028 inputShape = shape(inputs[i]);
3032 // Concat-1 has default value for axis is 1: https://github.com/onnx/onnx/blob/master/docs/Changelog.md#Concat-1
3033 int axis = layerParams.get<int>("axis", 1);
3034 for (size_t i = 0; i < inputs.size(); ++i)
3036 MatShape targetShape = inputShape;
3037 targetShape[axis] = shape(inputs[i])[axis];
3038 CV_CheckEQ(total(targetShape), total(shape(inputs[i])), "");
3039 inputs[i] = inputs[i].reshape(0, targetShape);
3041 runLayer(layerParams, inputs, concatenated);
3043 CV_Assert(concatenated.size() == 1);
3044 addConstant(node_proto.output(0), concatenated[0]);
3049 for (int i = 0; i < node_proto.input_size(); ++i)
3051 if (constBlobs.find(node_proto.input(i)) != constBlobs.end())
3053 LayerParams constParams;
3054 constParams.name = node_proto.input(i);
3055 constParams.type = "Const";
3056 constParams.blobs.push_back(getBlob(node_proto, i));
3058 opencv_onnx::NodeProto proto;
3059 proto.add_output(constParams.name);
3060 addLayer(constParams, proto);
3064 addLayer(layerParams, node_proto);
3067 // https://github.com/onnx/onnx/blob/master/docs/Operators.md#Resize
3068 void ONNXImporter::parseResize(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3070 for (int i = 1; i < node_proto.input_size(); i++)
3071 CV_Assert(layer_id.find(node_proto.input(i)) == layer_id.end());
3073 int depth = layerParams.get<int>("depth", CV_32F);
3074 layerParams.type += (depth == CV_8S) ? "Int8" : "";
3076 if (layerParams.has("coordinate_transformation_mode"))
3078 String interp_mode = layerParams.get<String>("coordinate_transformation_mode");
3079 CV_Assert_N(interp_mode != "tf_crop_and_resize", interp_mode != "tf_half_pixel_for_nn");
3081 layerParams.set("align_corners", interp_mode == "align_corners");
3082 if (layerParams.get<String>("mode") == "linear")
3084 layerParams.set("mode", interp_mode == "pytorch_half_pixel" || interp_mode == "half_pixel" ?
3085 "opencv_linear" : "bilinear");
3088 if (layerParams.get<String>("mode") == "linear" && framework_name == "pytorch")
3089 layerParams.set("mode", "opencv_linear");
3091 // opset-10: input = [X, scales]
3092 // opset-11: input = [X, roi, scales] or [x, roi, scales, sizes]
3093 // opset-13: may have empty input, [X, "", "", sizes] or [x, "", scales]
3094 int scalesInputId = node_proto.input_size() == 2 ? 1 : 2;
3095 const std::string& scale_name = node_proto.input(scalesInputId);
3097 if(!scale_name.empty())
3098 scales = getBlob(node_proto, scalesInputId);
3100 if (!scales.empty())
3102 CV_CheckEQ(scales.total(), (size_t)4, "HCHW layout is expected");
3103 layerParams.set("zoom_factor_y", scales.at<float>(2));
3104 layerParams.set("zoom_factor_x", scales.at<float>(3));
3106 else if (node_proto.input_size() >= 4) // opset-11 [x, roi, scales, sizes] or opset-13: input = [X, "", "", sizes]
3108 const std::string& inputSizes = node_proto.input(3);
3109 if (constBlobs.find(inputSizes) != constBlobs.end())
3111 Mat shapes = getBlob(inputSizes);
3112 CV_CheckEQ(shapes.total(), (size_t)4, "HCHW layout is expected");
3113 CV_CheckDepth(shapes.depth(), shapes.depth() == CV_32S || shapes.depth() == CV_32F, "");
3114 if (shapes.depth() == CV_32F)
3115 shapes.convertTo(shapes, CV_32S);
3116 layerParams.set("width", shapes.at<int>(3));
3117 layerParams.set("height", shapes.at<int>(2));
3121 CV_Error(Error::StsNotImplemented, cv::format("ONNX/Resize: doesn't support dynamic non-constant 'sizes' input: %s", inputSizes.c_str()));
3126 CV_Error(Error::StsNotImplemented, "ONNX/Resize: can't find neither 'scale' nor destination sizes parameters");
3128 replaceLayerParam(layerParams, "mode", "interpolation");
3129 addLayer(layerParams, node_proto);
3132 void ONNXImporter::parseUpsample(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3134 //fused from Resize Subgraph
3135 if (layerParams.has("coordinate_transformation_mode"))
3137 String interp_mode = layerParams.get<String>("coordinate_transformation_mode");
3138 CV_Assert_N(interp_mode != "tf_crop_and_resize", interp_mode != "tf_half_pixel_for_nn");
3140 layerParams.set("align_corners", interp_mode == "align_corners");
3141 if (layerParams.get<String>("mode") == "linear")
3143 layerParams.set("mode", interp_mode == "pytorch_half_pixel" ?
