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

Merge pull request #19428 from alalek:dnn_drop_misbehaved_clamp

Browse Source
pull/19465/head^2
Alexander Alekhin 4 years ago
parent 09d2ca17cf 83aa711346
commit 4f08bb5046
11 changed files with 57 additions and 52 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 26
      modules/dnn/include/opencv2/dnn/shape_utils.hpp
  2. 2
      modules/dnn/src/dnn.cpp
  3. 10
      modules/dnn/src/layers/concat_layer.cpp
  4. 12
      modules/dnn/src/layers/flatten_layer.cpp
  5. 6
      modules/dnn/src/layers/fully_connected_layer.cpp
  6. 8
      modules/dnn/src/layers/normalize_bbox_layer.cpp
  7. 9
      modules/dnn/src/layers/reshape_layer.cpp
  8. 2
      modules/dnn/src/layers/scale_layer.cpp
  9. 8
      modules/dnn/src/layers/slice_layer.cpp
  10. 10
      modules/dnn/src/layers/softmax_layer.cpp
  11. 16
      modules/dnn/src/onnx/onnx_importer.cpp

26
modules/dnn/include/opencv2/dnn/shape_utils.hpp
Unescape View File

@ -205,21 +205,33 @@ static inline std::ostream& operator<<(std::ostream &out, const MatShape& shape)
return out;
}
inline int clamp(int ax, int dims)
/// @brief Converts axis from `[-dims; dims)` (similar to Python's slice notation) to `[0; dims)` range.
static inline
int normalize_axis(int axis, int dims)
{
return ax < 0 ? ax + dims : ax;
CV_Check(axis, axis >= -dims && axis < dims, "");
axis = (axis < 0) ? (dims + axis) : axis;
CV_DbgCheck(axis, axis >= 0 && axis < dims, "");
return axis;
}
inline int clamp(int ax, const MatShape& shape)
static inline
int normalize_axis(int axis, const MatShape& shape)
{
return clamp(ax, (int)shape.size());
return normalize_axis(axis, (int)shape.size());
}
inline Range clamp(const Range& r, int axisSize)
static inline
Range normalize_axis_range(const Range& r, int axisSize)
{
Range clamped(std::max(r.start, 0),
if (r == Range::all())
return Range(0, axisSize);
CV_CheckGE(r.start, 0, "");
Range clamped(r.start,
r.end > 0 ? std::min(r.end, axisSize) : axisSize + r.end + 1);
CV_Assert_N(clamped.start < clamped.end, clamped.end <= axisSize);
CV_DbgCheckGE(clamped.start, 0, "");
CV_CheckLT(clamped.start, clamped.end, "");
CV_CheckLE(clamped.end, axisSize, "");
return clamped;
}

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

@ -2598,7 +2598,7 @@ struct Net::Impl : public detail::NetImplBase
// the concatenation optimization is applied with batch_size > 1.
// so, for now, we only apply this optimization in the most popular
// case batch_size == 1.
int axis = clamp(concatLayer->axis, output.dims);
int axis = normalize_axis(concatLayer->axis, output.dims);
if( output.total(0, axis) == 1 )
{
size_t i, ninputs = ld.inputBlobsId.size();

