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 if (layer_type == "ReduceSum" && node_proto.input_size() == 2)
1185 // TODO support the opset 13 of ReduceSum.
1186 // in opset 13, the ReduceSum has two input, it takes axes as input instead of attribute
1187 // details:https://github.com/onnx/onnx/issues/3420#issuecomment-844295687
1188 CV_Error(Error::StsNotImplemented, "Unsupported " + layer_type + " operation of opset 13, please try to "
1189 "re-export the onnx model with opset 11.");
1192 MatShape inpShape = outShapes[node_proto.input(0)];
1193 std::vector<bool> shouldDelete(inpShape.size(), false);
1195 if (layerParams.has("axes"))
1197 DictValue axes = layerParams.get("axes");
1198 for (int i = 0; i < axes.size(); i++)
1200 int axis = normalize_axis(axes.get<int>(i), inpShape.size());
1201 shouldDelete[axis] = true;
1206 for (int i = 0; i < inpShape.size(); i++)
1208 shouldDelete[i] = true;
1212 MatShape targetShape;
1213 for (int i = 0; i < inpShape.size(); ++i)
1215 if (!shouldDelete[i])
1217 targetShape.push_back(inpShape[i]);
1221 targetShape.push_back(1);
1225 if (targetShape.empty())
1226 targetShape.push_back(1);
1228 // Using PermuteLayer to move the deleted axis to the last.
1229 std::vector<int> perm(inpShape.size(), 0);
1230 for (int i = 0; i < inpShape.size(); i++)
1233 bool needPermuet = false;
1234 for (int i = 0; i < inpShape.size(); i++)
1236 if (shouldDelete[i])
1238 // find the first not deleted element.
1239 std::vector<bool>::iterator iter = std::find(shouldDelete.begin() + i, shouldDelete.end(), false);
1241 if (iter != shouldDelete.end())
1243 int index = iter - shouldDelete.begin();
1245 bool temp = shouldDelete[index];
1246 shouldDelete[index] = shouldDelete[i];
1247 shouldDelete[i] = temp;
1249 std::swap(perm[index], perm[i]);
1250 std::swap(inpShape[index], inpShape[i]);
1258 auto inputString= node_proto.input(0);
1261 LayerParams permuteLp;
1262 permuteLp.name = layerParams.name + "/permute";
1263 permuteLp.type = (depth == CV_8S) ? "PermuteInt8" : "Permute";
1264 permuteLp.set("order", DictValue::arrayInt(perm.data(), perm.size()));
1266 opencv_onnx::NodeProto protoPermute;
1267 protoPermute.add_input(inputString);
1268 protoPermute.add_output(permuteLp.name);
1269 addLayer(permuteLp, protoPermute);
1270 inputString = permuteLp.name;
1273 std::vector<int> deletedDims;
1274 for (int axis_i = 0; axis_i < inpShape.size(); ++axis_i)
1276 if (shouldDelete[axis_i])
1278 deletedDims.push_back(inpShape[axis_i]);
1282 LayerParams reduceLp = layerParams;
1283 reduceLp.name = layerParams.name + "/reduce";
1284 CV_Assert(layer_id.find(reduceLp.name) == layer_id.end());
1285 reduceLp.set("deleted_dims", DictValue::arrayInt(&deletedDims[0], deletedDims.size()));
1287 node_proto.set_input(0, inputString);
1288 node_proto.set_output(0, reduceLp.name);
1289 addLayer(reduceLp, node_proto);
1291 layerParams.type = (depth == CV_8S) ? "ReshapeInt8" : "Reshape";
1292 layerParams.set("dim", DictValue::arrayInt(&targetShape[0], targetShape.size()));
1294 node_proto.set_input(0, node_proto.output(0));
1295 node_proto.set_output(0, output_name);
1297 addLayer(layerParams, node_proto);
1300 void ONNXImporter::parseSlice(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1303 std::vector<int> begin;
1304 std::vector<int> end;
1305 std::vector<int> steps;
1306 int inp_size = node_proto.input_size();
1310 if (layerParams.has("axes")) {
1311 DictValue axes = layerParams.get("axes");
1312 for (int i = 1; i < axes.size(); ++i) {
1313 CV_Assert(axes.get<int>(i - 1) == axes.get<int>(i) - 1);
1315 axis = axes.get<int>(0);
1318 DictValue starts = layerParams.get("starts");
1319 DictValue ends = layerParams.get("ends");
1320 CV_Assert(starts.size() == ends.size());
1323 CV_CheckLE(axis, 1024, "Slice layer can't have more than 1024 axes"); // arbitrary limit
1324 begin.resize(axis, 0);
1325 end.resize(axis, INT_MAX);
1327 for (int i = 0; i < starts.size(); ++i)
1329 begin.push_back(starts.get<int>(i));
1330 end.push_back(ends.get<int>(i));
1332 } else { // inp_size > 1
1333 CV_Assert(inp_size >= 3);
1334 for (int i = 1; i < inp_size; i++) {
1335 CV_Assert(constBlobs.find(node_proto.input(i)) != constBlobs.end());
1337 Mat start_blob = getBlob(node_proto, 1);
1338 Mat end_blob = getBlob(node_proto, 2);
1339 CV_Assert(start_blob.total() == end_blob.total());
1342 Mat axes_blob = getBlob(node_proto, 3);
1343 const int* axes = (int*)axes_blob.data;
1344 for (int i = 1; i < axes_blob.total(); ++i) {
1345 CV_Assert(axes[i - 1] == axes[i] - 1);
1350 const int* starts = start_blob.ptr<int>();
1351 const int* ends = end_blob.ptr<int>();
1353 begin.resize(axis, 0);
1354 end.resize(axis, INT_MAX);
1356 std::copy(starts, starts + start_blob.total(), std::back_inserter(begin));
1357 std::copy(ends, ends + end_blob.total(), std::back_inserter(end));
1359 if (inp_size == 5) {
1360 CV_Assert(constBlobs.find(node_proto.input(4)) != constBlobs.end());
1361 Mat step_blob = getBlob(node_proto, 4);
1362 const int* steps_ptr = step_blob.ptr<int>();
1365 steps.resize(axis, 1);
1367 std::copy(steps_ptr, steps_ptr + step_blob.total(), std::back_inserter(steps));
1369 // Very strange application for Slice op with tensor reversing.
1370 // We just workaround it for 2d constants.
1371 if (constBlobs.find(node_proto.input(0)) != constBlobs.end() &&
1373 start_blob.at<int>(0) == -1 && step_blob.at<int>(0) == -1 &&
1374 end_blob.at<int>(0) == std::numeric_limits<int32_t>::min())
1376 Mat inp = getBlob(node_proto, 0);
1380 flip(inp, flipped, 0);
1381 addConstant(node_proto.output(0), flipped);
1387 layerParams.set("begin", DictValue::arrayInt(&begin[0], begin.size()));
1388 layerParams.set("end", DictValue::arrayInt(&end[0], end.size()));
1389 layerParams.set("axis", axis);
1392 layerParams.set("steps", DictValue::arrayInt(&steps[0], steps.size()));
1394 if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
1396 Mat inp = getBlob(node_proto, 0);
1397 std::vector<Mat> inputs, sliced;
1398 inputs.push_back(inp);
1399 runLayer(layerParams, inputs, sliced);
1400 CV_Assert(sliced.size() == 1);
1401 addConstant(node_proto.output(0), sliced[0]);
1404 addLayer(layerParams, node_proto);
1407 void ONNXImporter::parseSplit(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1409 if (layerParams.has("split"))
1411 DictValue splits = layerParams.get("split");
1412 const int numSplits = splits.size();
1413 CV_Assert(numSplits > 1);
1415 std::vector<int> slicePoints(numSplits - 1, splits.get<int>(0));
1416 for (int i = 1; i < splits.size() - 1; ++i)
1418 slicePoints[i] = slicePoints[i - 1] + splits.get<int>(i);
1420 layerParams.set("slice_point", DictValue::arrayInt(&slicePoints[0], slicePoints.size()));
1424 layerParams.set("num_split", node_proto.output_size());
1426 int depth = layerParams.get<int>("depth", CV_32F);
1427 layerParams.type = (depth == CV_8S) ? "SliceInt8" : "Slice";
1428 layerParams.set("axis", layerParams.get<float>("axis", 0));
1429 addLayer(layerParams, node_proto);
1432 void ONNXImporter::parseBias(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
1434 opencv_onnx::NodeProto node_proto = node_proto_;
1435 const std::string& layer_type = node_proto.op_type();
1436 bool isSub = layer_type == "Sub";
1438 if (layer_type == "Sum" && node_proto.input_size() == 1)
1440 layerParams.type = "Identity";
1441 addLayer(layerParams, node_proto);
1445 CV_Assert((node_proto.input_size() == 2) || (layer_type == "Sum" && node_proto.input_size() > 2));
1447 if (layer_type == "Sum" && node_proto.input_size() > 2)
1449 for (int i = 0; i < node_proto.input_size(); ++i)
1451 if (layer_id.find(node_proto.input(i)) == layer_id.end())
1453 CV_Error(Error::StsNotImplemented, "Sum of constants is not implemented for inputs > 2");
1458 bool is_const_0 = layer_id.find(node_proto.input(0)) == layer_id.end();
1459 bool is_const_1 = layer_id.find(node_proto.input(1)) == layer_id.end();
1460 if (is_const_0 && is_const_1)
1462 Mat blob_0 = getBlob(node_proto, 0);
1463 Mat blob_1 = getBlob(node_proto, 1);
1464 CV_Assert(blob_0.size == blob_1.size);
1465 Mat output = isSub ? (blob_0 - blob_1) : (blob_0 + blob_1);
1466 addConstant(node_proto.output(0), output);
1469 else if (is_const_0 || is_const_1)
1471 int const_blob_id = is_const_0 ? 0 : 1;
1472 int input_id = 1 - const_blob_id;
1473 Mat blob = getBlob(node_proto, const_blob_id);
1474 int blob_total = blob.total();
1476 const float inputScale = isSub && is_const_0 ? -1.f : 1.f;
1477 const float constScale = isSub && is_const_1 ? -1.f : 1.f;
1479 if (blob_total == 1) {
1480 layerParams.type = "Power";
1481 layerParams.set("scale", inputScale);
1482 layerParams.set("shift", constScale * blob.ptr<float>()[0]);
1485 MatShape inpShape = outShapes[node_proto.input(input_id)];
1486 if (shape(blob) == inpShape)
1488 LayerParams constParams;
1489 constParams.name = layerParams.name + "/const";
1490 constParams.type = "Const";
1491 constParams.blobs.push_back(blob);
1492 int id = dstNet.addLayer(constParams.name, constParams.type, constParams);
1493 layer_id.insert(std::make_pair(constParams.name, LayerInfo(id, 0)));
1494 outShapes[constParams.name] = shape(blob);
1496 layerParams.type = "Eltwise";
1497 float coeffs[] = {1., isSub ? -1.f : 1.f};
1498 layerParams.set("coeff", DictValue::arrayReal<float*>(coeffs, 2));
1499 node_proto.set_input(const_blob_id, constParams.name);
1503 if (inputScale < 0.f)
1505 addNegation(layerParams, node_proto, input_id);
1508 layerParams.type = "Scale";
1509 layerParams.set("bias_term", true);
1511 for (int i = 0; i < graph_proto.initializer_size(); i++)
1513 opencv_onnx::TensorProto tensor_proto = graph_proto.initializer(i);
1514 if (tensor_proto.name() == node_proto.input(const_blob_id))
1516 axis = inpShape.size() - tensor_proto.dims_size();
1520 layerParams.set("axis", axis);
1521 blob = blob.reshape(1, 1);
1522 layerParams.blobs.push_back(constScale * blob);
1526 else if (outShapes[node_proto.input(0)] == outShapes[node_proto.input(1)])
1528 layerParams.type = "Eltwise";
1531 static float subCoeffs[] = {1.f, -1.f};
1532 layerParams.set("coeff", DictValue::arrayReal<float*>(subCoeffs, 2));
1539 addNegation(layerParams, node_proto, 1);
1541 layerParams.type = "Scale";
1542 layerParams.set("bias_term", true);
1544 addLayer(layerParams, node_proto);
1547 void ONNXImporter::parsePow(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1549 if (layer_id.find(node_proto.input(1)) != layer_id.end())
1550 CV_Error(Error::StsNotImplemented, "Unsupported Pow op with variable power");
1552 Mat blob = getBlob(node_proto, 1);
1553 if (blob.total() != 1)
1554 CV_Error(Error::StsNotImplemented, "Pow op supports only scalar power");
1556 blob.convertTo(blob, CV_32F);
1557 layerParams.type = "Power";
1558 layerParams.set("power", blob.ptr<float>()[0]);
1559 addLayer(layerParams, node_proto);
1563 void ONNXImporter::parseMinMax(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1565 const std::string& layer_type = node_proto.op_type();
1566 layerParams.type = "Eltwise";
1567 layerParams.set("operation", layer_type == "Max" ? "max" : "min");
1568 addLayer(layerParams, node_proto);
1571 void ONNXImporter::parseNeg(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1573 layerParams.type = "Power";
1574 layerParams.set("scale", -1);
1575 addLayer(layerParams, node_proto);
1578 void ONNXImporter::parseConstant(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1580 CV_Assert(node_proto.input_size() == 0);
1581 CV_Assert(layerParams.blobs.size() == 1);
1582 addConstant(node_proto.output(0), layerParams.blobs[0]);
1585 void transformBlobs(std::vector<Mat>& blobs)
1590 std::vector<Mat> cudaWorkaround;
1591 cudaWorkaround.push_back(Wx.clone());
1592 cudaWorkaround.push_back(Wh.clone());
1593 cudaWorkaround.push_back(b.clone());
1595 const int numHidden = Wh.size[2];
1598 h0 = h0.reshape(1, h0.size[0] * h0.size[1]);
1600 c0 = c0.reshape(1, c0.size[0] * c0.size[1]);
1602 b = b.reshape(1, b.size[0]);
1603 Mat bx = b.colRange(0, b.cols / 2);
1604 Mat bh = b.colRange(b.cols / 2, b.cols);
1607 auto toIFOC = [] (Mat& in) {
1608 int first = in.size[0];
1609 int rest = in.total() / first / 4;
1610 // every weight blob contains weights for Input, Output, Forget and Cell gates
1611 Mat m = in.reshape(1, {first, 4, rest});
1612 Mat outputGate = m.col(1);
1613 Mat forgetGate = m.col(2);
1614 std::swap_ranges(outputGate.begin<float>(), outputGate.end<float>(), forgetGate.begin<float>());
1621 Wx = Wx.reshape(1, Wx.size[0] * Wx.size[1]);
1622 Wh = Wh.reshape(1, Wh.size[0] * Wh.size[1]);
1626 blobs[2] = b.reshape(1, 1);
1630 if (blobs.size() == 5) {
1631 // so that future patch removing copies can leave all indexing as is
1632 blobs.insert(blobs.begin(), cudaWorkaround.begin(), cudaWorkaround.end());
1637 blobs[5] = P.colRange(0, numHidden);
1638 blobs[5] = blobs[5].clone().reshape(1, blobs[5].total()); // Single column.