3144 "opencv_linear" : "bilinear");
3147 if (layerParams.get<String>("mode") == "linear" && framework_name == "pytorch")
3148 layerParams.set("mode", "opencv_linear");
3150 layerParams.type = "Resize";
3151 if (layerParams.has("scales"))
3154 DictValue scales = layerParams.get("scales");
3155 CV_Assert(scales.size() == 4);
3156 layerParams.set("zoom_factor_y", scales.getIntValue(2));
3157 layerParams.set("zoom_factor_x", scales.getIntValue(3));
3159 else if (layerParams.has("height_scale") && layerParams.has("width_scale"))
3162 replaceLayerParam(layerParams, "height_scale", "zoom_factor_y");
3163 replaceLayerParam(layerParams, "width_scale", "zoom_factor_x");
3168 const std::string& input1 = node_proto.input(1);
3169 if (constBlobs.find(input1) != constBlobs.end())
3171 Mat scales = getBlob(input1);
3172 CV_Assert(scales.total() == 4);
3173 layerParams.set("zoom_factor_y", scales.at<float>(2));
3174 layerParams.set("zoom_factor_x", scales.at<float>(3));
3177 replaceLayerParam(layerParams, "mode", "interpolation");
3178 addLayer(layerParams, node_proto);
3181 void ONNXImporter::parseSoftMax(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3183 const std::string& layer_type = node_proto.op_type();
3184 layerParams.type = "Softmax";
3185 layerParams.set("log_softmax", layer_type == "LogSoftmax");
3186 addLayer(layerParams, node_proto);
3189 void ONNXImporter::parseDetectionOutput(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
3191 opencv_onnx::NodeProto node_proto = node_proto_;
3192 CV_CheckEQ(node_proto.input_size(), 3, "");
3193 if (constBlobs.find(node_proto.input(2)) != constBlobs.end())
3195 Mat priors = getBlob(node_proto, 2);
3197 LayerParams constParams;
3198 constParams.name = layerParams.name + "/priors";
3199 constParams.type = "Const";
3200 constParams.blobs.push_back(priors);
3202 opencv_onnx::NodeProto priorsProto;
3203 priorsProto.add_output(constParams.name);
3204 addLayer(constParams, priorsProto);
3206 node_proto.set_input(2, constParams.name);
3208 addLayer(layerParams, node_proto);
3211 void ONNXImporter::parseCumSum(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3213 layerParams.type = "CumSum";
3216 const std::string& input1 = node_proto.input(1);
3218 if (constBlobs.find(input1) != constBlobs.end())
3220 Mat axis_blob = getBlob(input1);
3221 CV_Assert(axis_blob.total() == 1u);
3222 layerParams.set("axis", axis_blob.at<int>(0));
3225 addLayer(layerParams, node_proto);
3228 void ONNXImporter::parseDepthToSpace(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
3230 // We parse "DepthToSpace" and "SpaceToDepth" in this function.
3231 opencv_onnx::NodeProto node_proto = node_proto_;
3232 const std::string& layer_type = node_proto.op_type();
3233 CV_Assert(layer_type == "DepthToSpace" || layer_type == "SpaceToDepth");
3236 CV_Assert(layerParams.has("blocksize"));
3237 int blocksize = layerParams.get<int>("blocksize");
3238 CV_Assert(blocksize > 0);
3240 // Get mode, only for "DepthToSpace"
3241 std::string modeType = layerParams.get<std::string>("mode", "DCR");
3243 MatShape inpShape = outShapes[node_proto.input(0)];
3244 CV_Assert(inpShape.size() == 4);
3245 int N = inpShape[0], C = inpShape[1], H = inpShape[2], W = inpShape[3];
3247 // Implement DepthToSpace and SpaceToDepth by the Reshape and Permute layer.