10
modules/dnn/src/layers/concat_layer.cpp
Unescape View File

@ -72,7 +72,7 @@ public:
{
CV_Assert(inputs.size() > 0);
outputs.resize(1, inputs[0]);
int cAxis = clamp(axis, inputs[0]);
int cAxis = normalize_axis(axis, inputs[0]);
int axisSum = 0;
for (size_t i = 0; i < inputs.size(); i++)
@ -192,7 +192,7 @@ public:
inps.getUMatVector(inputs);
outs.getUMatVector(outputs);
int cAxis = clamp(axis, inputs[0].dims);
int cAxis = normalize_axis(axis, inputs[0].dims);
if (padding)
return false;
@ -246,7 +246,7 @@ public:
inputs_arr.getMatVector(inputs);
outputs_arr.getMatVector(outputs);
int cAxis = clamp(axis, inputs[0].dims);
int cAxis = normalize_axis(axis, inputs[0].dims);
Mat& outMat = outputs[0];
if (padding)
@ -306,7 +306,7 @@ public:
InferenceEngine::DataPtr input = infEngineDataNode(inputs[0]);
InferenceEngine::Builder::ConcatLayer ieLayer(name);
ieLayer.setAxis(clamp(axis, input->getDims().size()));
ieLayer.setAxis(normalize_axis(axis, input->getDims().size()));
ieLayer.setInputPorts(std::vector<InferenceEngine::Port>(inputs.size()));
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
}
@ -319,7 +319,7 @@ public:
{
InferenceEngine::DataPtr data = ngraphDataNode(inputs[0]);
const int numDims = data->getDims().size();
const int cAxis = clamp(axis, numDims);
const int cAxis = normalize_axis(axis, numDims);
std::vector<size_t> maxDims(numDims, 0);
CV_Assert(inputs.size() == nodes.size());

12
modules/dnn/src/layers/flatten_layer.cpp
Unescape View File

@ -82,8 +82,8 @@ public:
}
int numAxes = inputs[0].size();
int startAxis = clamp(_startAxis, numAxes);
int endAxis = clamp(_endAxis, numAxes);
int startAxis = normalize_axis(_startAxis, numAxes);
int endAxis = normalize_axis(_endAxis, numAxes);
CV_Assert(startAxis >= 0);
CV_Assert(endAxis >= startAxis && endAxis < (int)numAxes);
@ -113,8 +113,8 @@ public:
inputs_arr.getMatVector(inputs);
int numAxes = inputs[0].dims;
_startAxis = clamp(_startAxis, numAxes);
_endAxis = clamp(_endAxis, numAxes);
_startAxis = normalize_axis(_startAxis, numAxes);
_endAxis = normalize_axis(_endAxis, numAxes);
}
#ifdef HAVE_OPENCL
@ -186,8 +186,8 @@ virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >& inp
std::vector<size_t> dims = ieInpNode->get_shape();
int numAxes = dims.size();
int startAxis = clamp(_startAxis, numAxes);
int endAxis = clamp(_endAxis, numAxes);
int startAxis = normalize_axis(_startAxis, numAxes);
int endAxis = normalize_axis(_endAxis, numAxes);
CV_Assert(startAxis >= 0);
CV_Assert(endAxis >= startAxis && endAxis < numAxes);

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

@ -129,7 +129,7 @@ public:
CV_CheckEQ(blobs[0].dims, 2, "");
numOutput = blobs[0].size[0];
CV_Assert(!bias || (size_t)numOutput == blobs[1].total());
cAxis = clamp(axis, inputs[0]);
cAxis = normalize_axis(axis, inputs[0]);
}
MatShape outShape(cAxis + 1);
@ -352,7 +352,7 @@ public:
return true;
}
int axisCan = clamp(axis, inputs[0].dims);
int axisCan = normalize_axis(axis, inputs[0].dims);
int numOutput = blobs[0].size[0];
int innerSize = blobs[0].size[1];
int outerSize = total(shape(inputs[0]), 0, axisCan);
@ -473,7 +473,7 @@ public:
if (!blobs.empty())
{
int axisCan = clamp(axis, input[0].dims);
int axisCan = normalize_axis(axis, input[0].dims);
int outerSize = input[0].total(0, axisCan);
for (size_t i = 0; i < input.size(); i++)

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

@ -118,8 +118,8 @@ public:
const UMat& inp0 = inputs[0];
UMat& buffer = internals[0];
startAxis = clamp(startAxis, inp0.dims);
endAxis = clamp(endAxis, inp0.dims);
startAxis = normalize_axis(startAxis, inp0.dims);
endAxis = normalize_axis(endAxis, inp0.dims);
size_t num = total(shape(inp0.size), 0, startAxis);
size_t numPlanes = total(shape(inp0.size), startAxis, endAxis + 1);
@ -203,8 +203,8 @@ public:
const Mat& inp0 = inputs[0];
Mat& buffer = internals[0];
startAxis = clamp(startAxis, inp0.dims);
endAxis = clamp(endAxis, inp0.dims);
startAxis = normalize_axis(startAxis, inp0.dims);
endAxis = normalize_axis(endAxis, inp0.dims);
const float* inpData = inp0.ptr<float>();
float* outData = outputs[0].ptr<float>();

