opencv/modules/gapi/misc/python/pyopencv_gapi.hpp

1173 lines
40 KiB
C++
Raw Permalink Normal View History

#ifndef OPENCV_GAPI_PYOPENCV_GAPI_HPP
#define OPENCV_GAPI_PYOPENCV_GAPI_HPP
#ifdef HAVE_OPENCV_GAPI
#ifdef _MSC_VER
#pragma warning(disable: 4503) // "decorated name length exceeded"
#endif
2021-03-26 19:16:26 +08:00
#include <opencv2/gapi/cpu/gcpukernel.hpp>
#include <opencv2/gapi/python/python.hpp>
// NB: Python wrapper replaces :: with _ for classes
2021-12-24 23:04:11 +08:00
using gapi_GKernelPackage = cv::GKernelPackage;
using gapi_GNetPackage = cv::gapi::GNetPackage;
using gapi_ie_PyParams = cv::gapi::ie::PyParams;
2022-05-20 19:49:15 +08:00
using gapi_onnx_PyParams = cv::gapi::onnx::PyParams;
using gapi_ov_PyParams = cv::gapi::ov::PyParams;
using gapi_wip_IStreamSource_Ptr = cv::Ptr<cv::gapi::wip::IStreamSource>;
using detail_ExtractArgsCallback = cv::detail::ExtractArgsCallback;
using detail_ExtractMetaCallback = cv::detail::ExtractMetaCallback;
using vector_GNetParam = std::vector<cv::gapi::GNetParam>;
using vector_GMat = std::vector<cv::GMat>;
using gapi_streaming_queue_capacity = cv::gapi::streaming::queue_capacity;
using GStreamerSource_OutputType = cv::gapi::wip::GStreamerSource::OutputType;
using map_string_and_int = std::map<std::string, int>;
using map_string_and_string = std::map<std::string, std::string>;
using map_string_and_string = std::map<std::string, std::string>;
using map_string_and_vector_size_t = std::map<std::string, std::vector<size_t>>;
using map_string_and_vector_float = std::map<std::string, std::vector<float>>;
using map_int_and_double = std::map<int, double>;
using ep_OpenVINO = cv::gapi::onnx::ep::OpenVINO;
using ep_DirectML = cv::gapi::onnx::ep::DirectML;
using ep_CoreML = cv::gapi::onnx::ep::CoreML;
using ep_CUDA = cv::gapi::onnx::ep::CUDA;
using ep_TensorRT = cv::gapi::onnx::ep::TensorRT;
// NB: Python wrapper generate T_U for T<U>
// This behavior is only observed for inputs
using GOpaque_bool = cv::GOpaque<bool>;
using GOpaque_int = cv::GOpaque<int>;
using GOpaque_double = cv::GOpaque<double>;
using GOpaque_float = cv::GOpaque<double>;
using GOpaque_string = cv::GOpaque<std::string>;
2021-03-26 19:16:26 +08:00
using GOpaque_Point2i = cv::GOpaque<cv::Point>;
using GOpaque_Point2f = cv::GOpaque<cv::Point2f>;
using GOpaque_Size = cv::GOpaque<cv::Size>;
using GOpaque_Rect = cv::GOpaque<cv::Rect>;
using GArray_bool = cv::GArray<bool>;
using GArray_int = cv::GArray<int>;
using GArray_double = cv::GArray<double>;
using GArray_float = cv::GArray<double>;
using GArray_string = cv::GArray<std::string>;
2021-03-26 19:16:26 +08:00
using GArray_Point2i = cv::GArray<cv::Point>;
using GArray_Point2f = cv::GArray<cv::Point2f>;
using GArray_Point3f = cv::GArray<cv::Point3f>;
using GArray_Size = cv::GArray<cv::Size>;
using GArray_Rect = cv::GArray<cv::Rect>;
using GArray_Scalar = cv::GArray<cv::Scalar>;
using GArray_Mat = cv::GArray<cv::Mat>;
using GArray_GMat = cv::GArray<cv::GMat>;
using GArray_Prim = cv::GArray<cv::gapi::wip::draw::Prim>;
// FIXME: Python wrapper generate code without namespace std,
// so it cause error: "string wasn't declared"
// WA: Create using
using std::string;
namespace cv
{
namespace detail
{
class PyObjectHolder
{
public:
PyObjectHolder(PyObject* o, bool owner = true);
PyObject* get() const;
private:
class Impl;
std::shared_ptr<Impl> m_impl;
};
} // namespace detail
} // namespace cv
class cv::detail::PyObjectHolder::Impl
{
public:
Impl(PyObject* object, bool owner);
PyObject* get() const;
~Impl();
private:
PyObject* m_object;
};
cv::detail::PyObjectHolder::Impl::Impl(PyObject* object, bool owner)
: m_object(object)
{
// NB: Become an owner of that PyObject.
// Need to store this and get access
// after the caller which provide the object is out of range.
if (owner)
{
// NB: Impossible take ownership if object is NULL.
GAPI_Assert(object);
Py_INCREF(m_object);
}
}
cv::detail::PyObjectHolder::Impl::~Impl()
{
// NB: If NULL was set, don't decrease counter.
if (m_object)
{
Py_DECREF(m_object);
}
}
PyObject* cv::detail::PyObjectHolder::Impl::get() const
{
return m_object;
}
cv::detail::PyObjectHolder::PyObjectHolder(PyObject* object, bool owner)
: m_impl(new cv::detail::PyObjectHolder::Impl{object, owner})
{
}
PyObject* cv::detail::PyObjectHolder::get() const
{
return m_impl->get();
}
template<>
PyObject* pyopencv_from(const cv::detail::PyObjectHolder& v)
{
PyObject* o = cv::util::any_cast<cv::detail::PyObjectHolder>(v).get();
Py_INCREF(o);
return o;
}
// #FIXME: Is it possible to implement pyopencv_from/pyopencv_to for generic
// cv::variant<Types...> ?