1639 blobs[5] = Mat::diag(blobs[5]);
1641 blobs.push_back(P.colRange(numHidden, 2 * numHidden));
1642 blobs[6] = blobs[6].clone().reshape(1, blobs[6].total()); // Single column.
1643 blobs[6] = Mat::diag(blobs[6]);
1645 blobs.push_back(P.colRange(2 * numHidden, 3 * numHidden));
1646 blobs[7] = blobs[7].clone().reshape(1, blobs[7].total()); // Single column.
1647 blobs[7] = Mat::diag(blobs[7]);
1649 // so that future patch removing copies can leave all indexing as is
1650 blobs.insert(blobs.begin(), cudaWorkaround.begin(), cudaWorkaround.end());
1653 void ONNXImporter::lstm_extractConsts(LayerParams& layerParams, const opencv_onnx::NodeProto& lstm_proto, size_t idx, int* blobShape_, int size)
1655 MatShape blobShape(blobShape_, blobShape_ + size);
1657 if (idx < lstm_proto.input_size() && !lstm_proto.input(idx).empty())
1659 blob = getBlob(lstm_proto, idx);
1660 CV_Assert(shape(blob) == blobShape);
1664 blob = Mat(blobShape, CV_32FC1, 0.);
1666 layerParams.blobs.push_back(blob);
1669 void ONNXImporter::lstm_add_reshape(const std::string& input_name, const std::string& output_name, int* layerShape, size_t n)
1671 LayerParams reshapeLp;
1672 reshapeLp.name = cv::format("%s/reshape", input_name.c_str());
1673 reshapeLp.type = "Reshape";
1674 CV_Assert(layer_id.find(reshapeLp.name) == layer_id.end());
1676 reshapeLp.set("dim", DictValue::arrayInt(layerShape, n));
1678 opencv_onnx::NodeProto reshape_proto;
1679 reshape_proto.add_input(input_name);
1680 reshape_proto.add_output(output_name);
1681 addLayer(reshapeLp, reshape_proto);
1684 std::string ONNXImporter::lstm_add_slice(int index, const std::string& input_name, int* begin, int* end, size_t n)
1686 LayerParams sliceLP;
1687 sliceLP.name = cv::format("%s/slice_%d", input_name.c_str(), index);
1688 sliceLP.type = "Slice";
1689 CV_Assert(layer_id.find(sliceLP.name) == layer_id.end());
1691 sliceLP.set("begin", DictValue::arrayInt(begin, n));
1692 sliceLP.set("end", DictValue::arrayInt(end, n));
1693 sliceLP.set("axis", 0);
1695 opencv_onnx::NodeProto slice_proto;
1696 slice_proto.add_input(input_name);
1697 slice_proto.add_output(sliceLP.name);
1698 addLayer(sliceLP, slice_proto);
1700 return slice_proto.output(0);
1703 std::string ONNXImporter::lstm_fix_dims(LayerParams& layerParams, const opencv_onnx::NodeProto& lstm_proto,
1704 int batch_size, int num_directions, int hidden_size, bool need_y, const std::string& y_name,
1707 std::string reshape_output = cv::format("%s/reshape_%d", layerParams.name.c_str(), index);
1709 // reshape from Seq, Batch, Dirs*Hidden to Seq, Batch, Dirs, Hidden
1710 // to not confuse reshape with dynamic first dimension, zero means 'leave unchanged'
1711 int layerShape[] = {0, batch_size, num_directions, hidden_size};
1712 lstm_add_reshape(lstm_proto.output(index), reshape_output, layerShape, sizeof(layerShape) / sizeof(layerShape[0]));
1714 // permute from Seq, Batch, Dirs, Hidden to Seq, Dirs, Batch, Hidden
1715 LayerParams permuteLP;
1716 permuteLP.name = reshape_output + "/permute";
1717 permuteLP.type = "Permute";
1718 CV_Assert(layer_id.find(permuteLP.name) == layer_id.end());
1720 int order[] = {0, 2, 1, 3};
1721 permuteLP.set("order", DictValue::arrayInt(order, 4));
1723 opencv_onnx::NodeProto permute_proto;
1724 permute_proto.add_input(reshape_output);
1725 permute_proto.add_output((need_y && index == 0) ? y_name : static_cast<std::string>(permuteLP.name));
1726 addLayer(permuteLP, permute_proto);
1728 return permute_proto.output(0);
1731 void ONNXImporter::lstm_add_transform(int num_directions, int batch_size, int hidden_size,
1732 int index, const std::string& input_name, const std::string& output_name)
1734 if (num_directions == 1)
1736 // Slice: Yh = Y[-1, :, :, :]
1737 int begin[] = {-1}, end[] = {INT_MAX};
1738 std::string slice_output = lstm_add_slice(index, input_name, begin, end, sizeof(begin) / sizeof(begin[0]));
1740 // Reshape: 1x1xBxH -> 1xBxH
1741 int layerShape[] = {1, batch_size, hidden_size};
1742 lstm_add_reshape(slice_output, output_name, layerShape, sizeof(layerShape) / sizeof(layerShape[0]));
1746 // Slice: SxDxBxH -> last sequence, first direction
1747 int begin0[] = {-1, 0}, end0[] = {INT_MAX, 1};
1748 std::string slice_0 = lstm_add_slice(0, input_name, begin0, end0, sizeof(begin0) / sizeof(begin0[0]));
1750 // Slice: SxDxBxH -> first sequence, last direction
1751 int begin1[] = {0, -1}, end1[] = {1, INT_MAX};
1752 std::string slice_1 = lstm_add_slice(1, input_name, begin1, end1, sizeof(begin1) / sizeof(begin1[0]));
1754 LayerParams concatLP;
1755 concatLP.name = cv::format("%s/concat", input_name.c_str());
1756 concatLP.type = "Concat";
1757 CV_Assert(layer_id.find(concatLP.name) == layer_id.end());
1759 concatLP.set("axis", 1); // 1x1xBxH -> 1x2xBxH
1761 opencv_onnx::NodeProto concat_proto;
1762 concat_proto.add_input(slice_0);
1763 concat_proto.add_input(slice_1);
1764 concat_proto.add_output(concatLP.name);
1765 addLayer(concatLP, concat_proto);
1767 // Reshape: 1x2xBxH -> 2xBxH
1768 int layerShape[] = {2, batch_size, hidden_size};
1769 lstm_add_reshape(concat_proto.output(0), output_name, layerShape, sizeof(layerShape) / sizeof(layerShape[0]));
1773 void ONNXImporter::parseLSTM(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
1775 opencv_onnx::NodeProto lstm_proto = node_proto_;
1776 layerParams.name += "/lstm";
1778 // https://github.com/onnx/onnx/blob/main/docs/Operators.md#LSTM
1779 CV_Assert(lstm_proto.input_size() >= 3);
1780 for (size_t i = 1; i < 3; ++i)
1782 const std::string& name = lstm_proto.input(i);
1783 CV_Assert(!name.empty() && constBlobs.count(name) == 1);
1786 IterShape_t shapeIt = outShapes.find(lstm_proto.input(0));
1787 CV_Assert(shapeIt != outShapes.end());
1788 const MatShape x_shape = shapeIt->second;
1790 const int seq_length = x_shape[0];
1791 const int batch_size = x_shape[1];
1792 const int input_size = x_shape[2];
1793 const int hidden_size = layerParams.get<int>("hidden_size");
1794 const int num_directions = constBlobs[lstm_proto.input(1)].size[0];
1796 int w_size[] = {num_directions, 4*hidden_size, input_size};
1797 lstm_extractConsts(layerParams, lstm_proto, 1, w_size, sizeof(w_size) / sizeof(w_size[0])); // W
1799 int r_size[] = {num_directions, 4*hidden_size, hidden_size};
1800 lstm_extractConsts(layerParams, lstm_proto, 2, r_size, sizeof(r_size) / sizeof(r_size[0])); // R
1802 int b_size[] = {num_directions, 8*hidden_size};
1803 lstm_extractConsts(layerParams, lstm_proto, 3, b_size, sizeof(b_size) / sizeof(b_size[0])); // B
1805 if (4 < lstm_proto.input_size() && !lstm_proto.input(4).empty())
1807 Mat blob = getBlob(lstm_proto, 4);
1808 CV_Assert(blob.total() == batch_size);
1809 for (MatIterator_<int32_t> it = blob.begin<int32_t>(); it != blob.end<int32_t>(); ++it)
1811 CV_Assert(*it == seq_length);
1815 int h_size[] = {num_directions, batch_size, hidden_size};
1816 lstm_extractConsts(layerParams, lstm_proto, 5, h_size, sizeof(h_size) / sizeof(h_size[0])); // initial_h
1818 int c_size[] = {num_directions, batch_size, hidden_size};
1819 lstm_extractConsts(layerParams, lstm_proto, 6, c_size, sizeof(c_size) / sizeof(c_size[0])); // initial_c
1821 if (lstm_proto.input_size() > 7 && !lstm_proto.input(7).empty())
1823 layerParams.set("use_peephole", true);
1824 int p_size[] = {num_directions, 3 * hidden_size};
1825 lstm_extractConsts(layerParams, lstm_proto, 7, p_size, sizeof(p_size) / sizeof(p_size[0])); // P
1828 transformBlobs(layerParams.blobs);
1830 layerParams.set("is_onnx", true);
1831 layerParams.set("reverse", layerParams.get<String>("direction", "") == "reverse");
1832 layerParams.set("bidirectional", layerParams.get<String>("direction", "") == "bidirectional");
1834 bool need_yc = lstm_proto.output_size() > 2 && !lstm_proto.output(2).empty();
1835 bool need_yh = lstm_proto.output_size() > 1 && !lstm_proto.output(1).empty();
1836 bool need_y = lstm_proto.output_size() > 0 && !lstm_proto.output(0).empty();
1838 const std::string y_name = need_y ? lstm_proto.output(0) : "";
1839 const std::string yh_name = need_yh ? lstm_proto.output(1) : "";
1840 const std::string yc_name = need_yc ? lstm_proto.output(2) : "";
1842 layerParams.set("produce_cell_output", need_yc);
1844 lstm_proto.clear_output();
1845 if (need_y || need_yh)
1847 // give random names to LSTMLayer's outputs because every output needs postprocessing
1848 lstm_proto.add_output(cv::format("%s_y", layerParams.name.c_str()));
1852 lstm_proto.add_output(yc_name);
1855 addLayer(layerParams, lstm_proto);
1857 std::string y_output = lstm_fix_dims(layerParams, lstm_proto, batch_size, num_directions, hidden_size, need_y,
1861 lstm_add_transform(num_directions, batch_size, hidden_size, 0, y_output, yh_name);
1865 void ONNXImporter::parseGRU(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
1867 opencv_onnx::NodeProto node_proto = node_proto_;
1868 const std::string output_name = node_proto.output(0);
1869 LayerParams gruParams = layerParams;
1870 gruParams.name += "/gru";
1872 // https://pytorch.org/docs/stable/generated/torch.nn.GRU.html?highlight=gru#
1873 CV_Assert(node_proto.input_size() == 6);
1874 Mat Wx = getBlob(node_proto, 1);
1875 Mat Wh = getBlob(node_proto, 2);
1876 Mat b = getBlob(node_proto, 3);
1877 Mat h0 = getBlob(node_proto, 5);
1879 Wx = Wx.reshape(1, Wx.size[0] * Wx.size[1]);
1880 Wh = Wh.reshape(1, Wh.size[0] * Wh.size[1]);
1881 h0 = h0.reshape(1, h0.size[0] * h0.size[1]);
1882 b = b.reshape(1, b.size[0]);
1884 gruParams.blobs.resize(4);
1885 gruParams.blobs[0] = Wh;
1886 gruParams.blobs[1] = Wx;
1887 gruParams.blobs[2] = b;
1888 gruParams.blobs[3] = h0;
1889 gruParams.set("bidirectional", gruParams.get<String>("direction", "") == "bidirectional");
1891 node_proto.set_output(0, gruParams.name); // set different name so output shapes will be registered on that name
1892 addLayer(gruParams, node_proto);
1894 MatShape gruShape = outShapes[node_proto.output(0)];
1896 // Add fake 1 as it is done in ONNX
1897 gruShape.insert(gruShape.begin() + 1, 1);
1899 layerParams.type = "Reshape";
1900 layerParams.set("dim", DictValue::arrayInt(&gruShape[0], gruShape.size()));
1901 node_proto.set_input(0, gruParams.name); // redirect input to GRU
1902 node_proto.set_output(0, output_name); // keep origin GRU's name
1903 addLayer(layerParams, node_proto);
1906 void ONNXImporter::parseImageScaler(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1908 const float scale = layerParams.has("scale") ? layerParams.get<float>("scale") : 1.0f;
1909 layerParams.erase("scale");
1911 if (layerParams.has("bias"))
1913 layerParams.type = "Scale";
1914 layerParams.blobs.push_back(
1915 Mat(Size(1, layerParams.get("bias").size()), CV_32FC1, scale));
1917 layerParams.set("bias_term", true);
1918 Mat bias(1, layerParams.get("bias").size(), CV_32FC1);
1919 for (int j = 0; j < bias.total(); j++) {
1920 bias.at<float>(0, j) = layerParams.get("bias").getRealValue(j);
1922 layerParams.blobs.push_back(bias);
1923 layerParams.erase("bias");
1926 layerParams.set("scale", scale);
1927 layerParams.type = "Power";
1929 addLayer(layerParams, node_proto);
1932 void ONNXImporter::parseClip(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1934 layerParams.type = "ReLU6";
1935 float min_value = -FLT_MAX, max_value = FLT_MAX;
1936 int input_size = node_proto.input_size();
1937 CV_Check(input_size, 1 <= input_size && input_size <= 3, "");
1939 if (input_size >= 2 && !node_proto.input(1).empty())
1941 if (constBlobs.find(node_proto.input(1)) != constBlobs.end())
1942 min_value = getBlob(node_proto, 1).at<float>(0);
1944 CV_Error(Error::StsNotImplemented, "Non-constant min values in Clip are not supported");
1947 if (input_size == 3 && !node_proto.input(2).empty())
1949 if (constBlobs.find(node_proto.input(2)) != constBlobs.end())