3248 std::array<int, 6> shape0, perm;
3249 std::array<int, 4> shape1;
3251 if (layer_type == "DepthToSpace")
3253 if (modeType == "DCR")
3255 shape0 = {N, blocksize, blocksize, C/(blocksize * blocksize), H, W};
3256 perm = {0, 3, 4, 1, 5, 2};
3257 shape1 = {N, C/(blocksize * blocksize), H * blocksize, W * blocksize};
3259 else if (modeType == "CRD")
3261 shape0 = {N, C/(blocksize * blocksize), blocksize, blocksize, H, W};
3262 perm = {0, 1, 4, 2, 5, 3};
3263 shape1 = {N, C/(blocksize * blocksize), H * blocksize, W * blocksize};
3266 CV_Error(Error::StsNotImplemented, "The mode of " + modeType + " in " + layer_type + " Layer is not supported");
3268 else // SpaceToDepth
3270 shape0 = {N, C, H/blocksize, blocksize, W/blocksize, blocksize};
3271 perm = {0, 3, 5, 1, 2, 4};
3272 shape1 = {N, C * blocksize * blocksize, H/blocksize, W/blocksize};
3276 LayerParams reshapeLp;
3277 reshapeLp.name = layerParams.name + "/reshape";
3278 reshapeLp.type = "Reshape";
3279 CV_Assert(layer_id.find(reshapeLp.name) == layer_id.end());
3280 reshapeLp.set("dim", DictValue::arrayInt(shape0.data(), shape0.size()));
3282 opencv_onnx::NodeProto protoReshape;
3283 protoReshape.add_input(node_proto.input(0));
3284 protoReshape.add_output(reshapeLp.name);
3285 addLayer(reshapeLp, protoReshape);
3288 LayerParams permuteLp;
3289 permuteLp.name = layerParams.name + "/permute";
3290 permuteLp.type = "Permute";
3291 CV_Assert(layer_id.find(permuteLp.name) == layer_id.end());
3292 permuteLp.set("order", DictValue::arrayInt(perm.data(), perm.size()));
3294 opencv_onnx::NodeProto protoPermute;
3295 protoPermute.add_input(reshapeLp.name);
3296 protoPermute.add_output(permuteLp.name);
3297 addLayer(permuteLp, protoPermute);
3300 layerParams.type = "Reshape";
3301 layerParams.set("dim", DictValue::arrayInt(shape1.data(), shape1.size()));
3303 node_proto.set_input(0, permuteLp.name);
3304 addLayer(layerParams, node_proto);
3307 void ONNXImporter::parseSimpleLayers(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3309 for (int j = 0; j < node_proto.input_size(); j++) {
3310 if (layer_id.find(node_proto.input(j)) == layer_id.end())
3311 layerParams.blobs.push_back(getBlob(node_proto, j));
3313 addLayer(layerParams, node_proto);
3316 void ONNXImporter::parseCustomLayer(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3318 const std::string& name = layerParams.name;
3319 std::string& layer_type = layerParams.type;
3320 const std::string& layer_type_domain = node_proto.has_domain() ? node_proto.domain() : std::string();
3321 if (!layer_type_domain.empty() && layer_type_domain != str_domain_ai_onnx)
3323 // append ONNX domain name
3324 static bool DNN_CUSTOM_ONNX_TYPE_INCLUDE_DOMAIN_NAME = utils::getConfigurationParameterBool("OPENCV_DNN_CUSTOM_ONNX_TYPE_INCLUDE_DOMAIN_NAME", true);
3325 if (DNN_CUSTOM_ONNX_TYPE_INCLUDE_DOMAIN_NAME)
3327 layer_type = layer_type_domain + "." + layer_type;
3331 CV_LOG_IF_INFO(NULL, !LayerFactory::isLayerRegistered(layer_type), "DNN/ONNX: unknown node type, try using custom handler for node with " << node_proto.input_size() << " inputs and " << node_proto.output_size() << " outputs: "
3332 << cv::format("[%s]:(%s)", layer_type.c_str(), name.c_str())
3335 parseSimpleLayers(layerParams, node_proto);
3338 void ONNXImporter::parseQuantDequant(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3340 CV_Assert(node_proto.input_size() == 3);
3341 layerParams.type = (node_proto.op_type() == "QuantizeLinear") ? "Quantize" : "Dequantize";
3343 if (node_proto.op_type() == "DequantizeLinear")
3345 Mat scale = getBlob(node_proto, 1);
3346 Mat zeropoint = getBlob(node_proto, 2);
3348 layerParams.set("scales", DictValue::arrayReal(scale.ptr<float>(), 1));
3349 layerParams.set("zeropoints", DictValue::arrayInt(zeropoint.ptr<int8_t>(), 1));
3351 addLayer(layerParams, node_proto);