template <>
PyObject* pyopencv_from(const cv::gapi::wip::draw::Prim& prim)
{
switch (prim.index())
{
case cv::gapi::wip::draw::Prim::index_of<cv::gapi::wip::draw::Rect>():
return pyopencv_from(cv::util::get<cv::gapi::wip::draw::Rect>(prim));
case cv::gapi::wip::draw::Prim::index_of<cv::gapi::wip::draw::Text>():
return pyopencv_from(cv::util::get<cv::gapi::wip::draw::Text>(prim));
case cv::gapi::wip::draw::Prim::index_of<cv::gapi::wip::draw::Circle>():
return pyopencv_from(cv::util::get<cv::gapi::wip::draw::Circle>(prim));
case cv::gapi::wip::draw::Prim::index_of<cv::gapi::wip::draw::Line>():
return pyopencv_from(cv::util::get<cv::gapi::wip::draw::Line>(prim));
case cv::gapi::wip::draw::Prim::index_of<cv::gapi::wip::draw::Poly>():
return pyopencv_from(cv::util::get<cv::gapi::wip::draw::Poly>(prim));
case cv::gapi::wip::draw::Prim::index_of<cv::gapi::wip::draw::Mosaic>():
return pyopencv_from(cv::util::get<cv::gapi::wip::draw::Mosaic>(prim));
case cv::gapi::wip::draw::Prim::index_of<cv::gapi::wip::draw::Image>():
return pyopencv_from(cv::util::get<cv::gapi::wip::draw::Image>(prim));
}
util::throw_error(std::logic_error("Unsupported draw primitive type"));
}
template <>
PyObject* pyopencv_from(const cv::gapi::wip::draw::Prims& value)
{
return pyopencv_from_generic_vec(value);
}
template<>
bool pyopencv_to(PyObject* obj, cv::gapi::wip::draw::Prim& value, const ArgInfo&)
{
#define TRY_EXTRACT(Prim) \
if (PyObject_TypeCheck(obj, reinterpret_cast<PyTypeObject*>(pyopencv_gapi_wip_draw_##Prim##_TypePtr))) \
{ \
value = reinterpret_cast<pyopencv_gapi_wip_draw_##Prim##_t*>(obj)->v; \
return true; \
} \
TRY_EXTRACT(Rect)
TRY_EXTRACT(Text)
TRY_EXTRACT(Circle)
TRY_EXTRACT(Line)
TRY_EXTRACT(Mosaic)
TRY_EXTRACT(Image)
TRY_EXTRACT(Poly)
#undef TRY_EXTRACT
failmsg("Unsupported primitive type");
return false;
}
template <>
bool pyopencv_to(PyObject* obj, cv::gapi::wip::draw::Prims& value, const ArgInfo& info)
{
return pyopencv_to_generic_vec(obj, value, info);
}
template <>
bool pyopencv_to(PyObject* obj, cv::GMetaArg& value, const ArgInfo&)
{
#define TRY_EXTRACT(Meta) \
if (PyObject_TypeCheck(obj, \
reinterpret_cast<PyTypeObject*>(pyopencv_##Meta##_TypePtr))) \
{ \
value = reinterpret_cast<pyopencv_##Meta##_t*>(obj)->v; \
return true; \
} \
TRY_EXTRACT(GMatDesc)
TRY_EXTRACT(GScalarDesc)
TRY_EXTRACT(GArrayDesc)
TRY_EXTRACT(GOpaqueDesc)
#undef TRY_EXTRACT
failmsg("Unsupported cv::GMetaArg type");
return false;
}
template <>
bool pyopencv_to(PyObject* obj, cv::GMetaArgs& value, const ArgInfo& info)
{
return pyopencv_to_generic_vec(obj, value, info);
}
template<>
PyObject* pyopencv_from(const cv::GArg& value)
{
GAPI_Assert(value.kind != cv::detail::ArgKind::GOBJREF);
#define HANDLE_CASE(T, O) case cv::detail::OpaqueKind::CV_##T: \
{ \
return pyopencv_from(value.get<O>()); \
}
#define UNSUPPORTED(T) case cv::detail::OpaqueKind::CV_##T: break
switch (value.opaque_kind)
{
HANDLE_CASE(BOOL, bool);
HANDLE_CASE(INT, int);
HANDLE_CASE(INT64, int64_t);
HANDLE_CASE(UINT64, uint64_t);
HANDLE_CASE(DOUBLE, double);
HANDLE_CASE(FLOAT, float);
HANDLE_CASE(STRING, std::string);
HANDLE_CASE(POINT, cv::Point);
HANDLE_CASE(POINT2F, cv::Point2f);
HANDLE_CASE(POINT3F, cv::Point3f);
HANDLE_CASE(SIZE, cv::Size);
HANDLE_CASE(RECT, cv::Rect);
HANDLE_CASE(SCALAR, cv::Scalar);
HANDLE_CASE(MAT, cv::Mat);
HANDLE_CASE(UNKNOWN, cv::detail::PyObjectHolder);
HANDLE_CASE(DRAW_PRIM, cv::gapi::wip::draw::Prim);
#undef HANDLE_CASE
#undef UNSUPPORTED
}
util::throw_error(std::logic_error("Unsupported kernel input type"));
}
template<>
bool pyopencv_to(PyObject* obj, cv::GArg& value, const ArgInfo& info)
{
value = cv::GArg(cv::detail::PyObjectHolder(obj));
return true;
}
2021-03-26 19:16:26 +08:00
template <>
bool pyopencv_to(PyObject* obj, std::vector<cv::gapi::GNetParam>& value, const ArgInfo& info)
{
return pyopencv_to_generic_vec(obj, value, info);
}
2021-03-26 19:16:26 +08:00
template <>
PyObject* pyopencv_from(const std::vector<cv::gapi::GNetParam>& value)
{
return pyopencv_from_generic_vec(value);
}
2021-03-26 19:16:26 +08:00
template <>
bool pyopencv_to(PyObject* obj, std::vector<GCompileArg>& value, const ArgInfo& info)
{
return pyopencv_to_generic_vec(obj, value, info);
}
template <>
PyObject* pyopencv_from(const std::vector<GCompileArg>& value)
{
return pyopencv_from_generic_vec(value);
}
template<>
PyObject* pyopencv_from(const cv::detail::OpaqueRef& o)
{
switch (o.getKind())
{
case cv::detail::OpaqueKind::CV_BOOL : return pyopencv_from(o.rref<bool>());
case cv::detail::OpaqueKind::CV_INT : return pyopencv_from(o.rref<int>());
case cv::detail::OpaqueKind::CV_INT64 : return pyopencv_from(o.rref<int64_t>());
case cv::detail::OpaqueKind::CV_UINT64 : return pyopencv_from(o.rref<uint64_t>());
case cv::detail::OpaqueKind::CV_DOUBLE : return pyopencv_from(o.rref<double>());
case cv::detail::OpaqueKind::CV_FLOAT : return pyopencv_from(o.rref<float>());
case cv::detail::OpaqueKind::CV_STRING : return pyopencv_from(o.rref<std::string>());
case cv::detail::OpaqueKind::CV_POINT : return pyopencv_from(o.rref<cv::Point>());
case cv::detail::OpaqueKind::CV_POINT2F : return pyopencv_from(o.rref<cv::Point2f>());
case cv::detail::OpaqueKind::CV_POINT3F : return pyopencv_from(o.rref<cv::Point3f>());
case cv::detail::OpaqueKind::CV_SIZE : return pyopencv_from(o.rref<cv::Size>());
case cv::detail::OpaqueKind::CV_RECT : return pyopencv_from(o.rref<cv::Rect>());
case cv::detail::OpaqueKind::CV_UNKNOWN : return pyopencv_from(o.rref<cv::GArg>());
case cv::detail::OpaqueKind::CV_DRAW_PRIM : return pyopencv_from(o.rref<cv::gapi::wip::draw::Prim>());
case cv::detail::OpaqueKind::CV_SCALAR : break;
case cv::detail::OpaqueKind::CV_MAT : break;
}
PyErr_SetString(PyExc_TypeError, "Unsupported GOpaque type");
return NULL;
}
template <>
PyObject* pyopencv_from(const cv::detail::VectorRef& v)
{
switch (v.getKind())
{
case cv::detail::OpaqueKind::CV_BOOL : return pyopencv_from_generic_vec(v.rref<bool>());
case cv::detail::OpaqueKind::CV_INT : return pyopencv_from_generic_vec(v.rref<int>());
case cv::detail::OpaqueKind::CV_INT64 : return pyopencv_from_generic_vec(v.rref<int64_t>());
case cv::detail::OpaqueKind::CV_UINT64 : return pyopencv_from_generic_vec(v.rref<uint64_t>());
case cv::detail::OpaqueKind::CV_DOUBLE : return pyopencv_from_generic_vec(v.rref<double>());
case cv::detail::OpaqueKind::CV_FLOAT : return pyopencv_from_generic_vec(v.rref<float>());
case cv::detail::OpaqueKind::CV_STRING : return pyopencv_from_generic_vec(v.rref<std::string>());
case cv::detail::OpaqueKind::CV_POINT : return pyopencv_from_generic_vec(v.rref<cv::Point>());
case cv::detail::OpaqueKind::CV_POINT2F : return pyopencv_from_generic_vec(v.rref<cv::Point2f>());
case cv::detail::OpaqueKind::CV_POINT3F : return pyopencv_from_generic_vec(v.rref<cv::Point3f>());
case cv::detail::OpaqueKind::CV_SIZE : return pyopencv_from_generic_vec(v.rref<cv::Size>());
case cv::detail::OpaqueKind::CV_RECT : return pyopencv_from_generic_vec(v.rref<cv::Rect>());
case cv::detail::OpaqueKind::CV_SCALAR : return pyopencv_from_generic_vec(v.rref<cv::Scalar>());
case cv::detail::OpaqueKind::CV_MAT : return pyopencv_from_generic_vec(v.rref<cv::Mat>());
case cv::detail::OpaqueKind::CV_UNKNOWN : return pyopencv_from_generic_vec(v.rref<cv::GArg>());
case cv::detail::OpaqueKind::CV_DRAW_PRIM : return pyopencv_from_generic_vec(v.rref<cv::gapi::wip::draw::Prim>());
}
PyErr_SetString(PyExc_TypeError, "Unsupported GArray type");
return NULL;
}
template <>
PyObject* pyopencv_from(const GRunArg& v)
{
switch (v.index())
{
case GRunArg::index_of<cv::Mat>():
return pyopencv_from(util::get<cv::Mat>(v));
case GRunArg::index_of<cv::Scalar>():
return pyopencv_from(util::get<cv::Scalar>(v));
case GRunArg::index_of<cv::detail::VectorRef>():
return pyopencv_from(util::get<cv::detail::VectorRef>(v));
case GRunArg::index_of<cv::detail::OpaqueRef>():
return pyopencv_from(util::get<cv::detail::OpaqueRef>(v));
}
PyErr_SetString(PyExc_TypeError, "Failed to unpack GRunArgs. Index of variant is unknown");
return NULL;
}
template <typename T>
PyObject* pyopencv_from(const cv::optional<T>& opt)
{
if (!opt.has_value())
{
Py_RETURN_NONE;
}
return pyopencv_from(*opt);
}
template <>
PyObject* pyopencv_from(const GOptRunArg& v)
{
switch (v.index())
{
case GOptRunArg::index_of<cv::optional<cv::Mat>>():
return pyopencv_from(util::get<cv::optional<cv::Mat>>(v));
case GOptRunArg::index_of<cv::optional<cv::Scalar>>():
return pyopencv_from(util::get<cv::optional<cv::Scalar>>(v));
case GOptRunArg::index_of<optional<cv::detail::VectorRef>>():
return pyopencv_from(util::get<optional<cv::detail::VectorRef>>(v));
case GOptRunArg::index_of<optional<cv::detail::OpaqueRef>>():
return pyopencv_from(util::get<optional<cv::detail::OpaqueRef>>(v));
}
PyErr_SetString(PyExc_TypeError, "Failed to unpack GOptRunArg. Index of variant is unknown");
return NULL;
}
template<>
PyObject* pyopencv_from(const GRunArgs& value)
{
return value.size() == 1 ? pyopencv_from(value[0]) : pyopencv_from_generic_vec(value);
}
template<>
PyObject* pyopencv_from(const GOptRunArgs& value)
{
return value.size() == 1 ? pyopencv_from(value[0]) : pyopencv_from_generic_vec(value);
}
// FIXME: cv::variant should be wrapped once for all types.
