Ptr<Layer> createLayerFromCaffe<EltwiseLayer>(LayerParams& params)
{
EltwiseLayer::EltwiseOp op = EltwiseLayer::SUM;
if (params.has("operation"))
{
String operation = params.get<String>("operation").toLowerCase();
if (operation == "prod")
op = EltwiseLayer::PROD;
else if (operation == "sum")
op = EltwiseLayer::SUM;
else if (operation == "max")
op = EltwiseLayer::MAX;
else
CV_Error(cv::Error::StsBadArg, "Unknown operaticon type \"" + operation + "\"");
}
std::vector<int> coeffs;
if (params.has("coeff"))
{
DictValue paramCoeff = params.get("coeff");
coeffs.resize(paramCoeff.size(), 1);
for (int i = 0; i < paramCoeff.size(); i++)
{
coeffs[i] = paramCoeff.get<int>(i);
}
}
return Ptr<Layer>(EltwiseLayer::create(op, coeffs));
}
void testInPlaceActivation(LayerParams& lp)
{
EXPECT_FALSE(lp.name.empty());
LayerParams pool;
pool.set("pool", "ave");
pool.set("kernel_w", 2);
pool.set("kernel_h", 2);
pool.set("stride_w", 2);
pool.set("stride_h", 2);
pool.type = "Pooling";
Net net;
int poolId = net.addLayer(pool.name, pool.type, pool);
net.connect(0, 0, poolId, 0);
net.addLayerToPrev(lp.name, lp.type, lp);
Mat input({1, kNumChannels, 10, 10}, CV_32F);
randu(input, -1.0f, 1.0f);
net.setInput(input);
Mat outputDefault = net.forward(lp.name).clone();
net.setInput(input);
net.setPreferableBackend(DNN_BACKEND_HALIDE);
Mat outputHalide = net.forward(lp.name).clone();
normAssert(outputDefault, outputHalide);
}
TEST_P(ReLU, Accuracy)
{
float negativeSlope = get<0>(GetParam());
LayerParams lp;
lp.set("negative_slope", negativeSlope);
lp.type = "ReLU";
lp.name = "testLayer";
testInPlaceActivation(lp);
}
ResizeNearestNeighborLayerImpl(const LayerParams& params)
{
setParamsFrom(params);
CV_Assert(params.has("width"), params.has("height"));
outWidth = params.get<float>("width");
outHeight = params.get<float>("height");
alignCorners = params.get<bool>("align_corners", false);
if (alignCorners)
CV_Error(Error::StsNotImplemented, "Nearest neighborhood resize with align_corners=true is not implemented");
}
bool ONNXImporter::isCeilMode(const LayerParams& layerParams) {
if (!layerParams.has("pad_mode")) {
if (layerParams.has("pad_h")) {
return layerParams.get<int>("pad_h") != layerParams.get<int>("pad_b") ||
layerParams.get<int>("pad_w") != layerParams.get<int>("pad_r");
}
else
return false; // all pads == 0
}
return true;
}
TEST_P(Concat, Accuracy)
{
Vec3i inSize = get<0>(GetParam());
Vec3i numChannels = get<1>(GetParam());
Net net;
std::vector<int> convLayerIds;
convLayerIds.reserve(numChannels.channels);
for (int i = 0, n = numChannels.channels; i < n; ++i)
{
if (!numChannels[i])
break;
Mat weights({numChannels[i], inSize[0], 1, 1}, CV_32F);
randu(weights, -1.0f, 1.0f);
LayerParams convParam;
convParam.set("kernel_w", 1);
convParam.set("kernel_h", 1);
convParam.set("num_output", numChannels[i]);
convParam.set("bias_term", false);
convParam.type = "Convolution";
std::ostringstream ss;
ss << "convLayer" << i;
convParam.name = ss.str();
convParam.blobs.push_back(weights);
int layerId = net.addLayer(convParam.name, convParam.type, convParam);
convLayerIds.push_back(layerId);
net.connect(0, 0, layerId, 0);
}
LayerParams concatParam;
concatParam.type = "Concat";
concatParam.name = "testLayer";
int concatId = net.addLayer(concatParam.name, concatParam.type, concatParam);
net.connect(0, 0, concatId, 0);
for (int i = 0; i < convLayerIds.size(); ++i)
{
net.connect(convLayerIds[i], 0, concatId, i + 1);
}
Mat input({1, inSize[0], inSize[1], inSize[2]}, CV_32F);
randu(input, -1.0f, 1.0f);
net.setInput(input);
Mat outputDefault = net.forward(concatParam.name).clone();
net.setPreferableBackend(DNN_BACKEND_HALIDE);
Mat outputHalide = net.forward(concatParam.name).clone();
normAssert(outputDefault, outputHalide);
}
bool getParameterDict(const LayerParams ¶ms,
const std::string ¶meterName,
DictValue& result)
{
if (!params.has(parameterName))
{
return false;
}
result = params.get(parameterName);
return true;
}
TEST_P(Power, Accuracy)
{
float power = get<0>(GetParam())[0];
float scale = get<0>(GetParam())[1];
float shift = get<0>(GetParam())[2];
LayerParams lp;
lp.set("power", power);