3354 void ONNXImporter::parseQConv(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
3356 opencv_onnx::NodeProto node_proto = node_proto_;
3357 int ninputs = node_proto.input_size();
3358 CV_Assert(ninputs == 8 || ninputs == 9);
3360 Mat inp_sc = getBlob(node_proto, 1);
3361 Mat inp_zp = getBlob(node_proto, 2);
3363 if (layerParams.has("pad"))
3365 bool asymmetricPadding = false;
3366 DictValue pads = layerParams.get("pad");
3367 const int dims = pads.size() / 2;
3369 for (int i = 0; i < dims; ++i)
3371 if (pads.get<int>(i) != pads.get<int>(i + dims))
3373 asymmetricPadding = true;
3377 if (asymmetricPadding && pads.size() == 4)
3379 layerParams.erase("pad");
3380 std::vector<int> paddings(4, 0);
3381 for (int i = 0; i < dims; ++i)
3383 paddings.push_back(pads.get<int>(i));
3384 paddings.push_back(pads.get<int>(dims + i));
3387 padLp.name = layerParams.name + "/pad";
3388 padLp.type = "PaddingInt8";
3389 padLp.set("paddings", DictValue::arrayInt(&paddings[0], paddings.size()));
3390 padLp.set("depth", CV_8S);
3391 padLp.set("value", inp_zp.at<int8_t>(0));
3393 opencv_onnx::NodeProto proto;
3394 proto.add_input(node_proto.input(0));
3395 proto.add_output(padLp.name);
3397 addLayer(padLp, proto);
3398 node_proto.set_input(0, padLp.name);
3402 Mat weights = getBlob(node_proto, 3);
3403 int outCn = weights.size[0];
3404 Mat w_scale = getBlob(node_proto, 4);
3405 CV_Assert(w_scale.total() == 1 || w_scale.total() == outCn);
3406 bool per_channel = w_scale.total() == outCn ? true : false;
3407 Mat wt_sc = (w_scale.total() == outCn) ? w_scale : Mat(1, outCn, CV_32F, Scalar(w_scale.at<float>(0)));
3409 Mat out_sc = getBlob(node_proto, 6);
3410 Mat bias = (ninputs == 9) ? getBlob(node_proto, 8) : Mat::zeros(1, outCn, CV_32S);
3412 Mat weights_2d = weights.reshape(1, outCn);
3413 Mat biasFused(1, outCn, CV_32S);
3414 Mat outputMultiplier(1, outCn, CV_32F);
3415 for (int i = 0; i < outCn; i++)
3417 biasFused.at<int>(i) = bias.at<int>(i) - inp_zp.at<int8_t>(0)*(cv::sum(weights_2d.row(i))[0]);
3418 outputMultiplier.at<float>(i) = (inp_sc.at<float>(0) * wt_sc.at<float>(i)) / out_sc.at<float>(0);
3421 layerParams.type = "ConvolutionInt8";
3422 layerParams.set("num_output", outCn);
3423 layerParams.set("input_zeropoint", inp_zp.at<int8_t>(0));
3424 layerParams.set("input_scale",inp_sc.at<float>(0));
3425 layerParams.set("per_channel", per_channel);
3426 layerParams.blobs.push_back(weights);
3427 layerParams.blobs.push_back(biasFused);
3428 layerParams.blobs.push_back(outputMultiplier);
3429 addLayer(layerParams, node_proto);
3432 void ONNXImporter::parseQMatMul(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3434 int ninputs = node_proto.input_size();
3435 CV_Assert(ninputs == 8);
3437 if (constBlobs.find(node_proto.input(3)) == constBlobs.end())
3438 CV_Error(Error::StsNotImplemented, "Variable weights is not supported");
3440 int firstInpDims = outShapes[node_proto.input(0)].size();
3442 Mat inp_sc = getBlob(node_proto, 1);
3443 Mat inp_zp = getBlob(node_proto, 2);
3445 Mat weights = getBlob(node_proto, 3).t();
3446 int outCn = weights.size[0];
3447 int secondInpDims = weights.dims;
3449 Mat w_scale = getBlob(node_proto, 4);
3450 CV_Assert(w_scale.total() == 1 || w_scale.total() == outCn);
3451 bool per_channel = w_scale.total() == outCn ? true : false;
3452 Mat wt_sc = (w_scale.total() == outCn) ? w_scale : Mat(1, outCn, CV_32F, Scalar(w_scale.at<float>(0)));
3453 Mat out_sc = getBlob(node_proto, 6);
3455 Mat bias(1, outCn, CV_32S);
3456 Mat outputMultiplier(1, outCn, CV_32F);
3457 for (int i = 0; i < outCn; i++)
3459 bias.at<int>(i) = -inp_zp.at<int8_t>(0)*(cv::sum(weights.row(i))[0]);
3460 outputMultiplier.at<float>(i) = (inp_sc.at<float>(0) * wt_sc.at<float>(i)) / out_sc.at<float>(0);
3463 layerParams.type = "InnerProductInt8";
3464 layerParams.set("num_output", outCn);
3465 layerParams.set("axis", firstInpDims - secondInpDims + 1);