template <>
PyObject* pyopencv_from(const cv::util::variant<cv::GRunArgs, cv::GOptRunArgs>& v)
{
using RunArgs = cv::util::variant<cv::GRunArgs, cv::GOptRunArgs>;
switch (v.index())
{
case RunArgs::index_of<cv::GRunArgs>():
return pyopencv_from(util::get<cv::GRunArgs>(v));
case RunArgs::index_of<cv::GOptRunArgs>():
return pyopencv_from(util::get<cv::GOptRunArgs>(v));
}
PyErr_SetString(PyExc_TypeError, "Failed to recognize kind of RunArgs. Index of variant is unknown");
return NULL;
}
template <typename T>
void pyopencv_to_with_check(PyObject* from, T& to, const std::string& msg = "")
{
if (!pyopencv_to(from, to, ArgInfo("", false)))
{
cv::util::throw_error(std::logic_error(msg));
}
}
template <typename T>
void pyopencv_to_generic_vec_with_check(PyObject* from,
std::vector<T>& to,
const std::string& msg = "")
{
if (!pyopencv_to_generic_vec(from, to, ArgInfo("", false)))
{
cv::util::throw_error(std::logic_error(msg));
}
}
template <typename T>
static T extract_proto_args(PyObject* py_args)
{
using namespace cv;
GProtoArgs args;
Py_ssize_t size = PyList_Size(py_args);
args.reserve(size);
for (int i = 0; i < size; ++i)
{
PyObject* item = PyList_GetItem(py_args, i);
if (PyObject_TypeCheck(item, reinterpret_cast<PyTypeObject*>(pyopencv_GScalar_TypePtr)))
{
args.emplace_back(reinterpret_cast<pyopencv_GScalar_t*>(item)->v);
}
else if (PyObject_TypeCheck(item, reinterpret_cast<PyTypeObject*>(pyopencv_GMat_TypePtr)))
{
args.emplace_back(reinterpret_cast<pyopencv_GMat_t*>(item)->v);
}
else if (PyObject_TypeCheck(item, reinterpret_cast<PyTypeObject*>(pyopencv_GOpaqueT_TypePtr)))
{
args.emplace_back(reinterpret_cast<pyopencv_GOpaqueT_t*>(item)->v.strip());
}
else if (PyObject_TypeCheck(item, reinterpret_cast<PyTypeObject*>(pyopencv_GArrayT_TypePtr)))
{
args.emplace_back(reinterpret_cast<pyopencv_GArrayT_t*>(item)->v.strip());
}
else
{
util::throw_error(std::logic_error("Unsupported type for GProtoArgs"));
}
}
return T(std::move(args));
}
static cv::detail::OpaqueRef extract_opaque_ref(PyObject* from, cv::detail::OpaqueKind kind)
{
#define HANDLE_CASE(T, O) case cv::detail::OpaqueKind::CV_##T: \
{ \
O obj{}; \
pyopencv_to_with_check(from, obj, "Failed to obtain " # O); \
return cv::detail::OpaqueRef{std::move(obj)}; \
}
#define UNSUPPORTED(T) case cv::detail::OpaqueKind::CV_##T: break
switch (kind)
{
HANDLE_CASE(BOOL, bool);
HANDLE_CASE(INT, int);
HANDLE_CASE(INT64, int64_t);
HANDLE_CASE(UINT64, uint64_t);
HANDLE_CASE(DOUBLE, double);
HANDLE_CASE(FLOAT, float);
HANDLE_CASE(STRING, std::string);
HANDLE_CASE(POINT, cv::Point);
HANDLE_CASE(POINT2F, cv::Point2f);
HANDLE_CASE(POINT3F, cv::Point3f);
HANDLE_CASE(SIZE, cv::Size);
HANDLE_CASE(RECT, cv::Rect);
HANDLE_CASE(UNKNOWN, cv::GArg);
UNSUPPORTED(SCALAR);
UNSUPPORTED(MAT);
UNSUPPORTED(DRAW_PRIM);
#undef HANDLE_CASE
#undef UNSUPPORTED
2021-03-26 19:16:26 +08:00
}
util::throw_error(std::logic_error("Unsupported type for GOpaqueT"));
}
static cv::detail::VectorRef extract_vector_ref(PyObject* from, cv::detail::OpaqueKind kind)
{
#define HANDLE_CASE(T, O) case cv::detail::OpaqueKind::CV_##T: \
{ \
std::vector<O> obj; \
pyopencv_to_generic_vec_with_check(from, obj, "Failed to obtain vector of " # O); \
return cv::detail::VectorRef{std::move(obj)}; \
}
#define UNSUPPORTED(T) case cv::detail::OpaqueKind::CV_##T: break
switch (kind)
{
HANDLE_CASE(BOOL, bool);
HANDLE_CASE(INT, int);
HANDLE_CASE(INT64, int64_t);
HANDLE_CASE(UINT64, uint64_t);
HANDLE_CASE(DOUBLE, double);
HANDLE_CASE(FLOAT, float);
HANDLE_CASE(STRING, std::string);
HANDLE_CASE(POINT, cv::Point);
HANDLE_CASE(POINT2F, cv::Point2f);
HANDLE_CASE(POINT3F, cv::Point3f);
HANDLE_CASE(SIZE, cv::Size);
HANDLE_CASE(RECT, cv::Rect);
HANDLE_CASE(SCALAR, cv::Scalar);
HANDLE_CASE(MAT, cv::Mat);
HANDLE_CASE(UNKNOWN, cv::GArg);
HANDLE_CASE(DRAW_PRIM, cv::gapi::wip::draw::Prim);
#undef HANDLE_CASE
#undef UNSUPPORTED
}
2021-03-26 19:16:26 +08:00
util::throw_error(std::logic_error("Unsupported type for GArrayT"));
}
static cv::GRunArg extract_run_arg(const cv::GTypeInfo& info, PyObject* item)
{
switch (info.shape)
{
case cv::GShape::GMAT:
{
// NB: In case streaming it can be IStreamSource or cv::Mat
if (PyObject_TypeCheck(item,
reinterpret_cast<PyTypeObject*>(pyopencv_gapi_wip_IStreamSource_TypePtr)))
{
cv::gapi::wip::IStreamSource::Ptr source =
reinterpret_cast<pyopencv_gapi_wip_IStreamSource_t*>(item)->v;
return source;
}
cv::Mat obj;
pyopencv_to_with_check(item, obj, "Failed to obtain cv::Mat");
return obj;
}
case cv::GShape::GSCALAR:
{
cv::Scalar obj;
pyopencv_to_with_check(item, obj, "Failed to obtain cv::Scalar");
return obj;
}
case cv::GShape::GOPAQUE:
{
return extract_opaque_ref(item, info.kind);
}
case cv::GShape::GARRAY:
{
return extract_vector_ref(item, info.kind);
}
case cv::GShape::GFRAME:
{
2021-03-26 19:16:26 +08:00
// NB: Isn't supported yet.