lp.set("scale", scale);
lp.set("shift", shift);
lp.type = "Power";
lp.name = "testLayer";
testInPlaceActivation(lp);
}
Ptr<Layer> createLayerFromCaffe<PoolingLayer>(LayerParams ¶ms)
{
int type = PoolingLayer::MAX;
Size kernel, stride, pad;
bool globalPooling;
cv::String padMode;
if (params.has("pool"))
{
String pool = params.get<String>("pool").toLowerCase();
if (pool == "max")
type = PoolingLayer::MAX;
else if (pool == "ave")
type = PoolingLayer::AVE;
else if (pool == "stochastic")
type = PoolingLayer::STOCHASTIC;
else
CV_Error(Error::StsBadArg, "Unknown pooling type \"" + pool + "\"");
}
getPoolingKernelParams(params, kernel.height, kernel.width, globalPooling,
pad.height, pad.width, stride.height, stride.width, padMode);
//getCaffeConvParams(params, kernel, pad, stride);
if (!globalPooling)
return Ptr<Layer>(PoolingLayer::create(type, kernel, stride, pad, padMode));
else
return Ptr<Layer>(PoolingLayer::createGlobal(type));
}
ConvolutionLayer::ConvolutionLayer(LayerParams ¶ms) : Layer(params)
{
getKernelParams(params, kerH, kerW, padH, padW, strideH, strideW);
numOutput = params.get<int>("num_output");
bias = params.get<bool>("bias_term", true);
group = params.get<int>("group", 1);
CV_Assert(numOutput % group == 0);
CV_Assert(!bias || blobs.size() == 2);
CV_Assert( bias || blobs.size() == 1);
const Blob &wgtBlob = blobs[0];
CV_Assert(wgtBlob.dims() == 4 && wgtBlob.cols() == kerW && wgtBlob.rows() == kerH);
if (bias)
{
Blob &biasBlob = blobs[1];
CV_Assert(biasBlob.total() == (size_t)numOutput);
}
//TBD
useOpenCL = params.has("use_opencl");
#if HAVE_CBLAS
{
if (getBlasThreads() != cv::getThreadNum())
{
setBlasThreads(cv::getThreadNum());
}
}
#endif
}
TEST_P(Scale, Accuracy)
{
bool hasBias = get<0>(GetParam());
LayerParams lp;
lp.set("bias_term", hasBias);
lp.type = "Scale";
lp.name = "testLayer";
lp.blobs.push_back(Mat({kNumChannels}, CV_32F));
randu(lp.blobs[0], -1.0f, 1.0f);
if (hasBias)
{
lp.blobs.push_back(Mat({kNumChannels}, CV_32F));
randu(lp.blobs[1], -1.0f, 1.0f);
}
testInPlaceActivation(lp);
}
Ptr<Layer> createLayerFromCaffe<SliceLayer>(LayerParams& params)
{
int axis = params.get<int>("axis", 1);
if (!params.has("slice_point"))
{
return Ptr<Layer>(SliceLayer::create(axis));
}
else
{
const DictValue &indicesValue = params.get("slice_point");
std::vector<int> sliceIndices(indicesValue.size());
for (int i = 0; i < indicesValue.size(); i++)
sliceIndices[i] = indicesValue.get<int>(i);
return Ptr<Layer>(SliceLayer::create(axis, sliceIndices));
}
}
ReshapeLayerImpl(const LayerParams& params)
{
setParamsFrom(params);
int axis = params.get<int>("axis", 0);
int numAxes = params.get<int>("num_axes", -1);
CV_Assert(numAxes >= -1);
newShapeRange = (numAxes == -1) ? Range(axis, INT_MAX) : Range(axis, axis + numAxes);
newShapeDesc.clear();
if (params.has("dim"))
{
const DictValue ¶mShape = params.get("dim");
int i, dims = paramShape.size();
newShapeDesc.resize(dims);
for (i = 0; i < dims; i++)
newShapeDesc[i] = paramShape.get<int>(i);
}
}
////////////////////////////////////////////////////////////////////////////////
// Padding
////////////////////////////////////////////////////////////////////////////////
TEST(Padding_Halide, Accuracy)
{
static const int kNumRuns = 10;
std::vector<int> paddings(8);
for (int t = 0; t < kNumRuns; ++t)
{
for (int i = 0; i < paddings.size(); ++i)
paddings[i] = rand() % 5;
LayerParams lp;
lp.set("paddings", DictValue::arrayInt<int*>(&paddings[0], paddings.size()));
lp.type = "Padding";
lp.name = "testLayer";
Mat input({1 + rand() % 10, 1 + rand() % 10, 1 + rand() % 10, 1 + rand() % 10}, CV_32F);
test(lp, input);
}
}
PaddingLayerImpl(const LayerParams ¶ms)
{
setParamsFrom(params);
paddingValue = params.get<float>("value", 0);
inputDims = params.get<int>("input_dims", -1);
paddingType = params.get<String>("type", "constant");
CV_Assert(params.has("paddings"));
const DictValue& paddingsParam = params.get("paddings");
CV_Assert((paddingsParam.size() & 1) == 0);
paddings.resize(paddingsParam.size() / 2);
for (int i = 0; i < paddings.size(); ++i)
{
paddings[i].first = paddingsParam.get<int>(i * 2); // Pad before.