1950 max_value = getBlob(node_proto, 2).at<float>(0);
1952 CV_Error(Error::StsNotImplemented, "Non-constant max values in Clip are not supported");
1955 layerParams.set("min_value", layerParams.get<float>("min", min_value));
1956 layerParams.set("max_value", layerParams.get<float>("max", max_value));
1957 addLayer(layerParams, node_proto);
1960 void ONNXImporter::parseLeakyRelu(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1962 layerParams.type = "ReLU";
1963 layerParams.set("negative_slope", layerParams.get<float>("alpha", 0.01));
1964 addLayer(layerParams, node_proto);
1967 void ONNXImporter::parseRelu(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1969 layerParams.type = "ReLU";
1970 addLayer(layerParams, node_proto);
1973 void ONNXImporter::parseElu(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1975 layerParams.type = "ELU";
1976 addLayer(layerParams, node_proto);
1979 void ONNXImporter::parseTanh(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1981 layerParams.type = "TanH";
1982 addLayer(layerParams, node_proto);
1985 void ONNXImporter::parseAbs(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1987 layerParams.type = "AbsVal";
1988 addLayer(layerParams, node_proto);
1991 void ONNXImporter::parseCompare(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
1993 CV_Assert(node_proto.input_size() == 2);
1994 const std::string& layer_type = node_proto.op_type();
1996 bool is_const_0 = layer_id.find(node_proto.input(0)) == layer_id.end();
1997 bool is_const_1 = layer_id.find(node_proto.input(1)) == layer_id.end();
1999 if (is_const_0 || is_const_1)
2001 Mat blob = getBlob(node_proto, static_cast<int>(is_const_1));
2002 blob = blob.reshape(1, 1);
2003 layerParams.blobs.push_back(blob);
2006 layerParams.type = "Compare";
2008 if (layer_type == "Equal")
2009 layerParams.set("mode", "equal");
2010 else if (layer_type == "Greater")
2011 layerParams.set("mode", "greater");
2013 layerParams.set("mode", "less");
2014 addLayer(layerParams, node_proto);
2017 void ONNXImporter::parsePRelu(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2019 layerParams.type = "PReLU";
2020 layerParams.blobs.push_back(getBlob(node_proto, 1));
2021 addLayer(layerParams, node_proto);
2024 void ONNXImporter::parseLRN(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2026 replaceLayerParam(layerParams, "size", "local_size");
2027 addLayer(layerParams, node_proto);
2030 void ONNXImporter::parseInstanceNormalization(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
2032 opencv_onnx::NodeProto node_proto = node_proto_;
2033 if (node_proto.input_size() != 3)
2034 CV_Error(Error::StsNotImplemented,
2035 "Expected input, scale, bias");
2037 layerParams.blobs.resize(4);
2038 layerParams.blobs[2] = getBlob(node_proto, 1); // weightData
2039 layerParams.blobs[3] = getBlob(node_proto, 2); // biasData
2040 layerParams.set("has_bias", true);
2041 layerParams.set("has_weight", true);
2043 // Get number of channels in input
2044 int size = layerParams.blobs[2].total();
2045 layerParams.blobs[0] = Mat::zeros(size, 1, CV_32F); // mean
2046 layerParams.blobs[1] = Mat::ones(size, 1, CV_32F); // std
2048 LayerParams mvnParams;
2049 mvnParams.name = layerParams.name + "/MVN";
2050 mvnParams.type = "MVN";
2051 mvnParams.set("eps", layerParams.get<float>("epsilon"));
2052 layerParams.erase("epsilon");
2055 int id = dstNet.addLayer(mvnParams.name, mvnParams.type, mvnParams);
2057 IterLayerId_t layerId = layer_id.find(node_proto.input(0));
2058 CV_Assert(layerId != layer_id.end());
2059 dstNet.connect(layerId->second.layerId, layerId->second.outputId, id, 0);
2061 layer_id.insert(std::make_pair(mvnParams.name, LayerInfo(id, 0)));
2062 outShapes[mvnParams.name] = outShapes[node_proto.input(0)];
2064 //Replace Batch Norm's input to MVN
2065 node_proto.set_input(0, mvnParams.name);
2066 layerParams.type = "BatchNorm";
2067 addLayer(layerParams, node_proto);
2070 void ONNXImporter::parseBatchNormalization(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2072 if (node_proto.input_size() != 5)
2073 CV_Error(Error::StsNotImplemented,
2074 "Expected input, scale, bias, mean and var");
2076 layerParams.type = "BatchNorm";
2077 replaceLayerParam(layerParams, "epsilon", "eps");
2078 replaceLayerParam(layerParams, "spatial", "use_global_stats");
2080 Mat meanData = getBlob(node_proto, 3);
2081 Mat stdData = getBlob(node_proto, 4);
2083 layerParams.blobs.push_back(meanData);
2084 layerParams.blobs.push_back(stdData);
2086 if (!node_proto.input(1).empty()) {
2087 layerParams.set("has_weight", true);
2088 layerParams.blobs.push_back(getBlob(node_proto, 1)); // weightData
2090 layerParams.set("has_weight", false);
2093 if (!node_proto.input(2).empty()) {
2094 layerParams.set("has_bias", true);
2095 layerParams.blobs.push_back(getBlob(node_proto, 2)); // biasData
2097 layerParams.set("has_bias", false);
2099 addLayer(layerParams, node_proto);
2102 // A * B + C = Y, we require that the dimension of A is [m, k], and the dimension of B is [n, k].
2103 // And the dim of output Y is [m, n]
2104 void ONNXImporter::parseGemm(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2106 CV_Assert(node_proto.input_size() >= 2);
2107 layerParams.type = "InnerProduct";
2108 Mat weights = getBlob(node_proto, 1);
2110 if (!layerParams.get<int>("transB", 0))
2112 transpose(weights, weights);
2114 layerParams.blobs.push_back(weights);
2116 if (node_proto.input_size() == 3) {
2117 Mat bias = getBlob(node_proto, 2);
2118 layerParams.blobs.push_back(bias);
2120 if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
2122 Mat inputBuf = getBlob(node_proto, 0);
2124 LayerParams constParams;
2125 constParams.name = node_proto.input(0);
2126 constParams.type = "Const";
2127 constParams.blobs.push_back(inputBuf);
2129 opencv_onnx::NodeProto proto;
2130 proto.add_output(constParams.name);
2131 addLayer(constParams, proto);
2134 layerParams.set("num_output", layerParams.blobs[0].size[0]);
2135 layerParams.set("bias_term", node_proto.input_size() == 3);
2136 addLayer(layerParams, node_proto);
2139 void ONNXImporter::parseMatMul(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2141 CV_Assert(node_proto.input_size() == 2);
2142 layerParams.type = "InnerProduct";
2143 layerParams.set("bias_term", false);
2144 CV_Assert(constBlobs.find(node_proto.input(0)) == constBlobs.end());
2145 int firstInpDims = outShapes[node_proto.input(0)].size();
2148 if (constBlobs.find(node_proto.input(1)) != constBlobs.end())
2150 Mat blob = getBlob(node_proto, 1);
2151 secondInpDims = blob.dims;
2152 layerParams.blobs.push_back(blob.t());
2153 layerParams.set("num_output", layerParams.blobs[0].size[0]);
2155 secondInpDims = outShapes[node_proto.input(1)].size();
2157 layerParams.set("axis", firstInpDims - secondInpDims + 1);
2158 addLayer(layerParams, node_proto);
2161 void findBroadAxis(const MatShape& broadShape, const MatShape& outShape, size_t& axis, int& broadAxis)
2163 const size_t diff = outShape.size() - broadShape.size();
2165 // find the first non-one element of the broadcasting shape
2167 for (; axis < broadShape.size() && broadShape[axis] == 1; ++axis) {}
2169 // find the last non-one element of the broadcasting shape
2170 size_t endAxis = broadShape.size();
2171 for (; endAxis > axis && broadShape[endAxis - 1] == 1; --endAxis) {}
2173 // find one between axis and endAxis - as it needs to be broadcasted,
2174 // dimensions from the left of axis and from the right of endAxis will be handled by Scale layer
2176 for (size_t i = axis; i < endAxis; ++i)
2178 size_t outAxis = i + diff;
2179 if (outShape[outAxis] == broadShape[i])
2184 // ensure we need to broadcast only 1 dimension in the middle
2185 CV_Assert(broadShape[i] == 1 && broadAxis == -1);
2186 broadAxis = static_cast<int>(outAxis);
2193 void ONNXImporter::parseMul(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
2195 opencv_onnx::NodeProto node_proto = node_proto_;
2196 const std::string& layer_type = node_proto.op_type();
2197 const std::string output_name = node_proto.output(0);
2198 CV_Assert(node_proto.input_size() == 2);
2200 bool isDiv = layer_type == "Div";
2202 bool haveVariables = false;
2203 for (int i = 0; i < 2; ++i)
2205 if (constBlobs.find(node_proto.input(i)) != constBlobs.end())
2208 haveVariables = true;
2210 if (constId != -1 && haveVariables)
2212 Mat blob = getBlob(node_proto, constId);
2213 blob = blob.reshape(1, 1);
2214 if (blob.total() == 1) {
2215 float blob_value = blob.ptr<float>()[0];
2216 float coeff = blob_value;
2219 coeff = 1.f / blob_value;
2222 // Power layer calculates (x*scale + shift)^power, so const/x -> (x * (1/const) + 0)^(-1)
2223 layerParams.set("power", -1.f);
2226 layerParams.set("scale", coeff);
2227 layerParams.type = "Power";
2231 divide(1.0, blob, blob);
2232 layerParams.blobs.push_back(blob);
2233 layerParams.type = "Scale";
2236 else if (!haveVariables)
2238 Mat inp0 = getBlob(node_proto, 0);
2239 Mat inp1 = getBlob(node_proto, 1);
2241 if (inp0.size != inp1.size && (inp0.total() != 1 || inp1.total() != 1))
2242 CV_Error_(Error::StsNotImplemented, ("Different shapes case is not supported with constant inputs: %s", layer_type.c_str()));
2244 if (inp0.total() == 1 && inp1.total() == 1 && inp0.dims != inp1.dims)
2246 if (inp0.dims < inp1.dims)
2248 inp0 = inp0.reshape(1, inp1.dims, inp1.size);
2249 inp0.dims = inp1.dims;
2253 inp1 = inp1.reshape(1, inp0.dims, inp0.size);
2254 inp1.dims = inp0.dims;
2259 if (inp0.total() != inp1.total())
2261 if (inp0.total() == 1)
2263 float inp0_value = inp0.ptr<float>()[0];
2264 float coeff = isDiv ? 1.0 / inp0_value : inp0_value;
2265 multiply(inp1, coeff, out);
2269 float inp1_value = inp1.ptr<float>()[0];
2270 float coeff = isDiv ? 1.0 / inp1_value : inp1_value;
2271 multiply(inp0, coeff, out);
2277 out = isDiv ? inp0 / inp1 : inp0.mul(inp1);
2280 if (inp0.dims == 1 && inp1.dims == 1)
2281 out.dims = 1; // to workaround dims == 1
2282 addConstant(output_name, out);
2285 else if (outShapes[node_proto.input(0)] == outShapes[node_proto.input(1)])
2287 layerParams.type = "Eltwise";
2288 layerParams.set("operation", isDiv ? "div" : "prod");
2292 // Scale layer allocate output with the first input shape
2293 if (total(outShapes[node_proto.input(0)]) < total(outShapes[node_proto.input(1)]))
2295 opencv_onnx::NodeProto proto;
2296 proto.add_input(node_proto.input(1));
2297 proto.add_input(node_proto.input(0));
2298 proto.add_output(output_name);
2304 LayerParams powerParams;
2305 powerParams.name = layerParams.name + "/inv";
2306 powerParams.type = "Power";
2307 powerParams.set("power", -1);
2309 //Create Power layer
2310 int id = dstNet.addLayer(powerParams.name, powerParams.type, powerParams);
2312 IterLayerId_t layerId = layer_id.find(node_proto.input(1));
2313 CV_Assert(layerId != layer_id.end());
2314 dstNet.connect(layerId->second.layerId, layerId->second.outputId, id, 0);
2316 layer_id.insert(std::make_pair(powerParams.name, LayerInfo(id, 0)));
2317 outShapes[powerParams.name] = outShapes[node_proto.input(1)];
2319 //Replace input to Power
2320 node_proto.set_input(1, powerParams.name);
2323 const MatShape& broadShape = outShapes[node_proto.input(1)];
2324 const MatShape& outShape = outShapes[node_proto.input(0)];
2328 findBroadAxis(broadShape, outShape, axis, broadAxis);
2330 // if there is a one dimension in the middle that should be broadcasted, broadcast it
2331 if (broadAxis != -1)
2333 opencv_onnx::NodeProto concat_node_proto = node_proto;
2334 const std::string& input1 = concat_node_proto.input(1);
2336 expandMid(layerParams.name, concat_node_proto, input1, outShape[broadAxis]);
2338 LayerParams concatLP;
2339 concatLP.name = layerParams.name + "/concat";
2340 concatLP.set("axis", broadAxis);