break;
}
}
util::throw_error(std::logic_error("Unsupported output shape"));
}
static cv::GRunArgs extract_run_args(const cv::GTypesInfo& info, PyObject* py_args)
{
GAPI_Assert(PyList_Check(py_args));
cv::GRunArgs args;
Py_ssize_t list_size = PyList_Size(py_args);
args.reserve(list_size);
for (int i = 0; i < list_size; ++i)
{
args.push_back(extract_run_arg(info[i], PyList_GetItem(py_args, i)));
}
return args;
}
static cv::GMetaArg extract_meta_arg(const cv::GTypeInfo& info, PyObject* item)
{
switch (info.shape)
{
case cv::GShape::GMAT:
{
cv::Mat obj;
pyopencv_to_with_check(item, obj, "Failed to obtain cv::Mat");
return cv::GMetaArg{cv::descr_of(obj)};
}
case cv::GShape::GSCALAR:
{
cv::Scalar obj;
pyopencv_to_with_check(item, obj, "Failed to obtain cv::Scalar");
return cv::GMetaArg{cv::descr_of(obj)};
}
case cv::GShape::GARRAY:
{
return cv::GMetaArg{cv::empty_array_desc()};
}
case cv::GShape::GOPAQUE:
{
return cv::GMetaArg{cv::empty_gopaque_desc()};
}
case cv::GShape::GFRAME:
{
// NB: Isn't supported yet.
break;
}
}
util::throw_error(std::logic_error("Unsupported output shape"));
}
static cv::GMetaArgs extract_meta_args(const cv::GTypesInfo& info, PyObject* py_args)
{
GAPI_Assert(PyList_Check(py_args));
cv::GMetaArgs metas;
Py_ssize_t list_size = PyList_Size(py_args);
metas.reserve(list_size);
for (int i = 0; i < list_size; ++i)
{
metas.push_back(extract_meta_arg(info[i], PyList_GetItem(py_args, i)));
}
return metas;
}
static cv::GRunArgs run_py_kernel(cv::detail::PyObjectHolder kernel,
2021-03-26 19:16:26 +08:00
const cv::gapi::python::GPythonContext &ctx)
{
const auto& ins = ctx.ins;
const auto& in_metas = ctx.in_metas;
const auto& out_info = ctx.out_info;
PyGILState_STATE gstate;
gstate = PyGILState_Ensure();
cv::GRunArgs outs;
try
{
// NB: Doesn't increase reference counter (false),
// because PyObject already have ownership.
// In case exception decrement reference counter.
cv::detail::PyObjectHolder args(
PyTuple_New(ctx.m_state.has_value() ? ins.size() + 1 : ins.size()), false);
2021-03-26 19:16:26 +08:00
for (size_t i = 0; i < ins.size(); ++i)
{
// NB: If meta is monostate then object isn't associated with G-TYPE.
2021-03-26 19:16:26 +08:00
if (cv::util::holds_alternative<cv::util::monostate>(in_metas[i]))
{
PyTuple_SetItem(args.get(), i, pyopencv_from(ins[i]));
2021-03-26 19:16:26 +08:00
continue;
}
switch (in_metas[i].index())
{
case cv::GMetaArg::index_of<cv::GMatDesc>():
PyTuple_SetItem(args.get(), i, pyopencv_from(ins[i].get<cv::Mat>()));
2021-03-26 19:16:26 +08:00
break;
case cv::GMetaArg::index_of<cv::GScalarDesc>():
PyTuple_SetItem(args.get(), i, pyopencv_from(ins[i].get<cv::Scalar>()));
2021-03-26 19:16:26 +08:00
break;
case cv::GMetaArg::index_of<cv::GOpaqueDesc>():
PyTuple_SetItem(args.get(), i, pyopencv_from(ins[i].get<cv::detail::OpaqueRef>()));
2021-03-26 19:16:26 +08:00
break;
case cv::GMetaArg::index_of<cv::GArrayDesc>():
PyTuple_SetItem(args.get(), i, pyopencv_from(ins[i].get<cv::detail::VectorRef>()));
2021-03-26 19:16:26 +08:00
break;
case cv::GMetaArg::index_of<cv::GFrameDesc>():
util::throw_error(std::logic_error("GFrame isn't supported for custom operation"));
break;
}
}
2022-05-26 00:12:51 +08:00
if (ctx.m_state.has_value())
2022-05-26 00:12:51 +08:00
{
PyTuple_SetItem(args.get(), ins.size(), pyopencv_from(ctx.m_state.value()));
2022-05-26 00:12:51 +08:00
}
// NB: Doesn't increase reference counter (false).
// In case PyObject_CallObject return NULL, do nothing in destructor.
cv::detail::PyObjectHolder result(
PyObject_CallObject(kernel.get(), args.get()), false);
if (PyErr_Occurred())
{
PyErr_PrintEx(0);
PyErr_Clear();
throw std::logic_error("Python kernel failed with error!");
}
// NB: In fact it's impossible situation, because errors were handled above.
GAPI_Assert(result.get() && "Python kernel returned NULL!");
2021-03-26 19:16:26 +08:00
if (out_info.size() == 1)
{
outs = cv::GRunArgs{extract_run_arg(out_info[0], result.get())};
}
else if (out_info.size() > 1)
{
GAPI_Assert(PyTuple_Check(result.get()));
Py_ssize_t tuple_size = PyTuple_Size(result.get());
outs.reserve(tuple_size);
for (int i = 0; i < tuple_size; ++i)
{
outs.push_back(extract_run_arg(out_info[i], PyTuple_GetItem(result.get(), i)));
}
}
else
{
// Seems to be impossible case.
GAPI_Error("InternalError");
}
2021-03-26 19:16:26 +08:00
}
catch (...)