paddings[i].second = paddingsParam.get<int>(i * 2 + 1); // Pad after.
CV_Assert(paddings[i].first >= 0, paddings[i].second >= 0);
}
}
CropLayerImpl(const LayerParams& params)
{
setParamsFrom(params);
startAxis = params.get<int>("axis", 2);
const DictValue *paramOffset = params.ptr("offset");
if (paramOffset)
{
for (int i = 0; i < paramOffset->size(); i++)
offset.push_back(paramOffset->get<int>(i));
}
}
Ptr<Layer> createLayerFromCaffe<ReshapeLayer>(LayerParams ¶ms)
{
int axis = params.get<int>("axis", 0);
int numAxes = params.get<int>("num_axes", -1);
bool enableReordering = params.get<bool>("reorder_dims", false);
CV_Assert(numAxes >= -1);
Range applyingRange = (numAxes == -1) ? Range(axis, INT_MAX) : Range(axis, axis + numAxes);
Shape newShape;
if (params.has("dim"))
{
const DictValue ¶mShape = params.get("dim");
newShape = Shape::all(paramShape.size());
for (int i = 0; i < paramShape.size(); i++)
newShape[i] = paramShape.get<int>(i);
}
else
newShape = Shape::all(0);
return Ptr<Layer>(ReshapeLayer::create(newShape, applyingRange, enableReordering));
}
SplitLayerImpl(const LayerParams ¶ms)
{
setParamsFrom(params);
//TODO: maybe "top_count" param is useless because it can be determined by output connections number
if (params.has("top_count"))
{
outputsCount = params.get<int>("top_count");
CV_Assert(outputsCount >= 0);
}
else
{
outputsCount = -1;
}
}
Ptr<Layer> createLayerFromCaffe<CropLayer>(LayerParams& params)
{
int start_axis = params.get<int>("axis", 2);
DictValue *paramOffset = params.ptr("offset");
std::vector<int> offset;
if (paramOffset)
{
for (int i = 0; i < paramOffset->size(); i++)
offset.push_back(paramOffset->get<int>(i));
}
return Ptr<Layer>(CropLayer::create(start_axis, offset));
}
TEST_P(FullyConnected, Accuracy)
{
int inChannels = get<0>(GetParam());
Size inSize = get<1>(GetParam());
int outChannels = get<2>(GetParam());
bool hasBias = get<3>(GetParam());
Mat weights(outChannels, inChannels * inSize.height * inSize.width, CV_32F);
randu(weights, -1.0f, 1.0f);
Mat bias(1, outChannels, CV_32F);
randu(bias, -1.0f, 1.0f);
LayerParams lp;
lp.set("num_output", outChannels);
lp.set("bias_term", hasBias);
lp.blobs.push_back(weights);
lp.blobs.push_back(bias);
lp.type = "InnerProduct";
lp.name = "testLayer";
Mat input({1, inChannels, inSize.height, inSize.width}, CV_32F);
test(lp, input);
}
//////////////////////////////////////////////////////////////////////////////
// Max pooling - unpooling
//////////////////////////////////////////////////////////////////////////////
TEST(MaxPoolUnpool_Halide, Accuracy)
{
LayerParams pool;
pool.set("pool", "max");
pool.set("kernel_w", 2);
pool.set("kernel_h", 2);
pool.set("stride_w", 2);
pool.set("stride_h", 2);
pool.set("pad_w", 0);
pool.set("pad_h", 0);
pool.type = "Pooling";
pool.name = "testPool";
LayerParams unpool;
unpool.set("pool_k_w", 2);
unpool.set("pool_k_h", 2);
unpool.set("pool_stride_w", 2);
unpool.set("pool_stride_h", 2);
unpool.set("pool_pad_w", 0);
unpool.set("pool_pad_h", 0);
unpool.type = "MaxUnpool";
unpool.name = "testUnpool";
Net net;