2341 concatLP.type = "Concat";
2342 concat_node_proto.set_output(0, concatLP.name);
2344 addLayer(concatLP, concat_node_proto);
2345 node_proto.set_input(1, concatLP.name);
2348 CV_Assert(axis != outShape.size());
2349 layerParams.set("axis", static_cast<int>(axis));
2350 layerParams.type = "Scale";
2352 addLayer(layerParams, node_proto);
2355 void ONNXImporter::parseConv(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
2357 opencv_onnx::NodeProto node_proto = node_proto_;
2358 CV_Assert(node_proto.input_size() >= 2);
2359 layerParams.type = "Convolution";
2360 for (int j = 1; j < node_proto.input_size(); j++) {
2361 if (constBlobs.find(node_proto.input(j)) != constBlobs.end())
2363 layerParams.blobs.push_back(getBlob(node_proto, j));
2366 int outCn = layerParams.blobs.empty() ? outShapes[node_proto.input(1)][0] : layerParams.blobs[0].size[0];
2367 layerParams.set("num_output", outCn);
2369 // Check for asymmetric padding in Conv2D
2370 if (layerParams.has("pad"))
2372 bool asymmetricPadding = false;
2373 DictValue pads = layerParams.get("pad");
2374 const int dims = pads.size() / 2;
2375 for (int i = 0; i < dims; ++i)
2377 if (pads.get<int>(i) != pads.get<int>(i + dims))
2379 asymmetricPadding = true;
2383 if (asymmetricPadding && pads.size() == 4) // [pad_t, pad_l, pad_b, pad_r]
2385 layerParams.erase("pad");
2386 // No paddings required for N, C axis
2387 std::vector<int> paddings(4, 0);
2388 // Add paddings for H, W axis
2389 for (int i = 0; i < dims; ++i)
2391 paddings.push_back(pads.get<int>(i));
2392 paddings.push_back(pads.get<int>(dims + i));
2395 padLp.name = layerParams.name + "/pad";
2396 padLp.type = "Padding";
2397 padLp.set("paddings", DictValue::arrayInt(&paddings[0], paddings.size()));
2399 opencv_onnx::NodeProto proto;
2400 proto.add_input(node_proto.input(0));
2401 proto.add_output(padLp.name);
2403 addLayer(padLp, proto);
2404 node_proto.set_input(0, padLp.name);
2407 addLayer(layerParams, node_proto);
2410 void ONNXImporter::parseConvTranspose(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2412 CV_Assert(node_proto.input_size() >= 2);
2413 layerParams.type = "Deconvolution";
2414 for (int j = 1; j < node_proto.input_size(); j++) {
2415 layerParams.blobs.push_back(getBlob(node_proto, j));
2417 layerParams.set("num_output", layerParams.blobs[0].size[1] * layerParams.get<int>("group", 1));
2418 layerParams.set("bias_term", node_proto.input_size() == 3);
2420 if (!layerParams.has("kernel_size"))
2421 CV_Error(Error::StsNotImplemented,
2422 "Required attribute 'kernel_size' is not present.");
2424 if (layerParams.has("output_shape"))
2426 const DictValue& outShape = layerParams.get("output_shape");
2427 DictValue strides = layerParams.get("stride");
2428 DictValue kernel = layerParams.get("kernel_size");
2431 std::vector<int> adjust_pads;
2432 if (layerParams.has("pad_mode"))
2434 padMode = toUpperCase(layerParams.get<String>("pad_mode"));
2435 if (padMode != "SAME" && padMode != "VALID")
2436 CV_Error(Error::StsError, "Unsupported padding mode " + padMode);
2438 for (int i = 0; i < strides.size(); i++)
2440 int sz = outShape.get<int>(2 + i);
2441 int stride = strides.get<int>(i);
2442 adjust_pads.push_back(padMode == "SAME"? (sz - 1) % stride :
2443 (sz - kernel.get<int>(i)) % stride);
2445 layerParams.set("adj", DictValue::arrayInt(&adjust_pads[0], adjust_pads.size()));
2448 else if (layerParams.has("output_padding"))
2450 replaceLayerParam(layerParams, "output_padding", "adj");
2452 addLayer(layerParams, node_proto);
2455 void ONNXImporter::parseTranspose(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2457 int depth = layerParams.get<int>("depth", CV_32F);
2458 layerParams.type = (depth == CV_8S) ? "PermuteInt8" : "Permute";
2459 replaceLayerParam(layerParams, "perm", "order");
2460 if (!layerParams.has("order")) {
2461 MatShape inpShape = outShapes[node_proto.input(0)];
2462 size_t dims = inpShape.size();
2463 std::vector<int> perm(dims);
2464 for (size_t d = 0; d < dims; ++d)
2466 perm[d] = static_cast<int>(dims - 1 - d);
2468 layerParams.set("order", DictValue::arrayInt(perm.data(), perm.size()));
2471 CV_Assert(node_proto.input_size() == 1);
2472 if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
2474 std::vector<Mat> inputs(1, getBlob(node_proto, 0)), transposed;
2475 runLayer(layerParams, inputs, transposed);
2476 CV_Assert(transposed.size() == 1);
2477 addConstant(node_proto.output(0), transposed[0]);
2480 addLayer(layerParams, node_proto);
2483 void ONNXImporter::parseSqueeze(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2485 CV_Assert_N(node_proto.input_size() == 1, layerParams.has("axes"));
2486 DictValue axes_dict = layerParams.get("axes");
2487 MatShape inpShape = outShapes[node_proto.input(0)];
2489 std::vector<bool> maskedAxes(inpShape.size(), false);
2490 for (int i = 0; i < axes_dict.size(); ++i)
2492 int axis = axes_dict.getIntValue(i);
2493 CV_CheckLE(axis, static_cast<int>(inpShape.size()), "Squeeze axis");
2494 maskedAxes[axis] = inpShape[axis] == 1;
2497 for (int i = 0; i < inpShape.size(); ++i)
2500 outShape.push_back(inpShape[i]);
2502 if (outShape.size() != inpShape.size())
2504 layerParams.type = "Reshape";
2505 layerParams.set("dim", DictValue::arrayInt(&outShape[0], outShape.size()));
2506 if (hasDynamicShapes)
2508 std::vector<int> dynamicAxes;
2509 std::vector<int> inputIndices;
2510 for (int index = 0; index < inpShape.size(); ++index)
2512 if (!maskedAxes[index])
2513 inputIndices.push_back(index);
2515 for (int index = 0; index < outShape.size(); ++index)
2516 dynamicAxes.push_back(index);
2517 layerParams.set("dynamic_axes", DictValue::arrayInt(dynamicAxes.data(), dynamicAxes.size()));
2518 layerParams.set("input_indices", DictValue::arrayInt(inputIndices.data(), inputIndices.size()));
2522 layerParams.type = "Identity";
2524 if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
2526 Mat inp = getBlob(node_proto, 0);
2527 Mat out = inp.reshape(1, outShape);
2528 out.dims = outShape.size(); // to workaround dims == 1
2529 addConstant(node_proto.output(0), out);
2532 int depth = layerParams.get<int>("depth", CV_32F);
2533 layerParams.type += (depth == CV_8S) ? "Int8" : "";
2534 addLayer(layerParams, node_proto);
2537 void ONNXImporter::parseFlatten(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
2539 opencv_onnx::NodeProto node_proto = node_proto_;
2540 CV_CheckEQ(node_proto.input_size(), 1, "");
2541 int axis_ = layerParams.get<int>("axis", 1);
2542 if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
2544 Mat input = getBlob(node_proto, 0);
2545 int axis = normalize_axis(axis_, input.dims);
2547 int out_size[2] = {1, 1};
2548 for (int i = 0; i < axis; ++i)
2550 out_size[0] *= input.size[i];
2552 for (int i = axis; i < input.dims; ++i)
2554 out_size[1] *= input.size[i];
2557 Mat output = input.reshape(1, 2, out_size);
2558 addConstant(node_proto.output(0), output);
2561 IterShape_t shapeIt = outShapes.find(node_proto.input(0));
2562 CV_Assert(shapeIt != outShapes.end());
2563 MatShape inpShape = shapeIt->second;
2564 int axis = normalize_axis(axis_, inpShape.size());
2566 if (axis == 0 || axis == inpShape.size())
2568 LayerParams reshapeLp;
2569 reshapeLp.name = layerParams.name + "/reshape";
2570 reshapeLp.type = "Reshape";
2571 CV_Assert(layer_id.find(reshapeLp.name) == layer_id.end());
2573 inpShape.insert(axis == 0 ? inpShape.begin() : inpShape.end(), 1);
2574 reshapeLp.set("dim", DictValue::arrayInt(&inpShape[0], inpShape.size()));
2576 opencv_onnx::NodeProto proto;
2577 proto.add_input(node_proto.input(0));
2578 proto.add_output(reshapeLp.name);
2579 addLayer(reshapeLp, proto);
2580 node_proto.set_input(0, reshapeLp.name);
2584 LayerParams first_pass;
2585 first_pass.name = layerParams.name + "/flatten";
2586 CV_Assert(layer_id.find(first_pass.name) == layer_id.end());
2587 first_pass.type = "Flatten";
2588 first_pass.set("axis", 0);
2589 first_pass.set("end_axis", axis - 1);
2591 opencv_onnx::NodeProto proto;
2592 proto.add_input(node_proto.input(0));
2593 proto.add_output(first_pass.name);
2594 addLayer(first_pass, proto);
2596 layerParams.set("axis", 1);
2597 node_proto.set_input(0, first_pass.name);
2598 addLayer(layerParams, node_proto);
2601 void ONNXImporter::parseUnsqueeze(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2603 CV_Assert(node_proto.input_size() == 1 || node_proto.input_size() == 2);
2605 if (node_proto.input_size() == 2)
2607 Mat blob = getBlob(node_proto, 1);
2608 axes = DictValue::arrayInt(blob.ptr<int>(), blob.total());
2611 axes = layerParams.get("axes");
2613 if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
2616 Mat input = getBlob(node_proto, 0);
2618 std::vector<int> dims;
2619 for (int j = 0; j < input.dims; j++) {
2620 dims.push_back(input.size[j]);
2622 CV_Assert(axes.getIntValue(axes.size()-1) <= dims.size());
2623 for (int j = 0; j < axes.size(); j++) {
2624 const int idx = axes.getIntValue(j);
2625 CV_Assert(idx <= dims.size());
2626 dims.insert(dims.begin() + idx, 1);
2629 Mat out = input.reshape(0, dims);
2630 addConstant(node_proto.output(0), out);
2635 if (axes.size() != 1)
2636 CV_Error(Error::StsNotImplemented, "Multidimensional unsqueeze");
2638 int depth = layerParams.get<int>("depth", CV_32F);
2640 MatShape inpShape = outShapes[node_proto.input(0)];
2641 int axis = axes.getIntValue(0);
2642 CV_Assert(0 <= axis && axis <= inpShape.size());
2643 std::vector<int> outShape = inpShape;
2644 outShape.insert(outShape.begin() + axis, 1);
2645 layerParams.type = (depth == CV_8S) ? "ReshapeInt8" : "Reshape";
2646 layerParams.set("dim", DictValue::arrayInt(&outShape[0], outShape.size()));
2647 if (hasDynamicShapes)
2649 std::vector<int> dynamicAxes;
2650 std::vector<int> inputIndices;
2651 for (int index = 0; index < outShape.size(); ++index) {
2653 dynamicAxes.push_back(index);
2655 for (int index = 0; index < inpShape.size(); ++index)
2656 inputIndices.push_back(index);
2657 layerParams.set("dynamic_axes", DictValue::arrayInt(dynamicAxes.data(), dynamicAxes.size()));
2658 layerParams.set("input_indices", DictValue::arrayInt(inputIndices.data(), inputIndices.size()));
2660 addLayer(layerParams, node_proto);
2663 void ONNXImporter::parseExpand(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
2665 opencv_onnx::NodeProto node_proto = node_proto_;
2666 CV_CheckEQ(node_proto.input_size(), 2, "");
2667 const std::string& input0 = node_proto.input(0);
2668 const std::string& input1 = node_proto.input(1);
2669 const std::string output_name = node_proto.output(0);
2670 Mat newShapeMat = getBlob(input1);
2671 MatShape targetShape(newShapeMat.ptr<int>(), newShapeMat.ptr<int>() + newShapeMat.total());
2674 bool haveVariables = constBlobs.find(input0) == constBlobs.end();
2677 IterShape_t shapeIt = outShapes.find(input0);
2678 CV_Assert(shapeIt != outShapes.end());
2679 inpShape = shapeIt->second;
2683 inpShape = shape(getBlob(input0));
2686 String srcName = input0;
2687 // Unsqueeze and repeat along new axis
2688 if (targetShape.size() == inpShape.size() + 1)
2690 inpShape.insert(inpShape.begin(), targetShape.size() - inpShape.size(), 1);
2691 for (int i = 0; i < targetShape.size(); i++)
2693 if (abs(targetShape[i]) == 1)
2694 targetShape[i] = inpShape[i];
2698 LayerParams reshapeLp;
2699 reshapeLp.name = layerParams.name + "/reshape";
2700 reshapeLp.type = "Reshape";
2701 CV_Assert(layer_id.find(reshapeLp.name) == layer_id.end());
2702 reshapeLp.set("dim", DictValue::arrayInt(&inpShape[0], inpShape.size()));
2704 opencv_onnx::NodeProto proto;
2705 proto.add_input(node_proto.input(0));
2706 proto.add_output(reshapeLp.name);
2707 addLayer(reshapeLp, proto);
2708 srcName = reshapeLp.name;
2711 CV_CheckEQ(inpShape.size(), targetShape.size(), "Unsupported Expand op with different dims");