3466 layerParams.set("input_scale", inp_sc.at<float>(0));
3467 layerParams.set("input_zeropoint", inp_zp.at<int8_t>(0));
3468 layerParams.set("per_channel", per_channel);
3470 layerParams.blobs.push_back(weights);
3471 layerParams.blobs.push_back(bias);
3472 layerParams.blobs.push_back(outputMultiplier);
3473 addLayer(layerParams, node_proto);
3476 void ONNXImporter::parseQEltwise(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
3478 opencv_onnx::NodeProto node_proto = node_proto_;
3479 CV_Assert(node_proto.input_size() == 8);
3480 std::string op = (node_proto.op_type() == "QLinearAdd") ? "sum" : "prod";
3482 for (int i = 0; i < 4; i += 3)
3484 if (constBlobs.find(node_proto.input(i)) != constBlobs.end())
3488 Mat inp_0_sc = getBlob(node_proto, 1);
3489 Mat inp_0_zp = getBlob(node_proto, 2);
3491 Mat inp_1_sc = getBlob(node_proto, 4);
3492 Mat inp_1_zp = getBlob(node_proto, 5);
3494 // Set 2nd input as the const input
3497 cv::swap(inp_0_sc, inp_1_sc);
3498 cv::swap(inp_0_zp, inp_1_zp);
3501 float out_sc = getBlob(node_proto, 6).at<float>(0);
3502 int8_t out_zp = getBlob(node_proto, 7).at<int8_t>(0);
3504 std::vector<float> inp_scales = {inp_0_sc.at<float>(0), inp_1_sc.at<float>(0)};
3505 std::vector<int8_t> inp_zps = {inp_0_zp.at<int8_t>(0), inp_1_zp.at<int8_t>(0)};
3507 std::vector<float> coeffs;
3511 coeffs = {inp_scales[0]/out_sc, inp_scales[1]/out_sc};
3512 offset = out_zp - coeffs[0]*inp_zps[0] - coeffs[1]*inp_zps[1];
3516 coeffs = {inp_scales[0]/out_sc, inp_scales[1]};
3522 Mat blob = getBlob(node_proto, constId);
3523 if (blob.total() == 1)
3525 float val = inp_scales[1] * (blob.at<int8_t>(0) - inp_zps[1]);
3526 float scale = inp_scales[0] / out_sc;
3530 float shift = out_zp - scale*inp_zps[0];
3532 shift += (val/out_sc);
3534 LayerParams rescaleParams;
3535 rescaleParams.name = layerParams.name;
3536 rescaleParams.type = "Requantize";
3537 rescaleParams.set("depth", CV_8S);
3538 rescaleParams.set("scale", scale);
3539 rescaleParams.set("shift", shift);
3540 rescaleParams.set("isEltwise", true);
3541 addLayer(rescaleParams, node_proto);
3546 MatShape inpShape = outShapes[node_proto.input(3 - constId)];
3550 if (shape(blob) == inpShape)
3552 LayerParams constParams;
3553 constParams.name = layerParams.name + "/const";
3554 constParams.type = "ConstInt8";
3555 constParams.set("depth", CV_8S);
3556 constParams.set("scales", DictValue::arrayReal(inp_1_sc.ptr<float>(), 1));
3557 constParams.set("zeropoints", DictValue::arrayInt(inp_1_zp.ptr<int8_t>(), 1));
3558 constParams.blobs.push_back(blob);
3560 int id = dstNet.addLayer(constParams.name, constParams.type, CV_8S, constParams);
3561 layer_id.insert(std::make_pair(constParams.name, LayerInfo(id, 0)));
3562 outShapes[constParams.name] = shape(blob);
3563 node_proto.set_input(constId, constParams.name);
3565 layerParams.type = "EltwiseInt8";
3566 layerParams.set("operation", op);
3567 layerParams.set("coeff", DictValue::arrayReal(coeffs.data(), coeffs.size()));
3568 layerParams.set("offset", offset);
3572 layerParams.type = "ScaleInt8";
3573 layerParams.set("bias_term", op == "sum");
3575 for (int i = 0; i < graph_proto.initializer_size(); i++)
3577 opencv_onnx::TensorProto tensor_proto = graph_proto.initializer(i);
3578 if (tensor_proto.name() == node_proto.input(constId))
3580 axis = inpShape.size() - tensor_proto.dims_size();
3584 layerParams.set("axis", axis);
3585 blob = blob.reshape(1, 1);
3586 Mat blob_dequantized;
3587 blob.convertTo(blob_dequantized, CV_32F, inp_scales[1], -(inp_scales[1] * inp_zps[1]));
3588 layerParams.blobs.push_back(blob_dequantized);
3592 else if (outShapes[node_proto.input(0)] == outShapes[node_proto.input(3)])
3594 layerParams.type = "EltwiseInt8";
3595 layerParams.set("operation", op);
3596 layerParams.set("coeff", DictValue::arrayReal(coeffs.data(), coeffs.size()));