int poolId = net.addLayer(pool.name, pool.type, pool);
net.connect(0, 0, poolId, 0);
int unpoolId = net.addLayer(unpool.name, unpool.type, unpool);
net.connect(poolId, 0, unpoolId, 0);
net.connect(poolId, 1, unpoolId, 1);
Mat input({1, 1, 4, 4}, CV_32F);
randu(input, -1.0f, 1.0f);
net.setInput(input);
Mat outputDefault = net.forward("testUnpool").clone();
net.setPreferableBackend(DNN_BACKEND_HALIDE);
net.setInput(input);
Mat outputHalide = net.forward("testUnpool").clone();
normAssert(outputDefault, outputHalide);
}
TEST_P(BatchNorm, Accuracy)
{
bool hasWeights = get<0>(GetParam());
bool hasBias = get<1>(GetParam());
float epsilon = get<2>(GetParam());
LayerParams lp;
lp.set("has_weight", hasWeights);
lp.set("has_bias", hasBias);
lp.set("eps", epsilon);
lp.type = "BatchNorm";
lp.name = "testLayer";
lp.blobs.reserve(4);
for (int i = 0; i < 3; ++i)
lp.blobs.push_back(Mat({kNumChannels}, CV_32F));
if (hasBias || hasWeights)
lp.blobs.push_back(Mat({kNumChannels}, CV_32F));
for (Mat& m : lp.blobs)
randu(m, 0.0f, 1.0f);
testInPlaceActivation(lp);
}
Ptr<Layer> createLayerFromCaffe<SplitLayer>(LayerParams ¶ms)
{
int outputsCount;
//TODO: maybe "top_count" param is useless because it can be determined by output connections number
if (params.has("top_count"))
{
outputsCount = params.get<int>("top_count");
CV_Assert(outputsCount >= 0);
}
else
{
outputsCount = -1;
}
return Ptr<Layer>(SplitLayer::create(outputsCount));
}
TEST_P(MaxPooling, Accuracy)
{
int inChannels = get<0>(GetParam());
Size inSize = get<1>(GetParam());
Size kernel = get<2>(GetParam());
Size stride = get<3>(GetParam());
Size pad = get<4>(GetParam());
LayerParams lp;
lp.set("pool", "max");
lp.set("kernel_w", kernel.width);
lp.set("kernel_h", kernel.height);
lp.set("stride_w", stride.width);
lp.set("stride_h", stride.height);
lp.set("pad_w", pad.width);
lp.set("pad_h", pad.height);
lp.type = "Pooling";
lp.name = "testLayer";
Mat input({1, inChannels, inSize.height, inSize.width}, CV_32F);
test(lp, input);
}
PriorBoxLayerImpl(const LayerParams ¶ms)
{
setParamsFrom(params);
_minSize = getParameter<unsigned>(params, "min_size");
CV_Assert(_minSize > 0);
_flip = getParameter<bool>(params, "flip");
_clip = getParameter<bool>(params, "clip");
_aspectRatios.clear();
_aspectRatios.push_back(1.);
getAspectRatios(params);
getVariance(params);
_numPriors = _aspectRatios.size();
_maxSize = -1;
if (params.has("max_size"))
{
_maxSize = params.get("max_size").get<float>(0);
CV_Assert(_maxSize > _minSize);
_numPriors += 1;
}
if (params.has("step_h") || params.has("step_w")) {
CV_Assert(!params.has("step"));
_stepY = getParameter<float>(params, "step_h");
CV_Assert(_stepY > 0.);
_stepX = getParameter<float>(params, "step_w");
CV_Assert(_stepX > 0.);
} else if (params.has("step")) {
const float step = getParameter<float>(params, "step");
CV_Assert(step > 0);
_stepY = step;
_stepX = step;
} else {
_stepY = 0;
_stepX = 0;
}
}
TEST_P(AvePooling, Accuracy)
{
int inChannels = get<0>(GetParam());
Size outSize = get<1>(GetParam());; // Input size will be computed from parameters.