2713 std::vector<int> broadcast_axes;
2714 // shapes aren't right-aligned here because targetShape.size() == inpShape.size()
2715 for (int i = 0; i < targetShape.size(); i++)
2717 if (targetShape[i] != inpShape[i])
2719 if (inpShape[i] == 1)
2721 broadcast_axes.push_back(i);
2723 else if (targetShape[i] != 1)
2725 CV_Error(Error::StsError, format("Could not be broadcast by axis: %d", i));
2732 if (broadcast_axes.size() > 1)
2733 CV_Error(Error::StsNotImplemented, "Expand op doesn't support multiple axes for constant input");
2735 if (broadcast_axes.empty())
2737 addConstant(output_name, getBlob(node_proto, 0));
2741 Mat input = getBlob(node_proto, 0);
2742 input = input.reshape(0, total(inpShape, 0, broadcast_axes[0]));
2743 Mat output = cv::repeat(input, 1, targetShape[broadcast_axes[0]]);
2744 output = output.reshape(0, targetShape);
2745 addConstant(output_name, output);
2749 if (broadcast_axes.size() == 2 &&
2750 broadcast_axes[0] == broadcast_axes[1] - 1 && broadcast_axes[1] == inpShape.size() - 1)
2752 LayerParams constParams;
2753 constParams.name = layerParams.name + "/const";
2754 CV_Assert(layer_id.find(constParams.name) == layer_id.end());
2755 constParams.type = "Const";
2757 Mat inp = Mat::ones(newShapeMat.total(), newShapeMat.ptr<int>(), CV_32F);
2758 constParams.blobs.push_back(inp);
2760 opencv_onnx::NodeProto proto;
2761 proto.add_output(constParams.name);
2762 addLayer(constParams, proto);
2764 layerParams.type = "Scale";
2765 layerParams.set("bias_term", false);
2766 node_proto.set_input(0, constParams.name);
2767 node_proto.set_input(1, srcName);
2769 else if (broadcast_axes.size() == 1 && broadcast_axes[0] <= 1)
2771 expandMid(layerParams.name, node_proto, srcName, targetShape[broadcast_axes[0]]);
2773 layerParams.set("axis", broadcast_axes[0]);
2774 layerParams.type = "Concat";
2775 node_proto.set_output(0, output_name);
2777 else if (broadcast_axes.empty())
2779 layerParams.type = "Identity";
2782 CV_Error(Error::StsNotImplemented, "Unsupported Expand op");
2783 addLayer(layerParams, node_proto);
2786 void ONNXImporter::parseReshape(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2788 CV_Assert(node_proto.input_size() == 2 || layerParams.has("shape"));
2789 int depth = layerParams.get<int>("depth", CV_32F);
2790 layerParams.type += (depth == CV_8S) ? "Int8" : "";
2792 if (node_proto.input_size() == 2) {
2793 Mat blob = getBlob(node_proto, 1);
2794 CV_Assert(blob.type() == CV_32SC1);
2796 layerParams.set("dim", DictValue::arrayInt<int*>(blob.ptr<int>(), blob.total()));
2798 if (layer_id.find(node_proto.input(0)) == layer_id.end()) {
2799 std::vector<Mat> inputs(1, getBlob(node_proto, 0)), outputs;
2800 runLayer(layerParams, inputs, outputs);
2801 addConstant(node_proto.output(0), outputs[0]);
2806 DictValue shape = layerParams.get("shape");
2807 std::vector<int> dim;
2808 for (int j = 0; j < shape.size(); j++) {
2809 dim.push_back(shape.getIntValue(j));
2812 if (layer_id.find(node_proto.input(0)) == layer_id.end()) {
2813 Mat input = getBlob(node_proto, 0);
2814 Mat out = input.reshape(0, dim);
2815 addConstant(node_proto.output(0), out);
2818 replaceLayerParam(layerParams, "shape", "dim");
2820 addLayer(layerParams, node_proto);
2823 void ONNXImporter::parsePad(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2825 int depth = layerParams.get<int>("depth", CV_32F);
2826 layerParams.type = (depth == CV_8S) ? "PaddingInt8" : "Padding";
2827 replaceLayerParam(layerParams, "mode", "type");
2828 if (node_proto.input_size() == 3 || node_proto.input_size() == 2)
2830 // Paddings are in order begin0, begin1, .. beginN, end0, end1, ..., endN.
2831 // We need to shuffle it to begin0, end0, begin1, end1, ...
2832 Mat paddings = getBlob(node_proto, 1).reshape(1, 2);
2833 paddings = paddings.t();
2834 layerParams.set("paddings", DictValue::arrayInt(paddings.ptr<int>(), paddings.total()));
2836 if (node_proto.input_size() == 3)
2838 Mat value = getBlob(node_proto, 2);
2839 float padValue = (depth == CV_8S) ? (float)value.ptr<int8_t>()[0] : value.ptr<float>()[0];
2840 layerParams.set("value", padValue);
2843 addLayer(layerParams, node_proto);
2846 void ONNXImporter::parseShape(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2848 CV_Assert(node_proto.input_size() == 1);
2849 IterShape_t shapeIt = outShapes.find(node_proto.input(0));
2850 CV_Assert(shapeIt != outShapes.end());
2851 const MatShape& inpShape = shapeIt->second;
2853 int dims = static_cast<int>(inpShape.size());
2854 Mat shapeMat(dims, 1, CV_32S);
2855 bool isDynamicShape = false;
2856 for (int j = 0; j < dims; ++j)
2858 int sz = inpShape[j];
2859 isDynamicShape |= (sz == 0);
2860 shapeMat.at<int>(j) = sz;
2862 shapeMat.dims = 1; // FIXIT Mat 1D
2866 CV_LOG_ERROR(NULL, "DNN/ONNX(Shape): dynamic 'zero' shapes are not supported, input " << toString(inpShape, node_proto.input(0)));
2867 CV_Assert(!isDynamicShape); // not supported
2869 addConstant(node_proto.output(0), shapeMat);
2872 void ONNXImporter::parseCast(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2874 if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
2876 Mat blob = getBlob(node_proto, 0);
2878 switch (layerParams.get<int>("to"))
2880 case opencv_onnx::TensorProto_DataType_FLOAT: type = CV_32F; break;
2881 case opencv_onnx::TensorProto_DataType_UINT8: type = CV_8U; break;
2882 case opencv_onnx::TensorProto_DataType_UINT16: type = CV_16U; break;
2883 case opencv_onnx::TensorProto_DataType_FLOAT16: type = CV_16S; break;
2884 case opencv_onnx::TensorProto_DataType_INT8:
2885 case opencv_onnx::TensorProto_DataType_INT16:
2886 case opencv_onnx::TensorProto_DataType_INT32:
2887 case opencv_onnx::TensorProto_DataType_INT64: type = CV_32S; break;
2888 default: type = blob.type();
2891 blob.convertTo(dst, type);
2892 dst.dims = blob.dims;
2893 addConstant(node_proto.output(0), dst);
2897 layerParams.type = "Identity";
2898 addLayer(layerParams, node_proto);
2901 void ONNXImporter::parseConstantFill(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
2905 if (!layerParams.blobs.empty())
2907 CV_Assert(!layerParams.has("value"));
2908 depth = layerParams.blobs[0].depth();
2910 layerParams.blobs[0].convertTo(floats, CV_32F);
2911 fill_value = floats.at<float>(0, 0);
2914 fill_value = layerParams.get("value", 0);
2916 MatShape inpShape = getBlob(node_proto, 0);
2917 for (int i = 0; i < inpShape.size(); i++)
2918 CV_CheckGT(inpShape[i], 0, "");
2919 Mat tensor(inpShape.size(), &inpShape[0], depth, Scalar(fill_value));
2920 addConstant(node_proto.output(0), tensor);
2923 void ONNXImporter::parseGather(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
2925 opencv_onnx::NodeProto node_proto = node_proto_;
2926 CV_Assert(node_proto.input_size() == 2);
2927 Mat indexMat = getBlob(node_proto, 1);
2928 CV_Assert_N(indexMat.type() == CV_32S, indexMat.total() == 1);
2929 int index = indexMat.at<int>(0);
2930 int axis = layerParams.get<int>("axis", 0);
2932 if ((constBlobs.find(node_proto.input(0)) != constBlobs.end()))
2934 Mat input = getBlob(node_proto, 0);
2936 std::vector<cv::Range> ranges(input.dims, Range::all());
2937 ranges[axis] = Range(index, index + 1);
2939 out = input(ranges);
2940 MatShape outShape = shape(out);
2941 if (outShape.size() > 1)
2943 outShape.erase(outShape.begin() + axis);
2944 out.reshape(0, outShape);
2948 addConstant(node_proto.output(0), out);
2953 IterShape_t shapeIt = outShapes.find(node_proto.input(0));
2954 CV_Assert(shapeIt != outShapes.end());
2955 MatShape inpShape = shapeIt->second;
2957 LayerParams sliceLp;
2958 sliceLp.type = "Slice";
2959 sliceLp.name = inpShape.size() > 1 ? layerParams.name + "/slice" : layerParams.name;
2960 std::vector<int> begin(inpShape.size(), 0);
2961 std::vector<int> end(inpShape.size(), INT_MAX);
2962 begin[axis] = index;
2963 end[axis] = index + 1;
2965 cv::dnn::DictValue paramBegin = cv::dnn::DictValue::arrayInt(begin.data(), begin.size());
2966 cv::dnn::DictValue paramEnd = cv::dnn::DictValue::arrayInt(end.data(), end.size());
2967 sliceLp.set("begin", paramBegin);
2968 sliceLp.set("end", paramEnd);
2969 sliceLp.set("has_dynamic_shapes", hasDynamicShapes);
2971 if (inpShape.size() > 1)
2973 opencv_onnx::NodeProto proto;
2974 proto.add_input(node_proto.input(0));
2975 proto.add_output(sliceLp.name);
2976 addLayer(sliceLp, proto);
2978 inpShape.erase(inpShape.begin() + axis);
2979 layerParams.type = "Reshape";
2980 layerParams.set("axis", 0);
2981 layerParams.set("dim", DictValue::arrayInt(&inpShape[0], inpShape.size()));
2982 if (hasDynamicShapes)
2984 std::vector<int> dynamicAxes;
2985 std::vector<int> inputIndices;
2986 for (int index = 0; index < inpShape.size(); ++index)
2987 dynamicAxes.push_back(index);
2988 for (int index = 0; index < inpShape.size(); ++index)
2989 inputIndices.push_back(index);
2990 layerParams.set("dynamic_axes", DictValue::arrayInt(dynamicAxes.data(), dynamicAxes.size()));
2991 layerParams.set("input_indices", DictValue::arrayInt(inputIndices.data(), inputIndices.size()));
2993 node_proto.set_input(0, sliceLp.name);
2997 layerParams = sliceLp;
3000 addLayer(layerParams, node_proto);
3003 void ONNXImporter::parseConcat(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3005 bool hasVariableInps = false;
3006 for (int i = 0; i < node_proto.input_size(); ++i)
3008 if (layer_id.find(node_proto.input(i)) != layer_id.end())
3010 hasVariableInps = true;
3015 if (!hasVariableInps)
3017 std::vector<Mat> inputs(node_proto.input_size()), concatenated;
3018 // Due constant folding we can get inputs with different number of dimensions
3019 // Insert the missing dimension to inputs
3020 MatShape inputShape;
3021 for (size_t i = 0; i < inputs.size(); ++i)
3023 inputs[i] = getBlob(node_proto, i);
3024 if (inputs[i].size.dims() > inputShape.size())
3026 inputShape = shape(inputs[i]);
3030 // Concat-1 has default value for axis is 1: https://github.com/onnx/onnx/blob/master/docs/Changelog.md#Concat-1
3031 int axis = layerParams.get<int>("axis", 1);
3032 for (size_t i = 0; i < inputs.size(); ++i)
3034 MatShape targetShape = inputShape;
3035 targetShape[axis] = shape(inputs[i])[axis];
3036 CV_CheckEQ(total(targetShape), total(shape(inputs[i])), "");
3037 inputs[i] = inputs[i].reshape(0, targetShape);
3039 runLayer(layerParams, inputs, concatenated);
3041 CV_Assert(concatenated.size() == 1);
3042 addConstant(node_proto.output(0), concatenated[0]);
3047 for (int i = 0; i < node_proto.input_size(); ++i)
3049 if (constBlobs.find(node_proto.input(i)) != constBlobs.end())
3051 LayerParams constParams;
3052 constParams.name = node_proto.input(i);
3053 constParams.type = "Const";
3054 constParams.blobs.push_back(getBlob(node_proto, i));
3056 opencv_onnx::NodeProto proto;
3057 proto.add_output(constParams.name);
3058 addLayer(constParams, proto);
3062 addLayer(layerParams, node_proto);
3065 // https://github.com/onnx/onnx/blob/master/docs/Operators.md#Resize
3066 void ONNXImporter::parseResize(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3068 for (int i = 1; i < node_proto.input_size(); i++)
3069 CV_Assert(layer_id.find(node_proto.input(i)) == layer_id.end());
3071 int depth = layerParams.get<int>("depth", CV_32F);
3072 layerParams.type += (depth == CV_8S) ? "Int8" : "";
3074 if (layerParams.has("coordinate_transformation_mode"))
3076 String interp_mode = layerParams.get<String>("coordinate_transformation_mode");
3077 CV_Assert_N(interp_mode != "tf_crop_and_resize", interp_mode != "tf_half_pixel_for_nn");
3079 layerParams.set("align_corners", interp_mode == "align_corners");
3080 if (layerParams.get<String>("mode") == "linear")
3082 layerParams.set("mode", interp_mode == "pytorch_half_pixel" || interp_mode == "half_pixel" ?