3597 layerParams.set("offset", offset);
3601 layerParams.type = "ScaleInt8";
3602 layerParams.set("bias_term", op == "sum");
3605 layerParams.set("input_scales", DictValue::arrayReal(inp_scales.data(), inp_scales.size()));
3606 layerParams.set("input_zeropoints", DictValue::arrayInt(inp_zps.data(), inp_zps.size()));
3607 addLayer(layerParams, node_proto);
3610 void ONNXImporter::parseQLeakyRelu(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3612 CV_Assert(node_proto.input_size() == 5);
3614 float slope = layerParams.get<float>("alpha");
3615 float inp_sc = getBlob(node_proto, 1).at<float>(0);
3616 int8_t inp_zp = getBlob(node_proto, 2).at<int8_t>(0);
3617 float out_sc = getBlob(node_proto, 3).at<float>(0);
3618 int8_t out_zp = getBlob(node_proto, 4).at<int8_t>(0);
3620 Mat lookUpTable(1, 256, CV_8S);
3621 int8_t* table = lookUpTable.ptr<int8_t>();
3622 for (int i = -128; i < 128; i++)
3624 float x = inp_sc*(i - inp_zp);
3625 float y = x >= 0.f ? x : slope*x;
3626 int quantized = out_zp + cvRound(y/out_sc);
3627 table[i+128] = saturate_cast<int8_t>(quantized);
3630 layerParams.type = "ReLUInt8";
3631 layerParams.set("input_scale", inp_sc);
3632 layerParams.set("input_zeropoint", inp_zp);
3633 layerParams.set("slope", slope);
3634 layerParams.blobs.push_back(lookUpTable);
3635 addLayer(layerParams, node_proto);
3638 void ONNXImporter::parseQSigmoid(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3640 CV_Assert(node_proto.input_size() == 5);
3642 float inp_sc = getBlob(node_proto, 1).at<float>(0);
3643 int8_t inp_zp = getBlob(node_proto, 2).at<int8_t>(0);
3644 float out_sc = getBlob(node_proto, 3).at<float>(0);
3645 int8_t out_zp = getBlob(node_proto, 4).at<int8_t>(0);
3647 Mat lookUpTable(1, 256, CV_8S);
3648 int8_t* table = lookUpTable.ptr<int8_t>();
3649 for (int i = -128; i < 128; i++)
3651 float x = inp_sc*(i - inp_zp);
3652 float y = 1.f/(1.f + std::exp(-x));
3653 int quantized = out_zp + cvRound(y/out_sc);
3654 table[i+128] = saturate_cast<int8_t>(quantized);
3657 layerParams.type = "SigmoidInt8";
3658 layerParams.set("input_scale", inp_sc);
3659 layerParams.set("input_zeropoint", inp_zp);
3660 layerParams.blobs.push_back(lookUpTable);
3661 addLayer(layerParams, node_proto);
3664 void ONNXImporter::parseQAvgPool(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3666 CV_Assert(node_proto.input_size() == 5);
3667 float inp_sc = getBlob(node_proto, 1).at<float>(0);
3668 int8_t inp_zp = getBlob(node_proto, 2).at<int8_t>(0);
3669 float out_sc = getBlob(node_proto, 3).at<float>(0);
3671 layerParams.type = "PoolingInt8";
3672 layerParams.set("pool", "ave");
3673 layerParams.set("global_pooling", node_proto.op_type() == "QLinearGlobalAveragePool");
3674 layerParams.set("multiplier", inp_sc/out_sc);
3675 layerParams.set("input_zeropoint", inp_zp);
3676 addLayer(layerParams, node_proto);
3679 void ONNXImporter::parseQConcat(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
3681 opencv_onnx::NodeProto node_proto = node_proto_;
3682 layerParams.type = "ConcatInt8";
3683 int num_inputs = node_proto.input_size();
3685 float out_scale = getBlob(node_proto, 0).at<float>(0);
3686 int out_zp = getBlob(node_proto, 1).at<int8_t>(0);
3688 for (int i = 2; i < num_inputs; i += 3)
3690 float inp_scale = getBlob(node_proto, i + 1).at<float>(0);
3691 int inp_zp = getBlob(node_proto, i + 2).at<int8_t>(0);
3693 if (inp_scale != out_scale || inp_zp != out_zp)
3695 float scale = inp_scale/out_scale;
3696 float shift = out_zp - scale*inp_zp;
3698 if (constBlobs.find(node_proto.input(i)) != constBlobs.end())
3700 Mat blob = getBlob(node_proto, i);
3702 blob.convertTo(blob_rescaled, CV_8S, scale, shift);
3703 constBlobs[node_proto.input(i)] = blob_rescaled;
3707 LayerParams rescaleParams;
3708 rescaleParams.name = node_proto.input(i) + "/rescale";
3709 rescaleParams.type = "Requantize";