{
PyGILState_Release(gstate);
throw;
}
PyGILState_Release(gstate);
return outs;
}
static void unpackMetasToTuple(const cv::GMetaArgs& meta,
const cv::GArgs& gargs,
cv::detail::PyObjectHolder& tuple)
{
size_t idx = 0;
for (auto&& m : meta)
{
switch (m.index())
{
case cv::GMetaArg::index_of<cv::GMatDesc>():
PyTuple_SetItem(tuple.get(), idx, pyopencv_from(cv::util::get<cv::GMatDesc>(m)));
break;
case cv::GMetaArg::index_of<cv::GScalarDesc>():
PyTuple_SetItem(tuple.get(), idx,
pyopencv_from(cv::util::get<cv::GScalarDesc>(m)));
break;
case cv::GMetaArg::index_of<cv::GArrayDesc>():
PyTuple_SetItem(tuple.get(), idx,
pyopencv_from(cv::util::get<cv::GArrayDesc>(m)));
break;
case cv::GMetaArg::index_of<cv::GOpaqueDesc>():
PyTuple_SetItem(tuple.get(), idx,
pyopencv_from(cv::util::get<cv::GOpaqueDesc>(m)));
break;
case cv::GMetaArg::index_of<cv::util::monostate>():
PyTuple_SetItem(tuple.get(), idx, pyopencv_from(gargs[idx]));
break;
case cv::GMetaArg::index_of<cv::GFrameDesc>():
util::throw_error(
std::logic_error("GFrame isn't supported for custom operation"));
break;
}
++idx;
}
}
Merge pull request #23597 from dmatveev:dm/gapi_onnx_py_integration G-API: Integration branch for ONNX & Python-related changes #23597 # Changes overview ## 1. Expose ONNX backend's Normalization and Mean-value parameters in Python * Since Python G-API bindings rely on `Generic` infer to express Inference, the `Generic` specialization of `onnx::Params` was extended with new methods to control normalization (`/255`) and mean-value; these methods were exposed in the Python bindings * Found some questionable parts in the existing API which I'd like to review/discuss (see comments) UPD: 1. Thanks to @TolyaTalamanov normalization inconsistencies have been identified with `squeezenet1.0-9` ONNX model itself; tests using these model were updated to DISABLE normalization and NOT using mean/value. 2. Questionable parts were removed and tests still pass. ### Details (taken from @TolyaTalamanov's comment): `squeezenet1.0.*onnx` - doesn't require scaling to [0,1] and mean/std because the weights of the first convolution already scaled. ONNX documentation is broken. So the correct approach to use this models is: 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 but without normalization step: ``` # DON'T DO IT: # mean_vec = np.array([0.485, 0.456, 0.406]) # stddev_vec = np.array([0.229, 0.224, 0.225]) # norm_img_data = np.zeros(img_data.shape).astype('float32') # for i in range(img_data.shape[0]): # norm_img_data[i,:,:] = (img_data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i] # # add batch channel # norm_img_data = norm_img_data.reshape(1, 3, 224, 224).astype('float32') # return norm_img_data # INSTEAD return img_data.reshape(1, 3, 224, 224) ``` 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', False) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. --- `squeezenet1.1.*onnx` - requires scaling to [0,1] and mean/std - onnx documentation is correct. 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', True) // default net.cfgMeanStd('data_0', [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. ## 2. Expose Fluid & kernel package-related functionality in Python * `cv::gapi::combine()` * `cv::GKernelPackage::size()` (mainly for testing purposes) * `cv::gapi::imgproc::fluid::kernels()` Added a test for the above. ## 3. Fixed issues with Python stateful kernel handling Fixed error message when `outMeta()` of custom python operation fails. ## 4. Fixed various issues in Python tests 1. `test_gapi_streaming.py` - fixed behavior of Desync test to avoid sporadic issues 2. `test_gapi_infer_onnx.py` - fixed model lookup (it was still using the ONNX Zoo layout but was NOT using the proper env var we use to point to one). ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2023-05-30 22:52:17 +08:00
static cv::GArg run_py_setup(cv::detail::PyObjectHolder setup,
const cv::GMetaArgs &meta,
const cv::GArgs &gargs)
2022-05-26 00:12:51 +08:00
{
PyGILState_STATE gstate;
gstate = PyGILState_Ensure();
Merge pull request #23597 from dmatveev:dm/gapi_onnx_py_integration G-API: Integration branch for ONNX & Python-related changes #23597 # Changes overview ## 1. Expose ONNX backend's Normalization and Mean-value parameters in Python * Since Python G-API bindings rely on `Generic` infer to express Inference, the `Generic` specialization of `onnx::Params` was extended with new methods to control normalization (`/255`) and mean-value; these methods were exposed in the Python bindings * Found some questionable parts in the existing API which I'd like to review/discuss (see comments) UPD: 1. Thanks to @TolyaTalamanov normalization inconsistencies have been identified with `squeezenet1.0-9` ONNX model itself; tests using these model were updated to DISABLE normalization and NOT using mean/value. 2. Questionable parts were removed and tests still pass. ### Details (taken from @TolyaTalamanov's comment): `squeezenet1.0.*onnx` - doesn't require scaling to [0,1] and mean/std because the weights of the first convolution already scaled. ONNX documentation is broken. So the correct approach to use this models is: 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 but without normalization step: ``` # DON'T DO IT: # mean_vec = np.array([0.485, 0.456, 0.406]) # stddev_vec = np.array([0.229, 0.224, 0.225]) # norm_img_data = np.zeros(img_data.shape).astype('float32') # for i in range(img_data.shape[0]): # norm_img_data[i,:,:] = (img_data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i] # # add batch channel # norm_img_data = norm_img_data.reshape(1, 3, 224, 224).astype('float32') # return norm_img_data # INSTEAD return img_data.reshape(1, 3, 224, 224) ``` 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', False) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. --- `squeezenet1.1.*onnx` - requires scaling to [0,1] and mean/std - onnx documentation is correct. 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', True) // default net.cfgMeanStd('data_0', [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. ## 2. Expose Fluid & kernel package-related functionality in Python * `cv::gapi::combine()` * `cv::GKernelPackage::size()` (mainly for testing purposes) * `cv::gapi::imgproc::fluid::kernels()` Added a test for the above. ## 3. Fixed issues with Python stateful kernel handling Fixed error message when `outMeta()` of custom python operation fails. ## 4. Fixed various issues in Python tests 1. `test_gapi_streaming.py` - fixed behavior of Desync test to avoid sporadic issues 2. `test_gapi_infer_onnx.py` - fixed model lookup (it was still using the ONNX Zoo layout but was NOT using the proper env var we use to point to one). ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2023-05-30 22:52:17 +08:00
cv::GArg state;
2022-05-26 00:12:51 +08:00
try
{
// NB: Doesn't increase reference counter (false),
// because PyObject already have ownership.
// In case exception decrement reference counter.
cv::detail::PyObjectHolder args(PyTuple_New(meta.size()), false);
unpackMetasToTuple(meta, gargs, args);
2022-05-26 00:12:51 +08:00
Merge pull request #23597 from dmatveev:dm/gapi_onnx_py_integration G-API: Integration branch for ONNX & Python-related changes #23597 # Changes overview ## 1. Expose ONNX backend's Normalization and Mean-value parameters in Python * Since Python G-API bindings rely on `Generic` infer to express Inference, the `Generic` specialization of `onnx::Params` was extended with new methods to control normalization (`/255`) and mean-value; these methods were exposed in the Python bindings * Found some questionable parts in the existing API which I'd like to review/discuss (see comments) UPD: 1. Thanks to @TolyaTalamanov normalization inconsistencies have been identified with `squeezenet1.0-9` ONNX model itself; tests using these model were updated to DISABLE normalization and NOT using mean/value. 2. Questionable parts were removed and tests still pass. ### Details (taken from @TolyaTalamanov's comment): `squeezenet1.0.*onnx` - doesn't require scaling to [0,1] and mean/std because the weights of the first convolution already scaled. ONNX documentation is broken. So the correct approach to use this models is: 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 but without normalization step: ``` # DON'T DO IT: # mean_vec = np.array([0.485, 0.456, 0.406]) # stddev_vec = np.array([0.229, 0.224, 0.225]) # norm_img_data = np.zeros(img_data.shape).astype('float32') # for i in range(img_data.shape[0]): # norm_img_data[i,:,:] = (img_data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i] # # add batch channel # norm_img_data = norm_img_data.reshape(1, 3, 224, 224).astype('float32') # return norm_img_data # INSTEAD return img_data.reshape(1, 3, 224, 224) ``` 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', False) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. --- `squeezenet1.1.*onnx` - requires scaling to [0,1] and mean/std - onnx documentation is correct. 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', True) // default net.cfgMeanStd('data_0', [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. ## 2. Expose Fluid & kernel package-related functionality in Python * `cv::gapi::combine()` * `cv::GKernelPackage::size()` (mainly for testing purposes) * `cv::gapi::imgproc::fluid::kernels()` Added a test for the above. ## 3. Fixed issues with Python stateful kernel handling Fixed error message when `outMeta()` of custom python operation fails. ## 4. Fixed various issues in Python tests 1. `test_gapi_streaming.py` - fixed behavior of Desync test to avoid sporadic issues 2. `test_gapi_infer_onnx.py` - fixed model lookup (it was still using the ONNX Zoo layout but was NOT using the proper env var we use to point to one). ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2023-05-30 22:52:17 +08:00
PyObject *py_kernel_state = PyObject_CallObject(setup.get(), args.get());
2022-05-26 00:12:51 +08:00
if (PyErr_Occurred())
{
PyErr_PrintEx(0);
PyErr_Clear();
Merge pull request #23597 from dmatveev:dm/gapi_onnx_py_integration G-API: Integration branch for ONNX & Python-related changes #23597 # Changes overview ## 1. Expose ONNX backend's Normalization and Mean-value parameters in Python * Since Python G-API bindings rely on `Generic` infer to express Inference, the `Generic` specialization of `onnx::Params` was extended with new methods to control normalization (`/255`) and mean-value; these methods were exposed in the Python bindings * Found some questionable parts in the existing API which I'd like to review/discuss (see comments) UPD: 1. Thanks to @TolyaTalamanov normalization inconsistencies have been identified with `squeezenet1.0-9` ONNX model itself; tests using these model were updated to DISABLE normalization and NOT using mean/value. 2. Questionable parts were removed and tests still pass. ### Details (taken from @TolyaTalamanov's comment): `squeezenet1.0.*onnx` - doesn't require scaling to [0,1] and mean/std because the weights of the first convolution already scaled. ONNX documentation is broken. So the correct approach to use this models is: 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 but without normalization step: ``` # DON'T DO IT: # mean_vec = np.array([0.485, 0.456, 0.406]) # stddev_vec = np.array([0.229, 0.224, 0.225]) # norm_img_data = np.zeros(img_data.shape).astype('float32') # for i in range(img_data.shape[0]): # norm_img_data[i,:,:] = (img_data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i] # # add batch channel # norm_img_data = norm_img_data.reshape(1, 3, 224, 224).astype('float32') # return norm_img_data # INSTEAD return img_data.reshape(1, 3, 224, 224) ``` 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', False) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. --- `squeezenet1.1.*onnx` - requires scaling to [0,1] and mean/std - onnx documentation is correct. 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', True) // default net.cfgMeanStd('data_0', [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. ## 2. Expose Fluid & kernel package-related functionality in Python * `cv::gapi::combine()` * `cv::GKernelPackage::size()` (mainly for testing purposes) * `cv::gapi::imgproc::fluid::kernels()` Added a test for the above. ## 3. Fixed issues with Python stateful kernel handling Fixed error message when `outMeta()` of custom python operation fails. ## 4. Fixed various issues in Python tests 1. `test_gapi_streaming.py` - fixed behavior of Desync test to avoid sporadic issues 2. `test_gapi_infer_onnx.py` - fixed model lookup (it was still using the ONNX Zoo layout but was NOT using the proper env var we use to point to one). ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2023-05-30 22:52:17 +08:00
throw std::logic_error("Python kernel setup failed with error!");
2022-05-26 00:12:51 +08:00
}
// NB: In fact it's impossible situation, because errors were handled above.