9
modules/dnn/src/layers/reshape_layer.cpp
Unescape View File

@ -60,14 +60,7 @@ static void computeShapeByReshapeMask(const MatShape &srcShape,
int srcShapeSize = (int)srcShape.size();
int maskShapeSize = (int)maskShape.size();
if (srcRange == Range::all())
srcRange = Range(0, srcShapeSize);
else
{
int sz = srcRange.size();
srcRange.start = clamp(srcRange.start, srcShapeSize);
srcRange.end = srcRange.end == INT_MAX ? srcShapeSize : srcRange.start + sz;
}
srcRange = normalize_axis_range(srcRange, srcShapeSize);
bool explicitMask = !maskShape.empty(); // All mask values are positive.
for (int i = 0, n = maskShape.size(); i < n && explicitMask; ++i)

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

@ -240,7 +240,7 @@ public:
numChannels = blobs[0].total();
std::vector<size_t> shape(ieInpNode0->get_shape().size(), 1);
int cAxis = clamp(axis, shape.size());
int cAxis = normalize_axis(axis, shape.size());
shape[cAxis] = numChannels;
auto node = ieInpNode0;

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

@ -146,7 +146,7 @@ public:
for (int j = 0; j < sliceRanges[i].size(); ++j)
{
if (shapesInitialized || inpShape[j] > 0)
outputs[i][j] = clamp(sliceRanges[i][j], inpShape[j]).size();
outputs[i][j] = normalize_axis_range(sliceRanges[i][j], inpShape[j]).size();
}
}
}
@ -209,7 +209,7 @@ public:
// Clamp.
for (int j = 0; j < finalSliceRanges[i].size(); ++j)
{
finalSliceRanges[i][j] = clamp(finalSliceRanges[i][j], inpShape[j]);
finalSliceRanges[i][j] = normalize_axis_range(finalSliceRanges[i][j], inpShape[j]);
}
}
@ -601,7 +601,7 @@ public:
CV_Assert(inputs.size() == 2);
MatShape dstShape = inputs[0];
int start = clamp(axis, dstShape);
int start = normalize_axis(axis, dstShape);
for (int i = start; i < dstShape.size(); i++)
{
dstShape[i] = inputs[1][i];
@ -620,7 +620,7 @@ public:
const Mat &inpSzBlob = inputs[1];
int dims = inpBlob.dims;
int start_axis = clamp(axis, dims);
int start_axis = normalize_axis(axis, dims);
std::vector<int> offset_final(dims, 0);
if (offset.size() == 1)

10
modules/dnn/src/layers/softmax_layer.cpp
Unescape View File

@ -82,7 +82,7 @@ public:
{
bool inplace = Layer::getMemoryShapes(inputs, requiredOutputs, outputs, internals);
MatShape shape = inputs[0];
int cAxis = clamp(axisRaw, shape.size());
int cAxis = normalize_axis(axisRaw, shape.size());
shape[cAxis] = 1;
internals.assign(1, shape);
return inplace;
@ -115,7 +115,7 @@ public:
UMat& src = inputs[0];
UMat& dstMat = outputs[0];
int axis = clamp(axisRaw, src.dims);
int axis = normalize_axis(axisRaw, src.dims);
if (softmaxOp.empty())
{
@ -207,7 +207,7 @@ public:
const Mat &src = inputs[0];
Mat &dst = outputs[0];
int axis = clamp(axisRaw, src.dims);
int axis = normalize_axis(axisRaw, src.dims);
size_t outerSize = src.total(0, axis), channels = src.size[axis],
innerSize = src.total(axis + 1);
@ -318,7 +318,7 @@ public:
InferenceEngine::DataPtr input = infEngineDataNode(inputs[0]);
InferenceEngine::Builder::SoftMaxLayer ieLayer(name);
ieLayer.setAxis(clamp(axisRaw, input->getDims().size()));
ieLayer.setAxis(normalize_axis(axisRaw, input->getDims().size()));
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
}
@ -329,7 +329,7 @@ public:
const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE
{
auto& ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
int axis = clamp(axisRaw, ieInpNode->get_shape().size());
int axis = normalize_axis(axisRaw, ieInpNode->get_shape().size());
auto softmax = std::make_shared<ngraph::op::v1::Softmax>(ieInpNode, axis);
if (logSoftMax)
return Ptr<BackendNode>(new InfEngineNgraphNode(std::make_shared<ngraph::op::v0::Log>(softmax)));

