Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Image segmentation cpu bugfix and onnx output update #1783

Open
wants to merge 4 commits into
base: develop
Choose a base branch
from
+28 −51
Open

Image segmentation cpu bugfix and onnx output update #1783

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
d79b49d
Update model output processing
demoulinv Nov 21, 2024
92f919f
Remove cuda output allocator
demoulinv Nov 21, 2024
77a5760
bugfix in terminate function
demoulinv Nov 22, 2024
7c03abb
Let onnxRuntime handles output memory
demoulinv Jan 22, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
66 changes: 24 additions & 42 deletions src/aliceVision/segmentation/segmentation.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,12 +67,7 @@ bool Segmentation::initialize()
_ortSession = std::make_unique<Ort::Session>(*_ortEnvironment, _parameters.modelWeights.c_str(), ortSessionOptions);
#endif

Ort::MemoryInfo memInfoCuda("Cuda", OrtAllocatorType::OrtArenaAllocator, 0, OrtMemType::OrtMemTypeDefault);
Ort::Allocator cudaAllocator(*_ortSession, memInfoCuda);

_output.resize(_parameters.classes.size() * _parameters.modelHeight * _parameters.modelWidth);
_cudaInput = cudaAllocator.Alloc(_output.size() * sizeof(float));
_cudaOutput = cudaAllocator.Alloc(_output.size() * sizeof(float));
#endif
}
else
Expand All @@ -88,18 +83,6 @@ bool Segmentation::initialize()
return true;
}

bool Segmentation::terminate()
{
#if ALICEVISION_IS_DEFINED(ALICEVISION_HAVE_ONNX_GPU)
Ort::MemoryInfo mem_info_cuda("Cuda", OrtAllocatorType::OrtArenaAllocator, 0, OrtMemType::OrtMemTypeDefault);
Ort::Allocator cudaAllocator(*_ortSession, mem_info_cuda);
cudaAllocator.Free(_cudaInput);
cudaAllocator.Free(_cudaOutput);
#endif

return true;
}

bool Segmentation::processImage(image::Image<IndexT>& labels, const image::Image<image::RGBfColor>& source)
{
// Todo : handle orientation and small images smaller than model input
Expand Down Expand Up @@ -244,7 +227,7 @@ bool Segmentation::mergeLabels(image::Image<ScoredLabel>& labels, image::Image<S
return true;
}

bool Segmentation::labelsFromModelOutput(image::Image<ScoredLabel>& labels, const std::vector<float>& modelOutput)
bool Segmentation::labelsFromOutputTensor(image::Image<ScoredLabel>& labels, Ort::Value& modelOutput)
{
for (int outputY = 0; outputY < _parameters.modelHeight; outputY++)
{
Expand All @@ -255,10 +238,8 @@ bool Segmentation::labelsFromModelOutput(image::Image<ScoredLabel>& labels, cons

for (int classe = 0; classe < _parameters.classes.size(); classe++)
{
int classPos = classe * _parameters.modelWidth * _parameters.modelHeight;
int pos = classPos + outputY * _parameters.modelWidth + outputX;

float val = modelOutput[pos];
const std::vector<int64_t> coords = {0,classe,outputY,outputX};
const float val = modelOutput.At<float>(coords);
if (val > maxVal)
{
maxVal = val;
Expand All @@ -281,76 +262,77 @@ bool Segmentation::processTile(image::Image<ScoredLabel>& labels, const image::I
std::vector<const char*> inputNames{"input"};
std::vector<const char*> outputNames{"output"};
std::vector<int64_t> inputDimensions = {1, 3, _parameters.modelHeight, _parameters.modelWidth};
std::vector<int64_t> outputDimensions = {1, static_cast<int64_t>(_parameters.classes.size()), _parameters.modelHeight, _parameters.modelWidth};

std::vector<float> output(_parameters.classes.size() * _parameters.modelHeight * _parameters.modelWidth);
Ort::Value outputTensors =
Ort::Value::CreateTensor<float>(memInfo, output.data(), output.size(), outputDimensions.data(), outputDimensions.size());

std::vector<float> transformedInput;
imageToPlanes(transformedInput, source);