3710 rescaleParams.set("depth", CV_8S);
3711 rescaleParams.set("scale", scale);
3712 rescaleParams.set("shift", shift);
3713 rescaleParams.set("isEltwise", false);
3715 opencv_onnx::NodeProto proto;
3716 proto.add_input(node_proto.input(i));
3717 proto.add_output(rescaleParams.name);
3718 addLayer(rescaleParams, proto);
3719 node_proto.set_input(i, rescaleParams.name);
3724 bool hasVariableInps = false;
3725 for (int i = 2; i < num_inputs; i += 3)
3727 if (layer_id.find(node_proto.input(i)) != layer_id.end())
3729 hasVariableInps = true;
3734 if (!hasVariableInps)
3736 std::vector<Mat> inputs, concatenated;
3737 MatShape inputShape;
3738 for (size_t i = 2; i < num_inputs; i += 3)
3740 Mat blob = getBlob(node_proto, i);
3741 if (blob.size.dims() > inputShape.size())
3743 inputShape = shape(blob);
3745 inputs.push_back(blob);
3748 int axis = layerParams.get<int>("axis", 1);
3749 for (size_t i = 0; i < inputs.size(); ++i)
3751 MatShape targetShape = inputShape;
3752 targetShape[axis] = shape(inputs[i])[axis];
3753 CV_CheckEQ(total(targetShape), total(shape(inputs[i])), "");
3754 inputs[i] = inputs[i].reshape(0, targetShape);
3756 runLayer(layerParams, inputs, concatenated);
3757 CV_Assert(concatenated.size() == 1);
3758 addConstant(layerParams.name, concatenated[0]);
3763 for (int i = 2; i < num_inputs; i += 3)
3765 if (constBlobs.find(node_proto.input(i)) != constBlobs.end())
3767 LayerParams constParams;
3768 constParams.name = node_proto.input(i);
3769 constParams.type = "ConstInt8";
3770 constParams.blobs.push_back(getBlob(node_proto, i));
3771 constParams.set("depth", CV_8S);
3773 opencv_onnx::NodeProto proto;
3774 proto.add_output(constParams.name);
3775 addLayer(constParams, proto);
3779 addLayer(layerParams, node_proto);
3782 // Domain: ai.onnx (default)
3783 // URL: https://github.com/onnx/onnx/blob/master/docs/Operators.md
3784 void ONNXImporter::buildDispatchMap_ONNX_AI(int opset_version)
3786 CV_UNUSED(opset_version);
3787 DispatchMap dispatch;
3789 dispatch["ArgMax"] = dispatch["ArgMin"] = &ONNXImporter::parseArg;
3790 dispatch["MaxUnpool"] = &ONNXImporter::parseMaxUnpool;
3791 dispatch["MaxPool"] = &ONNXImporter::parseMaxPool;
3792 dispatch["AveragePool"] = &ONNXImporter::parseAveragePool;
3793 dispatch["GlobalAveragePool"] = dispatch["GlobalMaxPool"] = &ONNXImporter::parseGlobalPool;
3794 dispatch["ReduceMax"] = dispatch["ReduceMin"] = dispatch["ReduceMean"] = dispatch["ReduceSum"] = dispatch["ReduceMax"] =
3795 dispatch["ReduceMin"] = dispatch["ReduceSumSquare"] = dispatch["ReduceProd"] = dispatch["ReduceL1"] =
3796 dispatch["ReduceL2"] = dispatch["ReduceLogSum"] = dispatch["ReduceLogSumExp"] = &ONNXImporter::parseReduce;
3797 dispatch["Slice"] = &ONNXImporter::parseSlice;
3798 dispatch["Split"] = &ONNXImporter::parseSplit;
3799 dispatch["Add"] = dispatch["Sum"] = dispatch["Sub"] = &ONNXImporter::parseBias;
3800 dispatch["Pow"] = &ONNXImporter::parsePow;
3801 dispatch["Min"] = dispatch["Max"] = &ONNXImporter::parseMinMax;
3802 dispatch["Neg"] = &ONNXImporter::parseNeg;
3803 dispatch["Constant"] = &ONNXImporter::parseConstant;
3804 dispatch["LSTM"] = &ONNXImporter::parseLSTM;
3805 dispatch["GRU"] = &ONNXImporter::parseGRU;
3806 dispatch["ImageScaler"] = &ONNXImporter::parseImageScaler;
3807 dispatch["Clip"] = &ONNXImporter::parseClip;
3808 dispatch["LeakyRelu"] = &ONNXImporter::parseLeakyRelu;
3809 dispatch["Relu"] = &ONNXImporter::parseRelu;
3810 dispatch["Elu"] = &ONNXImporter::parseElu;
3811 dispatch["Tanh"] = &ONNXImporter::parseTanh;
3812 dispatch["Abs"] = &ONNXImporter::parseAbs;
3813 dispatch["Equal"] = dispatch["Greater"] = dispatch["Less"] = &ONNXImporter::parseCompare;
3814 dispatch["PRelu"] = &ONNXImporter::parsePRelu;
3815 dispatch["LRN"] = &ONNXImporter::parseLRN;
3816 dispatch["InstanceNormalization"] = &ONNXImporter::parseInstanceNormalization;