Merge pull request #23597 from dmatveev:dm/gapi_onnx_py_integration G-API: Integration branch for ONNX & Python-related changes #23597 # Changes overview ## 1. Expose ONNX backend's Normalization and Mean-value parameters in Python * Since Python G-API bindings rely on `Generic` infer to express Inference, the `Generic` specialization of `onnx::Params` was extended with new methods to control normalization (`/255`) and mean-value; these methods were exposed in the Python bindings * Found some questionable parts in the existing API which I'd like to review/discuss (see comments) UPD: 1. Thanks to @TolyaTalamanov normalization inconsistencies have been identified with `squeezenet1.0-9` ONNX model itself; tests using these model were updated to DISABLE normalization and NOT using mean/value. 2. Questionable parts were removed and tests still pass. ### Details (taken from @TolyaTalamanov's comment): `squeezenet1.0.*onnx` - doesn't require scaling to [0,1] and mean/std because the weights of the first convolution already scaled. ONNX documentation is broken. So the correct approach to use this models is: 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 but without normalization step: ``` # DON'T DO IT: # mean_vec = np.array([0.485, 0.456, 0.406]) # stddev_vec = np.array([0.229, 0.224, 0.225]) # norm_img_data = np.zeros(img_data.shape).astype('float32') # for i in range(img_data.shape[0]): # norm_img_data[i,:,:] = (img_data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i] # # add batch channel # norm_img_data = norm_img_data.reshape(1, 3, 224, 224).astype('float32') # return norm_img_data # INSTEAD return img_data.reshape(1, 3, 224, 224) ``` 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', False) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. --- `squeezenet1.1.*onnx` - requires scaling to [0,1] and mean/std - onnx documentation is correct. 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', True) // default net.cfgMeanStd('data_0', [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. ## 2. Expose Fluid & kernel package-related functionality in Python * `cv::gapi::combine()` * `cv::GKernelPackage::size()` (mainly for testing purposes) * `cv::gapi::imgproc::fluid::kernels()` Added a test for the above. ## 3. Fixed issues with Python stateful kernel handling Fixed error message when `outMeta()` of custom python operation fails. ## 4. Fixed various issues in Python tests 1. `test_gapi_streaming.py` - fixed behavior of Desync test to avoid sporadic issues 2. `test_gapi_infer_onnx.py` - fixed model lookup (it was still using the ONNX Zoo layout but was NOT using the proper env var we use to point to one). ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2023-05-30 22:52:17 +08:00
GAPI_Assert(py_kernel_state && "Python kernel setup returned NULL!");
2022-05-26 00:12:51 +08:00
Merge pull request #23597 from dmatveev:dm/gapi_onnx_py_integration G-API: Integration branch for ONNX & Python-related changes #23597 # Changes overview ## 1. Expose ONNX backend's Normalization and Mean-value parameters in Python * Since Python G-API bindings rely on `Generic` infer to express Inference, the `Generic` specialization of `onnx::Params` was extended with new methods to control normalization (`/255`) and mean-value; these methods were exposed in the Python bindings * Found some questionable parts in the existing API which I'd like to review/discuss (see comments) UPD: 1. Thanks to @TolyaTalamanov normalization inconsistencies have been identified with `squeezenet1.0-9` ONNX model itself; tests using these model were updated to DISABLE normalization and NOT using mean/value. 2. Questionable parts were removed and tests still pass. ### Details (taken from @TolyaTalamanov's comment): `squeezenet1.0.*onnx` - doesn't require scaling to [0,1] and mean/std because the weights of the first convolution already scaled. ONNX documentation is broken. So the correct approach to use this models is: 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 but without normalization step: ``` # DON'T DO IT: # mean_vec = np.array([0.485, 0.456, 0.406]) # stddev_vec = np.array([0.229, 0.224, 0.225]) # norm_img_data = np.zeros(img_data.shape).astype('float32') # for i in range(img_data.shape[0]): # norm_img_data[i,:,:] = (img_data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i] # # add batch channel # norm_img_data = norm_img_data.reshape(1, 3, 224, 224).astype('float32') # return norm_img_data # INSTEAD return img_data.reshape(1, 3, 224, 224) ``` 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', False) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. --- `squeezenet1.1.*onnx` - requires scaling to [0,1] and mean/std - onnx documentation is correct. 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', True) // default net.cfgMeanStd('data_0', [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. ## 2. Expose Fluid & kernel package-related functionality in Python * `cv::gapi::combine()` * `cv::GKernelPackage::size()` (mainly for testing purposes) * `cv::gapi::imgproc::fluid::kernels()` Added a test for the above. ## 3. Fixed issues with Python stateful kernel handling Fixed error message when `outMeta()` of custom python operation fails. ## 4. Fixed various issues in Python tests 1. `test_gapi_streaming.py` - fixed behavior of Desync test to avoid sporadic issues 2. `test_gapi_infer_onnx.py` - fixed model lookup (it was still using the ONNX Zoo layout but was NOT using the proper env var we use to point to one). ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2023-05-30 22:52:17 +08:00
if (!pyopencv_to(py_kernel_state, state, ArgInfo("arg", false)))
2022-05-26 00:12:51 +08:00
{
Merge pull request #23597 from dmatveev:dm/gapi_onnx_py_integration G-API: Integration branch for ONNX & Python-related changes #23597 # Changes overview ## 1. Expose ONNX backend's Normalization and Mean-value parameters in Python * Since Python G-API bindings rely on `Generic` infer to express Inference, the `Generic` specialization of `onnx::Params` was extended with new methods to control normalization (`/255`) and mean-value; these methods were exposed in the Python bindings * Found some questionable parts in the existing API which I'd like to review/discuss (see comments) UPD: 1. Thanks to @TolyaTalamanov normalization inconsistencies have been identified with `squeezenet1.0-9` ONNX model itself; tests using these model were updated to DISABLE normalization and NOT using mean/value. 2. Questionable parts were removed and tests still pass. ### Details (taken from @TolyaTalamanov's comment): `squeezenet1.0.*onnx` - doesn't require scaling to [0,1] and mean/std because the weights of the first convolution already scaled. ONNX documentation is broken. So the correct approach to use this models is: 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 but without normalization step: ``` # DON'T DO IT: # mean_vec = np.array([0.485, 0.456, 0.406]) # stddev_vec = np.array([0.229, 0.224, 0.225]) # norm_img_data = np.zeros(img_data.shape).astype('float32') # for i in range(img_data.shape[0]): # norm_img_data[i,:,:] = (img_data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i] # # add batch channel # norm_img_data = norm_img_data.reshape(1, 3, 224, 224).astype('float32') # return norm_img_data # INSTEAD return img_data.reshape(1, 3, 224, 224) ``` 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', False) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. --- `squeezenet1.1.*onnx` - requires scaling to [0,1] and mean/std - onnx documentation is correct. 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', True) // default net.cfgMeanStd('data_0', [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. ## 2. Expose Fluid & kernel package-related functionality in Python * `cv::gapi::combine()` * `cv::GKernelPackage::size()` (mainly for testing purposes) * `cv::gapi::imgproc::fluid::kernels()` Added a test for the above. ## 3. Fixed issues with Python stateful kernel handling Fixed error message when `outMeta()` of custom python operation fails. ## 4. Fixed various issues in Python tests 1. `test_gapi_streaming.py` - fixed behavior of Desync test to avoid sporadic issues 2. `test_gapi_infer_onnx.py` - fixed model lookup (it was still using the ONNX Zoo layout but was NOT using the proper env var we use to point to one). ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2023-05-30 22:52:17 +08:00
util::throw_error(std::logic_error("Failed to convert python state"));
2022-05-26 00:12:51 +08:00
}
}
catch (...)