Size kernel = get<2>(GetParam());
Size stride = get<3>(GetParam());
const int inWidth = (outSize.width - 1) * stride.width + kernel.width;
const int inHeight = (outSize.height - 1) * stride.height + kernel.height;
LayerParams lp;
lp.set("pool", "ave");
lp.set("kernel_w", kernel.width);
lp.set("kernel_h", kernel.height);
lp.set("stride_w", stride.width);
lp.set("stride_h", stride.height);
lp.type = "Pooling";
lp.name = "testLayer";
Mat input({1, inChannels, inHeight, inWidth}, CV_32F);
test(lp, input);
}
TEST_P(LRN, Accuracy)
{
int inChannels = get<0>(GetParam())[0];
Size inSize = Size(get<0>(GetParam())[1], get<0>(GetParam())[2]);
int localSize = get<1>(GetParam());
float alpha = get<2>(GetParam())[0];
float beta = get<2>(GetParam())[1];
float bias = get<2>(GetParam())[2];
bool normBySize = get<3>(GetParam());
std::string nrmType = get<4>(GetParam());
LayerParams lp;
lp.set("norm_region", nrmType);
lp.set("local_size", localSize);
lp.set("alpha", alpha);
lp.set("beta", beta);
lp.set("bias", bias);
lp.set("norm_by_size", normBySize);
lp.type = "LRN";
lp.name = "testLayer";
Mat input({1, inChannels, inSize.height, inSize.width}, CV_32F);
test(lp, input);
}
PriorBoxLayerImpl(const LayerParams ¶ms)
: _boxWidth(0), _boxHeight(0)
{
setParamsFrom(params);
_minSize = getParameter<float>(params, "min_size", 0, false, 0);
_flip = getParameter<bool>(params, "flip", 0, false, true);
_clip = getParameter<bool>(params, "clip", 0, false, true);
_bboxesNormalized = getParameter<bool>(params, "normalized_bbox", 0, false, true);
_scales.clear();
_aspectRatios.clear();
getAspectRatios(params);
getVariance(params);
getParams("scales", params, &_scales);
getParams("width", params, &_widths);
getParams("height", params, &_heights);
_explicitSizes = !_widths.empty();
CV_Assert(_widths.size() == _heights.size());
if (_explicitSizes)
{
CV_Assert(_aspectRatios.empty(), !params.has("min_size"), !params.has("max_size"));
_numPriors = _widths.size();
}
else
{
CV_Assert(!_aspectRatios.empty(), _minSize > 0);
_numPriors = _aspectRatios.size() + 1; // + 1 for an aspect ratio 1.0
}
_maxSize = -1;
if (params.has("max_size"))
{
_maxSize = params.get("max_size").get<float>(0);
CV_Assert(_maxSize > _minSize);
_numPriors += 1;
}
if (params.has("step_h") || params.has("step_w")) {
CV_Assert(!params.has("step"));
_stepY = getParameter<float>(params, "step_h");
CV_Assert(_stepY > 0.);
_stepX = getParameter<float>(params, "step_w");
CV_Assert(_stepX > 0.);
} else if (params.has("step")) {
const float step = getParameter<float>(params, "step");
CV_Assert(step > 0);
_stepY = step;
_stepX = step;
} else {
_stepY = 0;
_stepX = 0;
}
if (params.has("offset_h") || params.has("offset_w"))
{
CV_Assert(!params.has("offset"), params.has("offset_h"), params.has("offset_w"));
getParams("offset_h", params, &_offsetsY);
getParams("offset_w", params, &_offsetsX);
CV_Assert(_offsetsX.size() == _offsetsY.size());
_numPriors *= std::max((size_t)1, 2 * (_offsetsX.size() - 1));
}
else
{
float offset = getParameter<float>(params, "offset", 0, false, 0.5);
_offsetsX.assign(1, offset);
_offsetsY.assign(1, offset);
}
}
void populateNet(Net dstNet)
{
CV_TRACE_FUNCTION();
int layersSize = net.layer_size();
layerCounter.clear();
addedBlobs.clear();
addedBlobs.reserve(layersSize + 1);
//setup input layer names
std::vector<String> netInputs(net.input_size());
{
for (int inNum = 0; inNum < net.input_size(); inNum++)
{
addedBlobs.push_back(BlobNote(net.input(inNum), 0, inNum));
netInputs[inNum] = net.input(inNum);
}
}
for (int li = 0; li < layersSize; li++)
{
const caffe::LayerParameter &layer = net.layer(li);
String name = layer.name();
String type = layer.type();
LayerParams layerParams;
extractLayerParams(layer, layerParams);
extractBinaryLayerParams(layer, layerParams);
int repetitions = layerCounter[name]++;
if (repetitions)
name += String("_") + toString(repetitions);
if (type == "Input")
{
for (int outNum = 0; outNum < layer.top_size(); outNum++)
{
addOutput(layer, 0, outNum);
addedBlobs.back().outNum = netInputs.size();
netInputs.push_back(addedBlobs.back().name);
}
continue;
}
else if (type == "BatchNorm")
{
if (!layerParams.get<bool>("use_global_stats", true))
{
CV_Assert_N(layer.bottom_size() == 1, layer.top_size() == 1);
LayerParams mvnParams;
mvnParams.set("eps", layerParams.get<float>("eps", 1e-5));
std::string mvnName = name + "/mvn";
int repetitions = layerCounter[mvnName]++;
if (repetitions)
mvnName += String("_") + toString(repetitions);
int mvnId = dstNet.addLayer(mvnName, "MVN", mvnParams);
addInput(layer.bottom(0), mvnId, 0, dstNet);
addOutput(layer, mvnId, 0);
net.mutable_layer(li)->set_bottom(0, layer.top(0));
layerParams.blobs[0].setTo(0); // mean
layerParams.blobs[1].setTo(1); // std
}
}
else if ("ConvolutionDepthwise" == type)
{
type = "Convolution";
}
int id = dstNet.addLayer(name, type, layerParams);
for (int inNum = 0; inNum < layer.bottom_size(); inNum++)
addInput(layer.bottom(inNum), id, inNum, dstNet);
for (int outNum = 0; outNum < layer.top_size(); outNum++)
addOutput(layer, id, outNum);
}
dstNet.setInputsNames(netInputs);
addedBlobs.clear();
}
void ONNXImporter::populateNet(Net dstNet)
{
CV_Assert(model_proto.has_graph());
opencv_onnx::GraphProto graph_proto = model_proto.graph();
std::map<std::string, Mat> constBlobs = getGraphTensors(graph_proto);
// List of internal blobs shapes.