3817 dispatch["BatchNormalization"] = &ONNXImporter::parseBatchNormalization;
3818 dispatch["Gemm"] = &ONNXImporter::parseGemm;
3819 dispatch["MatMul"] = &ONNXImporter::parseMatMul;
3820 dispatch["Mul"] = dispatch["Div"] = &ONNXImporter::parseMul;
3821 dispatch["Conv"] = &ONNXImporter::parseConv;
3822 dispatch["ConvTranspose"] = &ONNXImporter::parseConvTranspose;
3823 dispatch["Transpose"] = &ONNXImporter::parseTranspose;
3824 dispatch["Squeeze"] = &ONNXImporter::parseSqueeze;
3825 dispatch["Flatten"] = &ONNXImporter::parseFlatten;
3826 dispatch["Unsqueeze"] = &ONNXImporter::parseUnsqueeze;
3827 dispatch["Expand"] = &ONNXImporter::parseExpand;
3828 dispatch["Reshape"] = &ONNXImporter::parseReshape;
3829 dispatch["Pad"] = &ONNXImporter::parsePad;
3830 dispatch["Shape"] = &ONNXImporter::parseShape;
3831 dispatch["Cast"] = &ONNXImporter::parseCast;
3832 dispatch["ConstantFill"] = dispatch["ConstantOfShape"] = &ONNXImporter::parseConstantFill;
3833 dispatch["Gather"] = &ONNXImporter::parseGather;
3834 dispatch["Concat"] = &ONNXImporter::parseConcat;
3835 dispatch["Resize"] = &ONNXImporter::parseResize;
3836 dispatch["Upsample"] = &ONNXImporter::parseUpsample;
3837 dispatch["SoftMax"] = dispatch["LogSoftmax"] = &ONNXImporter::parseSoftMax;
3838 dispatch["DetectionOutput"] = &ONNXImporter::parseDetectionOutput;
3839 dispatch["CumSum"] = &ONNXImporter::parseCumSum;
3840 dispatch["SpaceToDepth"] = dispatch["DepthToSpace"] = &ONNXImporter::parseDepthToSpace;
3842 std::vector<std::string> simpleLayers{"Acos", "Acosh", "Asin", "Asinh", "Atan", "Atanh", "Ceil", "Celu", "Cos",
3843 "Cosh", "Dropout", "Erf", "Exp", "Floor", "HardSigmoid", "HardSwish",
3844 "Identity", "Log", "Round", "Reciprocal", "Selu", "Sign", "Sigmoid", "Sin", "Sinh", "Softmax",
3845 "Softplus", "Softsign", "Shrink", "Sqrt", "Tan", "ThresholdedRelu"};
3846 for (const auto& name : simpleLayers)
3848 dispatch[name] = &ONNXImporter::parseSimpleLayers;
3851 // ai.onnx: opset 10+
3852 dispatch["QuantizeLinear"] = dispatch["DequantizeLinear"] = &ONNXImporter::parseQuantDequant;
3853 dispatch["QLinearConv"] = &ONNXImporter::parseQConv;
3854 dispatch["QLinearMatMul"] = &ONNXImporter::parseQMatMul;
3856 domain_dispatch_map[str_domain_ai_onnx] = dispatch;
3859 // Domain: com.microsoft
3860 // URL: https://github.com/microsoft/onnxruntime/blob/master/docs/ContribOperators.md
3861 void ONNXImporter::buildDispatchMap_COM_MICROSOFT(int opset_version)
3863 CV_UNUSED(opset_version);
3864 DispatchMap dispatch;
3866 dispatch["QLinearAdd"] = dispatch["QLinearMul"] = &ONNXImporter::parseQEltwise;
3867 dispatch["QLinearAveragePool"] = dispatch["QLinearGlobalAveragePool"] = &ONNXImporter::parseQAvgPool;
3868 dispatch["QLinearLeakyRelu"] = &ONNXImporter::parseQLeakyRelu;
3869 dispatch["QLinearSigmoid"] = &ONNXImporter::parseQSigmoid;
3870 dispatch["QLinearConcat"] = &ONNXImporter::parseQConcat;
3872 domain_dispatch_map["com.microsoft"] = dispatch;
3876 Net readNetFromONNX(const String& onnxFile)
3878 return detail::readNetDiagnostic<ONNXImporter>(onnxFile.c_str());
3881 Net readNetFromONNX(const char* buffer, size_t sizeBuffer)
3883 return detail::readNetDiagnostic<ONNXImporter>(buffer, sizeBuffer);
3886 Net readNetFromONNX(const std::vector<uchar>& buffer)
3888 return readNetFromONNX(reinterpret_cast<const char*>(buffer.data()), buffer.size());
3891 Mat readTensorFromONNX(const String& path)
3893 std::fstream input(path.c_str(), std::ios::in | std::ios::binary);
3896 CV_Error(Error::StsBadArg, cv::format("Can't read ONNX file: %s", path.c_str()));
3899 opencv_onnx::TensorProto tensor_proto = opencv_onnx::TensorProto();
3900 if (!tensor_proto.ParseFromIstream(&input))
3902 CV_Error(Error::StsUnsupportedFormat, cv::format("Failed to parse ONNX data: %s", path.c_str()));
3904 Mat mat = getMatFromTensor(tensor_proto);
3905 releaseONNXTensor(tensor_proto);
3909 CV__DNN_INLINE_NS_END