{
PyGILState_Release(gstate);
throw;
}
PyGILState_Release(gstate);
Merge pull request #23597 from dmatveev:dm/gapi_onnx_py_integration G-API: Integration branch for ONNX & Python-related changes #23597 # Changes overview ## 1. Expose ONNX backend's Normalization and Mean-value parameters in Python * Since Python G-API bindings rely on `Generic` infer to express Inference, the `Generic` specialization of `onnx::Params` was extended with new methods to control normalization (`/255`) and mean-value; these methods were exposed in the Python bindings * Found some questionable parts in the existing API which I'd like to review/discuss (see comments) UPD: 1. Thanks to @TolyaTalamanov normalization inconsistencies have been identified with `squeezenet1.0-9` ONNX model itself; tests using these model were updated to DISABLE normalization and NOT using mean/value. 2. Questionable parts were removed and tests still pass. ### Details (taken from @TolyaTalamanov's comment): `squeezenet1.0.*onnx` - doesn't require scaling to [0,1] and mean/std because the weights of the first convolution already scaled. ONNX documentation is broken. So the correct approach to use this models is: 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 but without normalization step: ``` # DON'T DO IT: # mean_vec = np.array([0.485, 0.456, 0.406]) # stddev_vec = np.array([0.229, 0.224, 0.225]) # norm_img_data = np.zeros(img_data.shape).astype('float32') # for i in range(img_data.shape[0]): # norm_img_data[i,:,:] = (img_data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i] # # add batch channel # norm_img_data = norm_img_data.reshape(1, 3, 224, 224).astype('float32') # return norm_img_data # INSTEAD return img_data.reshape(1, 3, 224, 224) ``` 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', False) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. --- `squeezenet1.1.*onnx` - requires scaling to [0,1] and mean/std - onnx documentation is correct. 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', True) // default net.cfgMeanStd('data_0', [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. ## 2. Expose Fluid & kernel package-related functionality in Python * `cv::gapi::combine()` * `cv::GKernelPackage::size()` (mainly for testing purposes) * `cv::gapi::imgproc::fluid::kernels()` Added a test for the above. ## 3. Fixed issues with Python stateful kernel handling Fixed error message when `outMeta()` of custom python operation fails. ## 4. Fixed various issues in Python tests 1. `test_gapi_streaming.py` - fixed behavior of Desync test to avoid sporadic issues 2. `test_gapi_infer_onnx.py` - fixed model lookup (it was still using the ONNX Zoo layout but was NOT using the proper env var we use to point to one). ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2023-05-30 22:52:17 +08:00
return state;
2022-05-26 00:12:51 +08:00
}
static GMetaArg get_meta_arg(PyObject* obj)
{
cv::GMetaArg arg;
if (!pyopencv_to(obj, arg, ArgInfo("arg", false)))
{
util::throw_error(std::logic_error("Unsupported output meta type"));
}
return arg;
}
static cv::GMetaArgs get_meta_args(PyObject* tuple)
{
size_t size = PyTuple_Size(tuple);
cv::GMetaArgs metas;
metas.reserve(size);
for (size_t i = 0; i < size; ++i)
{
metas.push_back(get_meta_arg(PyTuple_GetItem(tuple, i)));
}
return metas;
}
static GMetaArgs run_py_meta(cv::detail::PyObjectHolder out_meta,
const cv::GMetaArgs &meta,
const cv::GArgs &gargs)
{
PyGILState_STATE gstate;
gstate = PyGILState_Ensure();
cv::GMetaArgs out_metas;
try
{
// NB: Doesn't increase reference counter (false),
// because PyObject already have ownership.
// In case exception decrement reference counter.
cv::detail::PyObjectHolder args(PyTuple_New(meta.size()), false);
unpackMetasToTuple(meta, gargs, args);
// NB: Doesn't increase reference counter (false).
// In case PyObject_CallObject return NULL, do nothing in destructor.
cv::detail::PyObjectHolder result(
PyObject_CallObject(out_meta.get(), args.get()), false);
if (PyErr_Occurred())
{
PyErr_PrintEx(0);
PyErr_Clear();
throw std::logic_error("Python outMeta failed with error!");
}
// NB: In fact it's impossible situation, because errors were handled above.
GAPI_Assert(result.get() && "Python outMeta returned NULL!");
out_metas = PyTuple_Check(result.get()) ? get_meta_args(result.get())
: cv::GMetaArgs{get_meta_arg(result.get())};
}
catch (...)
{
PyGILState_Release(gstate);
throw;
}
PyGILState_Release(gstate);
return out_metas;
}
2021-03-26 19:16:26 +08:00
static PyObject* pyopencv_cv_gapi_kernels(PyObject* , PyObject* py_args, PyObject*)
{
using namespace cv;
2021-12-24 23:04:11 +08:00
GKernelPackage pkg;
2021-03-26 19:16:26 +08:00
Py_ssize_t size = PyTuple_Size(py_args);
2021-03-26 19:16:26 +08:00
for (int i = 0; i < size; ++i)
{
PyObject* user_kernel = PyTuple_GetItem(py_args, i);
PyObject* id_obj = PyObject_GetAttrString(user_kernel, "id");
if (!id_obj)
{
PyErr_SetString(PyExc_TypeError,
"Python kernel should contain id, please use cv.gapi.kernel to define kernel");
return NULL;
}
PyObject* out_meta = PyObject_GetAttrString(user_kernel, "outMeta");
if (!out_meta)
{
PyErr_SetString(PyExc_TypeError,
"Python kernel should contain outMeta, please use cv.gapi.kernel to define kernel");
return NULL;
}
PyObject* run = PyObject_GetAttrString(user_kernel, "run");
if (!run)
{
PyErr_SetString(PyExc_TypeError,
"Python kernel should contain run, please use cv.gapi.kernel to define kernel");
return NULL;
}
2022-05-26 00:12:51 +08:00
PyObject* setup = nullptr;
if (PyObject_HasAttrString(user_kernel, "setup")) {
setup = PyObject_GetAttrString(user_kernel, "setup");
}
2021-03-26 19:16:26 +08:00
std::string id;
if (!pyopencv_to(id_obj, id, ArgInfo("id", false)))
2021-03-26 19:16:26 +08:00
{
PyErr_SetString(PyExc_TypeError, "Failed to obtain string");
2021-03-26 19:16:26 +08:00
return NULL;
}
using namespace std::placeholders;
2022-05-26 00:12:51 +08:00
if (setup)
{
gapi::python::GPythonFunctor f(
id.c_str(), std::bind(run_py_meta, cv::detail::PyObjectHolder{out_meta}, _1, _2),
std::bind(run_py_kernel, cv::detail::PyObjectHolder{run}, _1),