3083 "opencv_linear" : "bilinear");
3086 if (layerParams.get<String>("mode") == "linear" && framework_name == "pytorch")
3087 layerParams.set("mode", "opencv_linear");
3089 // opset-10: input = [X, scales]
3090 // opset-11: input = [X, roi, scales] or [x, roi, scales, sizes]
3091 // opset-13: may have empty input, [X, "", "", sizes] or [x, "", scales]
3092 int scalesInputId = node_proto.input_size() == 2 ? 1 : 2;
3093 const std::string& scale_name = node_proto.input(scalesInputId);
3095 if(!scale_name.empty())
3096 scales = getBlob(node_proto, scalesInputId);
3098 if (!scales.empty())
3100 CV_CheckEQ(scales.total(), (size_t)4, "HCHW layout is expected");
3101 layerParams.set("zoom_factor_y", scales.at<float>(2));
3102 layerParams.set("zoom_factor_x", scales.at<float>(3));
3104 else if (node_proto.input_size() >= 4) // opset-11 [x, roi, scales, sizes] or opset-13: input = [X, "", "", sizes]
3106 const std::string& inputSizes = node_proto.input(3);
3107 if (constBlobs.find(inputSizes) != constBlobs.end())
3109 Mat shapes = getBlob(inputSizes);
3110 CV_CheckEQ(shapes.total(), (size_t)4, "HCHW layout is expected");
3111 CV_CheckDepth(shapes.depth(), shapes.depth() == CV_32S || shapes.depth() == CV_32F, "");
3112 if (shapes.depth() == CV_32F)
3113 shapes.convertTo(shapes, CV_32S);
3114 layerParams.set("width", shapes.at<int>(3));
3115 layerParams.set("height", shapes.at<int>(2));
3119 CV_Error(Error::StsNotImplemented, cv::format("ONNX/Resize: doesn't support dynamic non-constant 'sizes' input: %s", inputSizes.c_str()));
3124 CV_Error(Error::StsNotImplemented, "ONNX/Resize: can't find neither 'scale' nor destination sizes parameters");
3126 replaceLayerParam(layerParams, "mode", "interpolation");
3127 addLayer(layerParams, node_proto);
3130 void ONNXImporter::parseUpsample(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3132 //fused from Resize Subgraph
3133 if (layerParams.has("coordinate_transformation_mode"))
3135 String interp_mode = layerParams.get<String>("coordinate_transformation_mode");
3136 CV_Assert_N(interp_mode != "tf_crop_and_resize", interp_mode != "tf_half_pixel_for_nn");
3138 layerParams.set("align_corners", interp_mode == "align_corners");
3139 if (layerParams.get<String>("mode") == "linear")
3141 layerParams.set("mode", interp_mode == "pytorch_half_pixel" ?
3142 "opencv_linear" : "bilinear");
3145 if (layerParams.get<String>("mode") == "linear" && framework_name == "pytorch")
3146 layerParams.set("mode", "opencv_linear");
3148 layerParams.type = "Resize";
3149 if (layerParams.has("scales"))
3152 DictValue scales = layerParams.get("scales");
3153 CV_Assert(scales.size() == 4);
3154 layerParams.set("zoom_factor_y", scales.getIntValue(2));
3155 layerParams.set("zoom_factor_x", scales.getIntValue(3));
3157 else if (layerParams.has("height_scale") && layerParams.has("width_scale"))
3160 replaceLayerParam(layerParams, "height_scale", "zoom_factor_y");
3161 replaceLayerParam(layerParams, "width_scale", "zoom_factor_x");
3166 const std::string& input1 = node_proto.input(1);
3167 if (constBlobs.find(input1) != constBlobs.end())
3169 Mat scales = getBlob(input1);
3170 CV_Assert(scales.total() == 4);
3171 layerParams.set("zoom_factor_y", scales.at<float>(2));
3172 layerParams.set("zoom_factor_x", scales.at<float>(3));
3175 replaceLayerParam(layerParams, "mode", "interpolation");
3176 addLayer(layerParams, node_proto);
3179 void ONNXImporter::parseSoftMax(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3181 const std::string& layer_type = node_proto.op_type();
3182 layerParams.type = "Softmax";
3183 layerParams.set("log_softmax", layer_type == "LogSoftmax");
3184 addLayer(layerParams, node_proto);
3187 void ONNXImporter::parseDetectionOutput(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
3189 opencv_onnx::NodeProto node_proto = node_proto_;
3190 CV_CheckEQ(node_proto.input_size(), 3, "");
3191 if (constBlobs.find(node_proto.input(2)) != constBlobs.end())
3193 Mat priors = getBlob(node_proto, 2);
3195 LayerParams constParams;
3196 constParams.name = layerParams.name + "/priors";
3197 constParams.type = "Const";
3198 constParams.blobs.push_back(priors);
3200 opencv_onnx::NodeProto priorsProto;
3201 priorsProto.add_output(constParams.name);
3202 addLayer(constParams, priorsProto);
3204 node_proto.set_input(2, constParams.name);
3206 addLayer(layerParams, node_proto);
3209 void ONNXImporter::parseCumSum(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3211 layerParams.type = "CumSum";
3214 const std::string& input1 = node_proto.input(1);
3216 if (constBlobs.find(input1) != constBlobs.end())
3218 Mat axis_blob = getBlob(input1);
3219 CV_Assert(axis_blob.total() == 1u);
3220 layerParams.set("axis", axis_blob.at<int>(0));
3223 addLayer(layerParams, node_proto);
3226 void ONNXImporter::parseDepthToSpace(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
3228 // We parse "DepthToSpace" and "SpaceToDepth" in this function.
3229 opencv_onnx::NodeProto node_proto = node_proto_;
3230 const std::string& layer_type = node_proto.op_type();
3231 CV_Assert(layer_type == "DepthToSpace" || layer_type == "SpaceToDepth");
3234 CV_Assert(layerParams.has("blocksize"));
3235 int blocksize = layerParams.get<int>("blocksize");
3236 CV_Assert(blocksize > 0);
3238 // Get mode, only for "DepthToSpace"
3239 std::string modeType = layerParams.get<std::string>("mode", "DCR");
3241 MatShape inpShape = outShapes[node_proto.input(0)];
3242 CV_Assert(inpShape.size() == 4);
3243 int N = inpShape[0], C = inpShape[1], H = inpShape[2], W = inpShape[3];
3245 // Implement DepthToSpace and SpaceToDepth by the Reshape and Permute layer.