16
modules/dnn/src/onnx/onnx_importer.cpp
Unescape View File

@ -503,7 +503,7 @@ void ONNXImporter::handleNode(const opencv_onnx::NodeProto& node_proto_)
MatShape targetShape;
std::vector<bool> shouldDelete(inpShape.size(), false);
for (int i = 0; i < axes.size(); i++) {
int axis = clamp(axes.get<int>(i), inpShape.size());
int axis = normalize_axis(axes.get<int>(i), inpShape.size());
shouldDelete[axis] = true;
}
for (int axis = 0; axis < inpShape.size(); ++axis){
@ -515,7 +515,7 @@ void ONNXImporter::handleNode(const opencv_onnx::NodeProto& node_proto_)
if (inpShape.size() == 3 && axes.size() <= 2)
{
int axis = clamp(axes.get<int>(0), inpShape.size());
int axis = normalize_axis(axes.get<int>(0), inpShape.size());
CV_CheckNE(axis, 0, "");
LayerParams reshapeLp;
@ -539,8 +539,8 @@ void ONNXImporter::handleNode(const opencv_onnx::NodeProto& node_proto_)
avgLp.set("pool", pool);
if (axes.size() == 2)
{
CV_CheckEQ(clamp(axes.get<int>(0), inpShape.size()), 1, "Unsupported mode");
CV_CheckEQ(clamp(axes.get<int>(1), inpShape.size()), 2, "Unsupported mode");
CV_CheckEQ(normalize_axis(axes.get<int>(0), inpShape.size()), 1, "Unsupported mode");
CV_CheckEQ(normalize_axis(axes.get<int>(1), inpShape.size()), 2, "Unsupported mode");
avgLp.set("global_pooling", true);
}
else
@ -560,9 +560,9 @@ void ONNXImporter::handleNode(const opencv_onnx::NodeProto& node_proto_)
CV_Assert(axes.size() <= inpShape.size() - 2);
std::vector<int> kernel_size(inpShape.size() - 2, 1);
if (axes.size() == 1 && (clamp(axes.get<int>(0), inpShape.size()) <= 1))
if (axes.size() == 1 && (normalize_axis(axes.get<int>(0), inpShape.size()) <= 1))
{
int axis = clamp(axes.get<int>(0), inpShape.size());
int axis = normalize_axis(axes.get<int>(0), inpShape.size());
MatShape newShape = inpShape;
newShape[axis + 1] = total(newShape, axis + 1);
newShape.resize(axis + 2);
@ -584,7 +584,7 @@ void ONNXImporter::handleNode(const opencv_onnx::NodeProto& node_proto_)
else
{
for (int i = 0; i < axes.size(); i++) {
int axis = clamp(axes.get<int>(i), inpShape.size());
int axis = normalize_axis(axes.get<int>(i), inpShape.size());
CV_Assert_N(axis >= 2 + i, axis < inpShape.size());
kernel_size[axis - 2] = inpShape[axis];
}
@ -1376,7 +1376,7 @@ void ONNXImporter::handleNode(const opencv_onnx::NodeProto& node_proto_)
if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
{
Mat input = getBlob(node_proto, 0);
int axis = clamp(layerParams.get<int>("axis", 1), input.dims);
int axis = normalize_axis(layerParams.get<int>("axis", 1), input.dims);
std::vector<int> out_size(&input.size[0], &input.size[0] + axis);
out_size.push_back(input.total(axis));

Write Preview
Loading…
Cancel
Save