Ort::Value inputTensors =
Ort::Value::CreateTensor<float>(memInfo, transformedInput.data(), transformedInput.size(), inputDimensions.data(), inputDimensions.size());

std::vector<Ort::Value> outTensor;

try
{
_ortSession->Run(Ort::RunOptions{nullptr}, inputNames.data(), &inputTensors, 1, outputNames.data(), &outputTensors, 1);
outTensor = _ortSession->Run(Ort::RunOptions{nullptr}, inputNames.data(), &inputTensors, 1, outputNames.data(), 1);
}
catch (const Ort::Exception& exception)
{
ALICEVISION_LOG_ERROR("ERROR running model inference: " << exception.what());
return false;
}

if (!labelsFromModelOutput(labels, output))
if (!labelsFromOutputTensor(labels, outTensor[0]))
demoulinv marked this conversation as resolved.
Show resolved Hide resolved
{
return false;
}

std::vector<float> output(_parameters.classes.size() * _parameters.modelHeight * _parameters.modelWidth);
auto *outTData = outTensor.front().GetTensorMutableData<float>();
output.assign(outTData, outTData + _parameters.classes.size() * _parameters.modelHeight * _parameters.modelWidth);

return true;
}

bool Segmentation::processTileGPU(image::Image<ScoredLabel>& labels, const image::Image<image::RGBfColor>::Base& source)
{
ALICEVISION_LOG_TRACE("Process tile using gpu");
#if ALICEVISION_IS_DEFINED(ALICEVISION_HAVE_CUDA)
Ort::MemoryInfo mem_info_cuda("Cuda", OrtAllocatorType::OrtArenaAllocator, 0, OrtMemType::OrtMemTypeDefault);
Ort::Allocator cudaAllocator(*_ortSession, mem_info_cuda);
Ort::MemoryInfo memInfo = Ort::MemoryInfo::CreateCpu(OrtAllocatorType::OrtArenaAllocator, OrtMemType::OrtMemTypeDefault);

std::vector<const char*> inputNames{"input"};
std::vector<const char*> outputNames{"output"};
std::vector<int64_t> inputDimensions = {1, 3, _parameters.modelHeight, _parameters.modelWidth};
std::vector<int64_t> outputDimensions = {1, static_cast<int64_t>(_parameters.classes.size()), _parameters.modelHeight, _parameters.modelWidth};

Ort::Value outputTensors = Ort::Value::CreateTensor<float>(
mem_info_cuda, reinterpret_cast<float*>(_cudaOutput), _output.size(), outputDimensions.data(), outputDimensions.size());

std::vector<float> transformedInput;
imageToPlanes(transformedInput, source);

cudaMemcpy(_cudaInput, transformedInput.data(), sizeof(float) * transformedInput.size(), cudaMemcpyHostToDevice);
std::vector<Ort::Value> inputTensors;
inputTensors.emplace_back(Ort::Value::CreateTensor<float>(memInfo,
transformedInput.data(),
transformedInput.size(),
inputDimensions.data(),
inputDimensions.size()));

Ort::Value inputTensors = Ort::Value::CreateTensor<float>(
mem_info_cuda, reinterpret_cast<float*>(_cudaInput), transformedInput.size(), inputDimensions.data(), inputDimensions.size());
std::vector<Ort::Value> outTensor;

try
{
_ortSession->Run(Ort::RunOptions{nullptr}, inputNames.data(), &inputTensors, 1, outputNames.data(), &outputTensors, 1);
outTensor = _ortSession->Run(Ort::RunOptions{nullptr}, inputNames.data(), inputTensors.data(), 1, outputNames.data(), 1);
}
catch (const Ort::Exception& exception)
{
ALICEVISION_LOG_ERROR("ERROR running model inference: " << exception.what());
return false;
}

cudaMemcpy(_output.data(), _cudaOutput, sizeof(float) * _output.size(), cudaMemcpyDeviceToHost);

if (!labelsFromModelOutput(labels, _output))
if (!labelsFromOutputTensor(labels, outTensor[0]))
demoulinv marked this conversation as resolved.
Show resolved Hide resolved
{
return false;
}

auto *outTData = outTensor.front().GetTensorMutableData<float>();
_output.assign(outTData, outTData + _parameters.classes.size() * _parameters.modelHeight * _parameters.modelWidth);

#endif

return true;
Expand Down
13 changes: 4 additions & 9 deletions src/aliceVision/segmentation/segmentation.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,7 +57,7 @@ class Segmentation
}
}

virtual ~Segmentation() { terminate(); }
virtual ~Segmentation() {}

/**
* Process an input image to estimate segmentation
Expand All @@ -72,11 +72,6 @@ class Segmentation
*/
bool initialize();

/**
* Onnx destruction code
*/
bool terminate();

/**
* Assume the source image is the correct size
* @param labels the output label image
Expand All @@ -86,10 +81,10 @@ class Segmentation

/**
* Transform model output to a label image
* @param labels the output labels imaage
* @param modeloutput the model output vector
* @param labels the output labels image
* @param modeloutput the model output tensor
*/
bool labelsFromModelOutput(image::Image<ScoredLabel>& labels, const std::vector<float>& modelOutput);
bool labelsFromOutputTensor(image::Image<ScoredLabel>& labels, Ort::Value& modelOutput);

/**
* Process effectively a buffer of the model input size
Expand Down
Loading