3246 std::array<int, 6> shape0, perm;
3247 std::array<int, 4> shape1;
3249 if (layer_type == "DepthToSpace")
3251 if (modeType == "DCR")
3253 shape0 = {N, blocksize, blocksize, C/(blocksize * blocksize), H, W};
3254 perm = {0, 3, 4, 1, 5, 2};
3255 shape1 = {N, C/(blocksize * blocksize), H * blocksize, W * blocksize};
3257 else if (modeType == "CRD")
3259 shape0 = {N, C/(blocksize * blocksize), blocksize, blocksize, H, W};
3260 perm = {0, 1, 4, 2, 5, 3};
3261 shape1 = {N, C/(blocksize * blocksize), H * blocksize, W * blocksize};
3264 CV_Error(Error::StsNotImplemented, "The mode of " + modeType + " in " + layer_type + " Layer is not supported");
3266 else // SpaceToDepth
3268 shape0 = {N, C, H/blocksize, blocksize, W/blocksize, blocksize};
3269 perm = {0, 3, 5, 1, 2, 4};
3270 shape1 = {N, C * blocksize * blocksize, H/blocksize, W/blocksize};
3274 LayerParams reshapeLp;
3275 reshapeLp.name = layerParams.name + "/reshape";
3276 reshapeLp.type = "Reshape";
3277 CV_Assert(layer_id.find(reshapeLp.name) == layer_id.end());
3278 reshapeLp.set("dim", DictValue::arrayInt(shape0.data(), shape0.size()));
3280 opencv_onnx::NodeProto protoReshape;
3281 protoReshape.add_input(node_proto.input(0));
3282 protoReshape.add_output(reshapeLp.name);
3283 addLayer(reshapeLp, protoReshape);
3286 LayerParams permuteLp;
3287 permuteLp.name = layerParams.name + "/permute";
3288 permuteLp.type = "Permute";
3289 CV_Assert(layer_id.find(permuteLp.name) == layer_id.end());
3290 permuteLp.set("order", DictValue::arrayInt(perm.data(), perm.size()));
3292 opencv_onnx::NodeProto protoPermute;
3293 protoPermute.add_input(reshapeLp.name);
3294 protoPermute.add_output(permuteLp.name);
3295 addLayer(permuteLp, protoPermute);
3298 layerParams.type = "Reshape";
3299 layerParams.set("dim", DictValue::arrayInt(shape1.data(), shape1.size()));
3301 node_proto.set_input(0, permuteLp.name);
3302 addLayer(layerParams, node_proto);
3305 void ONNXImporter::parseSimpleLayers(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3307 for (int j = 0; j < node_proto.input_size(); j++) {
3308 if (layer_id.find(node_proto.input(j)) == layer_id.end())
3309 layerParams.blobs.push_back(getBlob(node_proto, j));
3311 addLayer(layerParams, node_proto);
3314 void ONNXImporter::parseCustomLayer(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3316 const std::string& name = layerParams.name;
3317 std::string& layer_type = layerParams.type;
3318 const std::string& layer_type_domain = node_proto.has_domain() ? node_proto.domain() : std::string();
3319 if (!layer_type_domain.empty() && layer_type_domain != str_domain_ai_onnx)
3321 // append ONNX domain name
3322 static bool DNN_CUSTOM_ONNX_TYPE_INCLUDE_DOMAIN_NAME = utils::getConfigurationParameterBool("OPENCV_DNN_CUSTOM_ONNX_TYPE_INCLUDE_DOMAIN_NAME", true);
3323 if (DNN_CUSTOM_ONNX_TYPE_INCLUDE_DOMAIN_NAME)
3325 layer_type = layer_type_domain + "." + layer_type;
3329 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: "
3330 << cv::format("[%s]:(%s)", layer_type.c_str(), name.c_str())
3333 parseSimpleLayers(layerParams, node_proto);
3336 void ONNXImporter::parseQuantDequant(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3338 CV_Assert(node_proto.input_size() == 3);
3339 layerParams.type = (node_proto.op_type() == "QuantizeLinear") ? "Quantize" : "Dequantize";
3341 if (node_proto.op_type() == "DequantizeLinear")
3343 Mat scale = getBlob(node_proto, 1);
3344 Mat zeropoint = getBlob(node_proto, 2);
3346 layerParams.set("scales", DictValue::arrayReal(scale.ptr<float>(), 1));
3347 layerParams.set("zeropoints", DictValue::arrayInt(zeropoint.ptr<int8_t>(), 1));
3349 addLayer(layerParams, node_proto);
3352 void ONNXImporter::parseQConv(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
3354 opencv_onnx::NodeProto node_proto = node_proto_;
3355 int ninputs = node_proto.input_size();
3356 CV_Assert(ninputs == 8 || ninputs == 9);
3358 Mat inp_sc = getBlob(node_proto, 1);
3359 Mat inp_zp = getBlob(node_proto, 2);
3361 if (layerParams.has("pad"))
3363 bool asymmetricPadding = false;
3364 DictValue pads = layerParams.get("pad");
3365 const int dims = pads.size() / 2;
3367 for (int i = 0; i < dims; ++i)
3369 if (pads.get<int>(i) != pads.get<int>(i + dims))
3371 asymmetricPadding = true;
3375 if (asymmetricPadding && pads.size() == 4)
3377 layerParams.erase("pad");
3378 std::vector<int> paddings(4, 0);
3379 for (int i = 0; i < dims; ++i)
3381 paddings.push_back(pads.get<int>(i));
3382 paddings.push_back(pads.get<int>(dims + i));
3385 padLp.name = layerParams.name + "/pad";
3386 padLp.type = "PaddingInt8";
3387 padLp.set("paddings", DictValue::arrayInt(&paddings[0], paddings.size()));
3388 padLp.set("depth", CV_8S);
3389 padLp.set("value", inp_zp.at<int8_t>(0));
3391 opencv_onnx::NodeProto proto;
3392 proto.add_input(node_proto.input(0));
3393 proto.add_output(padLp.name);
3395 addLayer(padLp, proto);
3396 node_proto.set_input(0, padLp.name);
3400 Mat weights = getBlob(node_proto, 3);
3401 int outCn = weights.size[0];
3402 Mat w_scale = getBlob(node_proto, 4);
3403 CV_Assert(w_scale.total() == 1 || w_scale.total() == outCn);
3404 bool per_channel = w_scale.total() == outCn ? true : false;
3405 Mat wt_sc = (w_scale.total() == outCn) ? w_scale : Mat(1, outCn, CV_32F, Scalar(w_scale.at<float>(0)));
3407 Mat out_sc = getBlob(node_proto, 6);
3408 Mat bias = (ninputs == 9) ? getBlob(node_proto, 8) : Mat::zeros(1, outCn, CV_32S);
3410 Mat weights_2d = weights.reshape(1, outCn);
3411 Mat biasFused(1, outCn, CV_32S);
3412 Mat outputMultiplier(1, outCn, CV_32F);
3413 for (int i = 0; i < outCn; i++)
3415 biasFused.at<int>(i) = bias.at<int>(i) - inp_zp.at<int8_t>(0)*(cv::sum(weights_2d.row(i))[0]);
3416 outputMultiplier.at<float>(i) = (inp_sc.at<float>(0) * wt_sc.at<float>(i)) / out_sc.at<float>(0);
3419 layerParams.type = "ConvolutionInt8";
3420 layerParams.set("num_output", outCn);
3421 layerParams.set("input_zeropoint", inp_zp.at<int8_t>(0));
3422 layerParams.set("input_scale",inp_sc.at<float>(0));
3423 layerParams.set("per_channel", per_channel);
3424 layerParams.blobs.push_back(weights);
3425 layerParams.blobs.push_back(biasFused);
3426 layerParams.blobs.push_back(outputMultiplier);
3427 addLayer(layerParams, node_proto);
3430 void ONNXImporter::parseQMatMul(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3432 int ninputs = node_proto.input_size();
3433 CV_Assert(ninputs == 8);
3435 if (constBlobs.find(node_proto.input(3)) == constBlobs.end())
3436 CV_Error(Error::StsNotImplemented, "Variable weights is not supported");
3438 int firstInpDims = outShapes[node_proto.input(0)].size();
3440 Mat inp_sc = getBlob(node_proto, 1);
3441 Mat inp_zp = getBlob(node_proto, 2);
3443 Mat weights = getBlob(node_proto, 3).t();
3444 int outCn = weights.size[0];
3445 int secondInpDims = weights.dims;
3447 Mat w_scale = getBlob(node_proto, 4);
3448 CV_Assert(w_scale.total() == 1 || w_scale.total() == outCn);
3449 bool per_channel = w_scale.total() == outCn ? true : false;
3450 Mat wt_sc = (w_scale.total() == outCn) ? w_scale : Mat(1, outCn, CV_32F, Scalar(w_scale.at<float>(0)));
3451 Mat out_sc = getBlob(node_proto, 6);
3453 Mat bias(1, outCn, CV_32S);
3454 Mat outputMultiplier(1, outCn, CV_32F);
3455 for (int i = 0; i < outCn; i++)
3457 bias.at<int>(i) = -inp_zp.at<int8_t>(0)*(cv::sum(weights.row(i))[0]);
3458 outputMultiplier.at<float>(i) = (inp_sc.at<float>(0) * wt_sc.at<float>(i)) / out_sc.at<float>(0);
3461 layerParams.type = "InnerProductInt8";
3462 layerParams.set("num_output", outCn);
3463 layerParams.set("axis", firstInpDims - secondInpDims + 1);
3464 layerParams.set("input_scale", inp_sc.at<float>(0));
3465 layerParams.set("input_zeropoint", inp_zp.at<int8_t>(0));
3466 layerParams.set("per_channel", per_channel);
3468 layerParams.blobs.push_back(weights);
3469 layerParams.blobs.push_back(bias);
3470 layerParams.blobs.push_back(outputMultiplier);
3471 addLayer(layerParams, node_proto);
3474 void ONNXImporter::parseQEltwise(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
3476 opencv_onnx::NodeProto node_proto = node_proto_;
3477 CV_Assert(node_proto.input_size() == 8);
3478 std::string op = (node_proto.op_type() == "QLinearAdd") ? "sum" : "prod";
3480 for (int i = 0; i < 4; i += 3)
3482 if (constBlobs.find(node_proto.input(i)) != constBlobs.end())
3486 Mat inp_0_sc = getBlob(node_proto, 1);
3487 Mat inp_0_zp = getBlob(node_proto, 2);
3489 Mat inp_1_sc = getBlob(node_proto, 4);
3490 Mat inp_1_zp = getBlob(node_proto, 5);
3492 // Set 2nd input as the const input
3495 cv::swap(inp_0_sc, inp_1_sc);
3496 cv::swap(inp_0_zp, inp_1_zp);
3499 float out_sc = getBlob(node_proto, 6).at<float>(0);
3500 int8_t out_zp = getBlob(node_proto, 7).at<int8_t>(0);
3502 std::vector<float> inp_scales = {inp_0_sc.at<float>(0), inp_1_sc.at<float>(0)};
3503 std::vector<int8_t> inp_zps = {inp_0_zp.at<int8_t>(0), inp_1_zp.at<int8_t>(0)};
3505 std::vector<float> coeffs;
3509 coeffs = {inp_scales[0]/out_sc, inp_scales[1]/out_sc};
3510 offset = out_zp - coeffs[0]*inp_zps[0] - coeffs[1]*inp_zps[1];
3514 coeffs = {inp_scales[0]/out_sc, inp_scales[1]};
3520 Mat blob = getBlob(node_proto, constId);
3521 if (blob.total() == 1)
3523 float val = inp_scales[1] * (blob.at<int8_t>(0) - inp_zps[1]);
3524 float scale = inp_scales[0] / out_sc;
3528 float shift = out_zp - scale*inp_zps[0];
3530 shift += (val/out_sc);
3532 LayerParams rescaleParams;
3533 rescaleParams.name = layerParams.name;
3534 rescaleParams.type = "Requantize";
3535 rescaleParams.set("depth", CV_8S);
3536 rescaleParams.set("scale", scale);
3537 rescaleParams.set("shift", shift);
3538 rescaleParams.set("isEltwise", true);
3539 addLayer(rescaleParams, node_proto);
3544 MatShape inpShape = outShapes[node_proto.input(3 - constId)];
3548 if (shape(blob) == inpShape)
3550 LayerParams constParams;
3551 constParams.name = layerParams.name + "/const";
3552 constParams.type = "ConstInt8";
3553 constParams.set("depth", CV_8S);
3554 constParams.set("scales", DictValue::arrayReal(inp_1_sc.ptr<float>(), 1));
3555 constParams.set("zeropoints", DictValue::arrayInt(inp_1_zp.ptr<int8_t>(), 1));
3556 constParams.blobs.push_back(blob);
3558 int id = dstNet.addLayer(constParams.name, constParams.type, CV_8S, constParams);
3559 layer_id.insert(std::make_pair(constParams.name, LayerInfo(id, 0)));
3560 outShapes[constParams.name] = shape(blob);
3561 node_proto.set_input(constId, constParams.name);
3563 layerParams.type = "EltwiseInt8";
3564 layerParams.set("operation", op);
3565 layerParams.set("coeff", DictValue::arrayReal(coeffs.data(), coeffs.size()));
3566 layerParams.set("offset", offset);
3570 layerParams.type = "ScaleInt8";
3571 layerParams.set("bias_term", op == "sum");
3573 for (int i = 0; i < graph_proto.initializer_size(); i++)
3575 opencv_onnx::TensorProto tensor_proto = graph_proto.initializer(i);
3576 if (tensor_proto.name() == node_proto.input(constId))
3578 axis = inpShape.size() - tensor_proto.dims_size();
3582 layerParams.set("axis", axis);
3583 blob = blob.reshape(1, 1);
3584 Mat blob_dequantized;
3585 blob.convertTo(blob_dequantized, CV_32F, inp_scales[1], -(inp_scales[1] * inp_zps[1]));
3586 layerParams.blobs.push_back(blob_dequantized);
3590 else if (outShapes[node_proto.input(0)] == outShapes[node_proto.input(3)])
3592 layerParams.type = "EltwiseInt8";
3593 layerParams.set("operation", op);
3594 layerParams.set("coeff", DictValue::arrayReal(coeffs.data(), coeffs.size()));
3595 layerParams.set("offset", offset);
3599 layerParams.type = "ScaleInt8";
3600 layerParams.set("bias_term", op == "sum");