std::map<std::string, MatShape> outShapes;
// Add all the inputs shapes. It includes as constant blobs as network's inputs shapes.
for (int i = 0; i < graph_proto.input_size(); ++i)
{
opencv_onnx::ValueInfoProto valueInfoProto = graph_proto.input(i);
CV_Assert(valueInfoProto.has_type());
opencv_onnx::TypeProto typeProto = valueInfoProto.type();
CV_Assert(typeProto.has_tensor_type());
opencv_onnx::TypeProto::Tensor tensor = typeProto.tensor_type();
CV_Assert(tensor.has_shape());
opencv_onnx::TensorShapeProto tensorShape = tensor.shape();
MatShape inpShape(tensorShape.dim_size());
for (int j = 0; j < inpShape.size(); ++j)
{
inpShape[j] = tensorShape.dim(j).dim_value();
}
outShapes[valueInfoProto.name()] = inpShape;
}
std::string framework_name;
if (model_proto.has_producer_name()) {
framework_name = model_proto.producer_name();
}
// create map with network inputs (without const blobs)
std::map<std::string, LayerInfo> layer_id;
std::map<std::string, LayerInfo>::iterator layerId;
std::map<std::string, MatShape>::iterator shapeIt;
// fill map: push layer name, layer id and output id
std::vector<String> netInputs;
for (int j = 0; j < graph_proto.input_size(); j++)
{
const std::string& name = graph_proto.input(j).name();
if (constBlobs.find(name) == constBlobs.end()) {
netInputs.push_back(name);
layer_id.insert(std::make_pair(name, LayerInfo(0, netInputs.size() - 1)));
}
}
dstNet.setInputsNames(netInputs);
int layersSize = graph_proto.node_size();
LayerParams layerParams;
opencv_onnx::NodeProto node_proto;
for(int li = 0; li < layersSize; li++)
{
node_proto = graph_proto.node(li);
layerParams = getLayerParams(node_proto);
CV_Assert(node_proto.output_size() >= 1);
layerParams.name = node_proto.output(0);
std::string layer_type = node_proto.op_type();
layerParams.type = layer_type;
if (layer_type == "MaxPool")
{
layerParams.type = "Pooling";
layerParams.set("pool", "MAX");
layerParams.set("ceil_mode", isCeilMode(layerParams));
}
else if (layer_type == "AveragePool")
{
layerParams.type = "Pooling";
layerParams.set("pool", "AVE");
layerParams.set("ceil_mode", isCeilMode(layerParams));
layerParams.set("ave_pool_padded_area", framework_name == "pytorch");
}
else if (layer_type == "GlobalAveragePool")
{
layerParams.type = "Pooling";
layerParams.set("pool", "AVE");
layerParams.set("global_pooling", true);
}
else if (layer_type == "Add" || layer_type == "Sum")
{
if (layer_id.find(node_proto.input(1)) == layer_id.end())
{
Mat blob = getBlob(node_proto, constBlobs, 1);
blob = blob.reshape(1, 1);
if (blob.total() == 1) {
layerParams.type = "Power";
layerParams.set("shift", blob.at<float>(0));
}
else {
layerParams.type = "Scale";
layerParams.set("bias_term", true);
layerParams.blobs.push_back(blob);
}
}
else {
layerParams.type = "Eltwise";
}
}
else if (layer_type == "Sub")
{
Mat blob = getBlob(node_proto, constBlobs, 1);
if (blob.total() == 1) {
layerParams.type = "Power";
layerParams.set("shift", -blob.at<float>(0));
}
else {
layerParams.type = "Scale";
layerParams.set("has_bias", true);
layerParams.blobs.push_back(-1.0f * blob.reshape(1, 1));
}
}
else if (layer_type == "Div")
{
Mat blob = getBlob(node_proto, constBlobs, 1);
CV_Assert_N(blob.type() == CV_32F, blob.total());
if (blob.total() == 1)
{
layerParams.set("scale", 1.0f / blob.at<float>(0));
layerParams.type = "Power";
}
else
{
layerParams.type = "Scale";
divide(1.0, blob, blob);
layerParams.blobs.push_back(blob);
layerParams.set("bias_term", false);
}
}
else if (layer_type == "Constant")
{
CV_Assert(node_proto.input_size() == 0);
CV_Assert(layerParams.blobs.size() == 1);
constBlobs.insert(std::make_pair(layerParams.name, layerParams.blobs[0]));
continue;
}
else if (layer_type == "ImageScaler")
{
const float scale = layerParams.has("scale") ? layerParams.get<float>("scale") : 1.0f;
layerParams.erase("scale");
if (layerParams.has("bias"))
{
layerParams.type = "Scale";
layerParams.blobs.push_back(
Mat(Size(1, layerParams.get("bias").size()), CV_32FC1, scale));
layerParams.set("bias_term", true);
Mat bias(1, layerParams.get("bias").size(), CV_32FC1);
for (int j = 0; j < bias.total(); j++) {
bias.at<float>(0, j) = layerParams.get("bias").getRealValue(j);
}
layerParams.blobs.push_back(bias);
layerParams.erase("bias");
}
else {
layerParams.set("scale", scale);
layerParams.type = "Power";
}
}
else if (layer_type == "LeakyRelu")
{
layerParams.type = "ReLU";
replaceLayerParam(layerParams, "alpha", "negative_slope");
}
else if (layer_type == "LRN")
{
replaceLayerParam(layerParams, "size", "local_size");
}
else if (layer_type == "BatchNormalization")
{
if (node_proto.input_size() != 5)
CV_Error(Error::StsNotImplemented,
"Expected input, scale, bias, mean and var");
layerParams.type = "BatchNorm";
replaceLayerParam(layerParams, "epsilon", "eps");
replaceLayerParam(layerParams, "spatial", "use_global_stats");
Mat meanData = getBlob(node_proto, constBlobs, 3);
Mat stdData = getBlob(node_proto, constBlobs, 4);
layerParams.blobs.push_back(meanData);
layerParams.blobs.push_back(stdData);
if (!node_proto.input(1).empty()) {
layerParams.set("has_weight", true);
layerParams.blobs.push_back(getBlob(node_proto, constBlobs, 1)); // weightData
} else {
layerParams.set("has_weight", false);
}
if (!node_proto.input(2).empty()) {
layerParams.set("has_bias", true);
layerParams.blobs.push_back(getBlob(node_proto, constBlobs, 2)); // biasData
} else {
layerParams.set("has_bias", false);
}
}
else if (layer_type == "Gemm")
{
CV_Assert(node_proto.input_size() >= 2);
layerParams.type = "InnerProduct";
Mat weights = getBlob(node_proto, constBlobs, 1);
int ind_num_out = 0;
if (layerParams.has("transB") && !layerParams.get<int>("transB")) {
transpose(weights, weights);
ind_num_out = 1;
}
layerParams.blobs.push_back(weights);
if (node_proto.input_size() == 3) {
Mat bias = getBlob(node_proto, constBlobs, 2);
layerParams.blobs.push_back(bias);
}
layerParams.set("num_output", layerParams.blobs[0].size[ind_num_out]);
layerParams.set("bias_term", node_proto.input_size() == 3);
}
else if (layer_type == "MatMul")
{
CV_Assert(node_proto.input_size() == 2);
layerParams.type = "InnerProduct";
Mat blob = getBlob(node_proto, constBlobs, 1);
layerParams.blobs.push_back(blob.t());
layerParams.set("bias_term", false);
layerParams.set("num_output", layerParams.blobs[0].size[0]);
}
else if (layer_type == "Mul")
{
CV_Assert(node_proto.input_size() == 2);
if (layer_id.find(node_proto.input(1)) == layer_id.end()) {
Mat blob = getBlob(node_proto, constBlobs, 1);
blob = blob.reshape(1, 1);
if (blob.total() == 1) {
layerParams.set("scale", blob.at<float>(0));
layerParams.type = "Power";
}
else {
layerParams.blobs.push_back(blob);
layerParams.type = "Scale";
}
}
else {
layerParams.type = "Eltwise";
layerParams.set("operation", "prod");
}
}
else if (layer_type == "Conv")
{
CV_Assert(node_proto.input_size() >= 2);
layerParams.type = "Convolution";
for (int j = 1; j < node_proto.input_size(); j++) {
layerParams.blobs.push_back(getBlob(node_proto, constBlobs, j));
}
layerParams.set("num_output", layerParams.blobs[0].size[0]);
layerParams.set("bias_term", node_proto.input_size() == 3);
}
else if (layer_type == "ConvTranspose")
{
CV_Assert(node_proto.input_size() >= 2);
layerParams.type = "Deconvolution";
for (int j = 1; j < node_proto.input_size(); j++) {
layerParams.blobs.push_back(getBlob(node_proto, constBlobs, j));
}
layerParams.set("num_output", layerParams.blobs[0].size[1]);
layerParams.set("bias_term", node_proto.input_size() == 3);
}
else if (layer_type == "Transpose")
{
layerParams.type = "Permute";
replaceLayerParam(layerParams, "perm", "order");
}
else if (layer_type == "Unsqueeze")
{
CV_Assert(node_proto.input_size() == 1);
Mat input = getBlob(node_proto, constBlobs, 0);
DictValue axes = layerParams.get("axes");
std::vector<int> dims;
for (int j = 0; j < input.dims; j++) {