3603 layerParams.set("input_scales", DictValue::arrayReal(inp_scales.data(), inp_scales.size()));
3604 layerParams.set("input_zeropoints", DictValue::arrayInt(inp_zps.data(), inp_zps.size()));
3605 addLayer(layerParams, node_proto);
3608 void ONNXImporter::parseQLeakyRelu(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3610 CV_Assert(node_proto.input_size() == 5);
3612 float slope = layerParams.get<float>("alpha");
3613 float inp_sc = getBlob(node_proto, 1).at<float>(0);
3614 int8_t inp_zp = getBlob(node_proto, 2).at<int8_t>(0);
3615 float out_sc = getBlob(node_proto, 3).at<float>(0);
3616 int8_t out_zp = getBlob(node_proto, 4).at<int8_t>(0);
3618 Mat lookUpTable(1, 256, CV_8S);
3619 int8_t* table = lookUpTable.ptr<int8_t>();
3620 for (int i = -128; i < 128; i++)
3622 float x = inp_sc*(i - inp_zp);
3623 float y = x >= 0.f ? x : slope*x;
3624 int quantized = out_zp + cvRound(y/out_sc);
3625 table[i+128] = saturate_cast<int8_t>(quantized);
3628 layerParams.type = "ReLUInt8";
3629 layerParams.set("input_scale", inp_sc);
3630 layerParams.set("input_zeropoint", inp_zp);
3631 layerParams.set("slope", slope);
3632 layerParams.blobs.push_back(lookUpTable);
3633 addLayer(layerParams, node_proto);
3636 void ONNXImporter::parseQSigmoid(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3638 CV_Assert(node_proto.input_size() == 5);
3640 float inp_sc = getBlob(node_proto, 1).at<float>(0);
3641 int8_t inp_zp = getBlob(node_proto, 2).at<int8_t>(0);
3642 float out_sc = getBlob(node_proto, 3).at<float>(0);
3643 int8_t out_zp = getBlob(node_proto, 4).at<int8_t>(0);
3645 Mat lookUpTable(1, 256, CV_8S);
3646 int8_t* table = lookUpTable.ptr<int8_t>();
3647 for (int i = -128; i < 128; i++)
3649 float x = inp_sc*(i - inp_zp);
3650 float y = 1.f/(1.f + std::exp(-x));
3651 int quantized = out_zp + cvRound(y/out_sc);
3652 table[i+128] = saturate_cast<int8_t>(quantized);
3655 layerParams.type = "SigmoidInt8";
3656 layerParams.set("input_scale", inp_sc);
3657 layerParams.set("input_zeropoint", inp_zp);
3658 layerParams.blobs.push_back(lookUpTable);
3659 addLayer(layerParams, node_proto);
3662 void ONNXImporter::parseQAvgPool(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto)
3664 CV_Assert(node_proto.input_size() == 5);
3665 float inp_sc = getBlob(node_proto, 1).at<float>(0);
3666 int8_t inp_zp = getBlob(node_proto, 2).at<int8_t>(0);
3667 float out_sc = getBlob(node_proto, 3).at<float>(0);
3669 layerParams.type = "PoolingInt8";
3670 layerParams.set("pool", "ave");
3671 layerParams.set("global_pooling", node_proto.op_type() == "QLinearGlobalAveragePool");
3672 layerParams.set("multiplier", inp_sc/out_sc);
3673 layerParams.set("input_zeropoint", inp_zp);
3674 addLayer(layerParams, node_proto);
3677 void ONNXImporter::parseQConcat(LayerParams& layerParams, const opencv_onnx::NodeProto& node_proto_)
3679 opencv_onnx::NodeProto node_proto = node_proto_;
3680 layerParams.type = "ConcatInt8";
3681 int num_inputs = node_proto.input_size();
3683 float out_scale = getBlob(node_proto, 0).at<float>(0);
3684 int out_zp = getBlob(node_proto, 1).at<int8_t>(0);
3686 for (int i = 2; i < num_inputs; i += 3)
3688 float inp_scale = getBlob(node_proto, i + 1).at<float>(0);
3689 int inp_zp = getBlob(node_proto, i + 2).at<int8_t>(0);
3691 if (inp_scale != out_scale || inp_zp != out_zp)
3693 float scale = inp_scale/out_scale;
3694 float shift = out_zp - scale*inp_zp;
3696 if (constBlobs.find(node_proto.input(i)) != constBlobs.end())
3698 Mat blob = getBlob(node_proto, i);
3700 blob.convertTo(blob_rescaled, CV_8S, scale, shift);
3701 constBlobs[node_proto.input(i)] = blob_rescaled;
3705 LayerParams rescaleParams;
3706 rescaleParams.name = node_proto.input(i) + "/rescale";
3707 rescaleParams.type = "Requantize";
3708 rescaleParams.set("depth", CV_8S);
3709 rescaleParams.set("scale", scale);
3710 rescaleParams.set("shift", shift);
3711 rescaleParams.set("isEltwise", false);
3713 opencv_onnx::NodeProto proto;
3714 proto.add_input(node_proto.input(i));
3715 proto.add_output(rescaleParams.name);
3716 addLayer(rescaleParams, proto);
3717 node_proto.set_input(i, rescaleParams.name);
3722 bool hasVariableInps = false;
3723 for (int i = 2; i < num_inputs; i += 3)
3725 if (layer_id.find(node_proto.input(i)) != layer_id.end())
3727 hasVariableInps = true;
3732 if (!hasVariableInps)
3734 std::vector<Mat> inputs, concatenated;
3735 MatShape inputShape;
3736 for (size_t i = 2; i < num_inputs; i += 3)
3738 Mat blob = getBlob(node_proto, i);
3739 if (blob.size.dims() > inputShape.size())
3741 inputShape = shape(blob);
3743 inputs.push_back(blob);
3746 int axis = layerParams.get<int>("axis", 1);
3747 for (size_t i = 0; i < inputs.size(); ++i)
3749 MatShape targetShape = inputShape;
3750 targetShape[axis] = shape(inputs[i])[axis];
3751 CV_CheckEQ(total(targetShape), total(shape(inputs[i])), "");
3752 inputs[i] = inputs[i].reshape(0, targetShape);
3754 runLayer(layerParams, inputs, concatenated);
3755 CV_Assert(concatenated.size() == 1);
3756 addConstant(layerParams.name, concatenated[0]);
3761 for (int i = 2; i < num_inputs; i += 3)
3763 if (constBlobs.find(node_proto.input(i)) != constBlobs.end())
3765 LayerParams constParams;
3766 constParams.name = node_proto.input(i);
3767 constParams.type = "ConstInt8";
3768 constParams.blobs.push_back(getBlob(node_proto, i));
3769 constParams.set("depth", CV_8S);
3771 opencv_onnx::NodeProto proto;
3772 proto.add_output(constParams.name);
3773 addLayer(constParams, proto);
3777 addLayer(layerParams, node_proto);
3780 // Domain: ai.onnx (default)
3781 // URL: https://github.com/onnx/onnx/blob/master/docs/Operators.md
3782 void ONNXImporter::buildDispatchMap_ONNX_AI(int opset_version)
3784 CV_UNUSED(opset_version);
3785 DispatchMap dispatch;
3787 dispatch["ArgMax"] = dispatch["ArgMin"] = &ONNXImporter::parseArg;
3788 dispatch["MaxUnpool"] = &ONNXImporter::parseMaxUnpool;
3789 dispatch["MaxPool"] = &ONNXImporter::parseMaxPool;
3790 dispatch["AveragePool"] = &ONNXImporter::parseAveragePool;
3791 dispatch["GlobalAveragePool"] = dispatch["GlobalMaxPool"] = &ONNXImporter::parseGlobalPool;
3792 dispatch["ReduceMax"] = dispatch["ReduceMin"] = dispatch["ReduceMean"] = dispatch["ReduceSum"] = dispatch["ReduceMax"] =
3793 dispatch["ReduceMin"] = dispatch["ReduceSumSquare"] = dispatch["ReduceProd"] = dispatch["ReduceL1"] =
3794 dispatch["ReduceL2"] = dispatch["ReduceLogSum"] = dispatch["ReduceLogSumExp"] = &ONNXImporter::parseReduce;
3795 dispatch["Slice"] = &ONNXImporter::parseSlice;
3796 dispatch["Split"] = &ONNXImporter::parseSplit;
3797 dispatch["Add"] = dispatch["Sum"] = dispatch["Sub"] = &ONNXImporter::parseBias;
3798 dispatch["Pow"] = &ONNXImporter::parsePow;
3799 dispatch["Min"] = dispatch["Max"] = &ONNXImporter::parseMinMax;
3800 dispatch["Neg"] = &ONNXImporter::parseNeg;
3801 dispatch["Constant"] = &ONNXImporter::parseConstant;
3802 dispatch["LSTM"] = &ONNXImporter::parseLSTM;
3803 dispatch["GRU"] = &ONNXImporter::parseGRU;
3804 dispatch["ImageScaler"] = &ONNXImporter::parseImageScaler;
3805 dispatch["Clip"] = &ONNXImporter::parseClip;
3806 dispatch["LeakyRelu"] = &ONNXImporter::parseLeakyRelu;
3807 dispatch["Relu"] = &ONNXImporter::parseRelu;
3808 dispatch["Elu"] = &ONNXImporter::parseElu;
3809 dispatch["Tanh"] = &ONNXImporter::parseTanh;
3810 dispatch["Abs"] = &ONNXImporter::parseAbs;
3811 dispatch["Equal"] = dispatch["Greater"] = dispatch["Less"] = &ONNXImporter::parseCompare;
3812 dispatch["PRelu"] = &ONNXImporter::parsePRelu;
3813 dispatch["LRN"] = &ONNXImporter::parseLRN;
3814 dispatch["InstanceNormalization"] = &ONNXImporter::parseInstanceNormalization;
3815 dispatch["BatchNormalization"] = &ONNXImporter::parseBatchNormalization;
3816 dispatch["Gemm"] = &ONNXImporter::parseGemm;
3817 dispatch["MatMul"] = &ONNXImporter::parseMatMul;
3818 dispatch["Mul"] = dispatch["Div"] = &ONNXImporter::parseMul;
3819 dispatch["Conv"] = &ONNXImporter::parseConv;
3820 dispatch["ConvTranspose"] = &ONNXImporter::parseConvTranspose;
3821 dispatch["Transpose"] = &ONNXImporter::parseTranspose;
3822 dispatch["Squeeze"] = &ONNXImporter::parseSqueeze;
3823 dispatch["Flatten"] = &ONNXImporter::parseFlatten;
3824 dispatch["Unsqueeze"] = &ONNXImporter::parseUnsqueeze;
3825 dispatch["Expand"] = &ONNXImporter::parseExpand;
3826 dispatch["Reshape"] = &ONNXImporter::parseReshape;
3827 dispatch["Pad"] = &ONNXImporter::parsePad;
3828 dispatch["Shape"] = &ONNXImporter::parseShape;
3829 dispatch["Cast"] = &ONNXImporter::parseCast;
3830 dispatch["ConstantFill"] = dispatch["ConstantOfShape"] = &ONNXImporter::parseConstantFill;
3831 dispatch["Gather"] = &ONNXImporter::parseGather;
3832 dispatch["Concat"] = &ONNXImporter::parseConcat;
3833 dispatch["Resize"] = &ONNXImporter::parseResize;
3834 dispatch["Upsample"] = &ONNXImporter::parseUpsample;
3835 dispatch["SoftMax"] = dispatch["LogSoftmax"] = &ONNXImporter::parseSoftMax;
3836 dispatch["DetectionOutput"] = &ONNXImporter::parseDetectionOutput;
3837 dispatch["CumSum"] = &ONNXImporter::parseCumSum;
3838 dispatch["SpaceToDepth"] = dispatch["DepthToSpace"] = &ONNXImporter::parseDepthToSpace;
3840 std::vector<std::string> simpleLayers{"Acos", "Acosh", "Asin", "Asinh", "Atan", "Atanh", "Ceil", "Celu", "Cos",
3841 "Cosh", "Dropout", "Erf", "Exp", "Floor", "HardSigmoid", "HardSwish",
3842 "Identity", "Log", "Round", "Reciprocal", "Selu", "Sign", "Sigmoid", "Sin", "Sinh", "Softmax",
3843 "Softplus", "Softsign", "Shrink", "Sqrt", "Tan", "ThresholdedRelu"};
3844 for (const auto& name : simpleLayers)
3846 dispatch[name] = &ONNXImporter::parseSimpleLayers;
3849 // ai.onnx: opset 10+
3850 dispatch["QuantizeLinear"] = dispatch["DequantizeLinear"] = &ONNXImporter::parseQuantDequant;
3851 dispatch["QLinearConv"] = &ONNXImporter::parseQConv;
3852 dispatch["QLinearMatMul"] = &ONNXImporter::parseQMatMul;
3854 domain_dispatch_map[str_domain_ai_onnx] = dispatch;
3857 // Domain: com.microsoft
3858 // URL: https://github.com/microsoft/onnxruntime/blob/master/docs/ContribOperators.md
3859 void ONNXImporter::buildDispatchMap_COM_MICROSOFT(int opset_version)
3861 CV_UNUSED(opset_version);
3862 DispatchMap dispatch;
3864 dispatch["QLinearAdd"] = dispatch["QLinearMul"] = &ONNXImporter::parseQEltwise;
3865 dispatch["QLinearAveragePool"] = dispatch["QLinearGlobalAveragePool"] = &ONNXImporter::parseQAvgPool;
3866 dispatch["QLinearLeakyRelu"] = &ONNXImporter::parseQLeakyRelu;
3867 dispatch["QLinearSigmoid"] = &ONNXImporter::parseQSigmoid;
3868 dispatch["QLinearConcat"] = &ONNXImporter::parseQConcat;
3870 domain_dispatch_map["com.microsoft"] = dispatch;
3874 Net readNetFromONNX(const String& onnxFile)
3876 return detail::readNetDiagnostic<ONNXImporter>(onnxFile.c_str());
3879 Net readNetFromONNX(const char* buffer, size_t sizeBuffer)
3881 return detail::readNetDiagnostic<ONNXImporter>(buffer, sizeBuffer);
3884 Net readNetFromONNX(const std::vector<uchar>& buffer)
3886 return readNetFromONNX(reinterpret_cast<const char*>(buffer.data()), buffer.size());
3889 Mat readTensorFromONNX(const String& path)
3891 std::fstream input(path.c_str(), std::ios::in | std::ios::binary);
3894 CV_Error(Error::StsBadArg, cv::format("Can't read ONNX file: %s", path.c_str()));
3897 opencv_onnx::TensorProto tensor_proto = opencv_onnx::TensorProto();
3898 if (!tensor_proto.ParseFromIstream(&input))
3900 CV_Error(Error::StsUnsupportedFormat, cv::format("Failed to parse ONNX data: %s", path.c_str()));
3902 Mat mat = getMatFromTensor(tensor_proto);
3903 releaseONNXTensor(tensor_proto);
3907 CV__DNN_INLINE_NS_END