dims.push_back(input.size[j]);
}
CV_Assert(axes.getIntValue(axes.size()-1) <= dims.size());
for (int j = 0; j < axes.size(); j++) {
dims.insert(dims.begin() + axes.getIntValue(j), 1);
}
Mat out = input.reshape(0, dims);
constBlobs.insert(std::make_pair(layerParams.name, out));
continue;
}
else if (layer_type == "Reshape")
{
CV_Assert(node_proto.input_size() == 2 || layerParams.has("shape"));
if (node_proto.input_size() == 2) {
Mat blob = getBlob(node_proto, constBlobs, 1);
CV_Assert(blob.type() == CV_32SC1);
if (layer_id.find(node_proto.input(0)) == layer_id.end()) {
Mat input = getBlob(node_proto, constBlobs, 0);
Mat out = input.reshape(0, static_cast<std::vector<int> >(blob));
constBlobs.insert(std::make_pair(layerParams.name, out));
continue;
}
layerParams.set("dim", DictValue::arrayInt<int*>(
blob.ptr<int>(), blob.total() ));
}
else {
DictValue shape = layerParams.get("shape");
std::vector<int> dim;
for (int j = 0; j < shape.size(); j++) {
dim.push_back(shape.getIntValue(j));
}
if (layer_id.find(node_proto.input(0)) == layer_id.end()) {
Mat input = getBlob(node_proto, constBlobs, 0);
Mat out = input.reshape(0, dim);
constBlobs.insert(std::make_pair(layerParams.name, out));
continue;
}
replaceLayerParam(layerParams, "shape", "dim");
}
}
else if (layer_type == "Pad")
{
layerParams.type = "Padding";
}
else if (layer_type == "Shape")
{
CV_Assert(node_proto.input_size() == 1);
shapeIt = outShapes.find(node_proto.input(0));
CV_Assert(shapeIt != outShapes.end());
MatShape inpShape = shapeIt->second;
Mat shapeMat(inpShape.size(), 1, CV_32S);
for (int j = 0; j < inpShape.size(); ++j)
shapeMat.at<int>(j) = inpShape[j];
shapeMat.dims = 1;
constBlobs.insert(std::make_pair(layerParams.name, shapeMat));
continue;
}
else if (layer_type == "Gather")
{
CV_Assert(node_proto.input_size() == 2);
CV_Assert(layerParams.has("axis"));
Mat input = getBlob(node_proto, constBlobs, 0);
Mat indexMat = getBlob(node_proto, constBlobs, 1);
CV_Assert_N(indexMat.type() == CV_32S, indexMat.total() == 1);
int index = indexMat.at<int>(0);
int axis = layerParams.get<int>("axis");
std::vector<cv::Range> ranges(input.dims, Range::all());
ranges[axis] = Range(index, index + 1);
Mat out = input(ranges);
constBlobs.insert(std::make_pair(layerParams.name, out));
continue;
}
else if (layer_type == "Concat")
{
bool hasVariableInps = false;
for (int i = 0; i < node_proto.input_size(); ++i)
{
if (layer_id.find(node_proto.input(i)) != layer_id.end())
{
hasVariableInps = true;
break;
}
}
if (!hasVariableInps)
{
std::vector<Mat> inputs(node_proto.input_size()), concatenated;
for (size_t i = 0; i < inputs.size(); ++i)
{
inputs[i] = getBlob(node_proto, constBlobs, i);
}
Ptr<Layer> concat = ConcatLayer::create(layerParams);
runLayer(concat, inputs, concatenated);
CV_Assert(concatenated.size() == 1);
constBlobs.insert(std::make_pair(layerParams.name, concatenated[0]));
continue;
}
}
else
{
for (int j = 0; j < node_proto.input_size(); j++) {
if (layer_id.find(node_proto.input(j)) == layer_id.end())
layerParams.blobs.push_back(getBlob(node_proto, constBlobs, j));
}
}
int id = dstNet.addLayer(layerParams.name, layerParams.type, layerParams);
layer_id.insert(std::make_pair(layerParams.name, LayerInfo(id, 0)));
std::vector<MatShape> layerInpShapes, layerOutShapes, layerInternalShapes;
for (int j = 0; j < node_proto.input_size(); j++) {
layerId = layer_id.find(node_proto.input(j));
if (layerId != layer_id.end()) {
dstNet.connect(layerId->second.layerId, layerId->second.outputId, id, j);
// Collect input shapes.
shapeIt = outShapes.find(node_proto.input(j));
CV_Assert(shapeIt != outShapes.end());
layerInpShapes.push_back(shapeIt->second);
}
}
// Compute shape of output blob for this layer.
Ptr<Layer> layer = dstNet.getLayer(id);
layer->getMemoryShapes(layerInpShapes, 0, layerOutShapes, layerInternalShapes);
CV_Assert(!layerOutShapes.empty());
outShapes[layerParams.name] = layerOutShapes[0];
}
}