Merge pull request #23597 from dmatveev:dm/gapi_onnx_py_integration G-API: Integration branch for ONNX & Python-related changes #23597 # Changes overview ## 1. Expose ONNX backend's Normalization and Mean-value parameters in Python * Since Python G-API bindings rely on `Generic` infer to express Inference, the `Generic` specialization of `onnx::Params` was extended with new methods to control normalization (`/255`) and mean-value; these methods were exposed in the Python bindings * Found some questionable parts in the existing API which I'd like to review/discuss (see comments) UPD: 1. Thanks to @TolyaTalamanov normalization inconsistencies have been identified with `squeezenet1.0-9` ONNX model itself; tests using these model were updated to DISABLE normalization and NOT using mean/value. 2. Questionable parts were removed and tests still pass. ### Details (taken from @TolyaTalamanov's comment): `squeezenet1.0.*onnx` - doesn't require scaling to [0,1] and mean/std because the weights of the first convolution already scaled. ONNX documentation is broken. So the correct approach to use this models is: 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 but without normalization step: ``` # DON'T DO IT: # mean_vec = np.array([0.485, 0.456, 0.406]) # stddev_vec = np.array([0.229, 0.224, 0.225]) # norm_img_data = np.zeros(img_data.shape).astype('float32') # for i in range(img_data.shape[0]): # norm_img_data[i,:,:] = (img_data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i] # # add batch channel # norm_img_data = norm_img_data.reshape(1, 3, 224, 224).astype('float32') # return norm_img_data # INSTEAD return img_data.reshape(1, 3, 224, 224) ``` 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', False) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. --- `squeezenet1.1.*onnx` - requires scaling to [0,1] and mean/std - onnx documentation is correct. 1. ONNX: apply preprocessing from the documentation: https://github.com/onnx/models/blob/main/vision/classification/imagenet_preprocess.py#L8-L44 2. G-API: Convert image from BGR to RGB and then pass to `apply` as-is with configuring parameters: ``` net = cv.gapi.onnx.params('squeezenet', model_filename) net.cfgNormalize('data_0', True) // default net.cfgMeanStd('data_0', [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ``` **Note**: Results might be difference because `G-API` doesn't apply central crop but just do resize to model resolution. ## 2. Expose Fluid & kernel package-related functionality in Python * `cv::gapi::combine()` * `cv::GKernelPackage::size()` (mainly for testing purposes) * `cv::gapi::imgproc::fluid::kernels()` Added a test for the above. ## 3. Fixed issues with Python stateful kernel handling Fixed error message when `outMeta()` of custom python operation fails. ## 4. Fixed various issues in Python tests 1. `test_gapi_streaming.py` - fixed behavior of Desync test to avoid sporadic issues 2. `test_gapi_infer_onnx.py` - fixed model lookup (it was still using the ONNX Zoo layout but was NOT using the proper env var we use to point to one). ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake
2023-05-30 22:52:17 +08:00
std::bind(run_py_setup, cv::detail::PyObjectHolder{setup}, _1, _2));
2022-05-26 00:12:51 +08:00
pkg.include(f);
}
else
{
gapi::python::GPythonFunctor f(
id.c_str(), std::bind(run_py_meta, cv::detail::PyObjectHolder{out_meta}, _1, _2),
std::bind(run_py_kernel, cv::detail::PyObjectHolder{run}, _1));
pkg.include(f);
}
2021-03-26 19:16:26 +08:00
}
return pyopencv_from(pkg);
}
static PyObject* pyopencv_cv_gapi_op(PyObject* , PyObject* py_args, PyObject*)
{
using namespace cv;
Py_ssize_t size = PyTuple_Size(py_args);
std::string id;
if (!pyopencv_to(PyTuple_GetItem(py_args, 0), id, ArgInfo("id", false)))
{
PyErr_SetString(PyExc_TypeError, "Failed to obtain: operation id must be a string");
return NULL;
}
PyObject* outMeta = PyTuple_GetItem(py_args, 1);
cv::GArgs args;
for (int i = 2; i < size; i++)
{
PyObject* item = PyTuple_GetItem(py_args, i);
if (PyObject_TypeCheck(item,
reinterpret_cast<PyTypeObject*>(pyopencv_GMat_TypePtr)))
{
args.emplace_back(reinterpret_cast<pyopencv_GMat_t*>(item)->v);
}
else if (PyObject_TypeCheck(item,
reinterpret_cast<PyTypeObject*>(pyopencv_GScalar_TypePtr)))
{
args.emplace_back(reinterpret_cast<pyopencv_GScalar_t*>(item)->v);
}
else if (PyObject_TypeCheck(item,
reinterpret_cast<PyTypeObject*>(pyopencv_GOpaqueT_TypePtr)))
{
auto&& arg = reinterpret_cast<pyopencv_GOpaqueT_t*>(item)->v.arg();
#define HC(T, K) case cv::GOpaqueT::Storage:: index_of<cv::GOpaque<T>>(): \
args.emplace_back(cv::util::get<cv::GOpaque<T>>(arg)); \
break; \
SWITCH(arg.index(), GOPAQUE_TYPE_LIST_G, HC)
#undef HC
}
else if (PyObject_TypeCheck(item,
reinterpret_cast<PyTypeObject*>(pyopencv_GArrayT_TypePtr)))
{
auto&& arg = reinterpret_cast<pyopencv_GArrayT_t*>(item)->v.arg();
#define HC(T, K) case cv::GArrayT::Storage:: index_of<cv::GArray<T>>(): \
args.emplace_back(cv::util::get<cv::GArray<T>>(arg)); \
break; \
SWITCH(arg.index(), GARRAY_TYPE_LIST_G, HC)
#undef HC
}
else
{
args.emplace_back(cv::GArg(cv::detail::PyObjectHolder{item}));
}
}
cv::GKernel::M outMetaWrapper = std::bind(run_py_meta,
cv::detail::PyObjectHolder{outMeta},
std::placeholders::_1,
std::placeholders::_2);
return pyopencv_from(cv::gapi::wip::op(id, outMetaWrapper, std::move(args)));
}
template<>
bool pyopencv_to(PyObject* obj, cv::detail::ExtractArgsCallback& value, const ArgInfo&)
{
cv::detail::PyObjectHolder holder{obj};
value = cv::detail::ExtractArgsCallback{[=](const cv::GTypesInfo& info)
{
PyGILState_STATE gstate;
gstate = PyGILState_Ensure();
cv::GRunArgs args;
try
{
args = extract_run_args(info, holder.get());
}
catch (...)
{
PyGILState_Release(gstate);
throw;
}
PyGILState_Release(gstate);
return args;
}};
return true;
}
template<>
bool pyopencv_to(PyObject* obj, cv::detail::ExtractMetaCallback& value, const ArgInfo&)
{
cv::detail::PyObjectHolder holder{obj};
value = cv::detail::ExtractMetaCallback{[=](const cv::GTypesInfo& info)
{
PyGILState_STATE gstate;
gstate = PyGILState_Ensure();
cv::GMetaArgs args;
try
{
args = extract_meta_args(info, holder.get());
}
catch (...)
{
PyGILState_Release(gstate);
throw;
}
PyGILState_Release(gstate);
return args;
}};
return true;
}
template<typename T>
struct PyOpenCV_Converter<cv::GArray<T>>
{
static PyObject* from(const cv::GArray<T>& p)
{
return pyopencv_from(cv::GArrayT(p));
}
static bool to(PyObject *obj, cv::GArray<T>& value, const ArgInfo& info)
{
if (PyObject_TypeCheck(obj, reinterpret_cast<PyTypeObject*>(pyopencv_GArrayT_TypePtr)))
{
auto& array = reinterpret_cast<pyopencv_GArrayT_t*>(obj)->v;
try
{
value = cv::util::get<cv::GArray<T>>(array.arg());
}
catch (...)
{
return false;
}
return true;
}
return false;
}
};
template<typename T>
struct PyOpenCV_Converter<cv::GOpaque<T>>
{
static PyObject* from(const cv::GOpaque<T>& p)
{
return pyopencv_from(cv::GOpaqueT(p));
}
static bool to(PyObject *obj, cv::GOpaque<T>& value, const ArgInfo& info)
{
if (PyObject_TypeCheck(obj, reinterpret_cast<PyTypeObject*>(pyopencv_GOpaqueT_TypePtr)))
{
auto& opaque = reinterpret_cast<pyopencv_GOpaqueT_t*>(obj)->v;
try
{
value = cv::util::get<cv::GOpaque<T>>(opaque.arg());
}
catch (...)
{
return false;
}
return true;
}
return false;
}
};
template<>
bool pyopencv_to(PyObject* obj, cv::GProtoInputArgs& value, const ArgInfo& info)
{
try
{
value = extract_proto_args<cv::GProtoInputArgs>(obj);
return true;
}
catch (...)
{
failmsg("Can't parse cv::GProtoInputArgs");
return false;
}
}
template<>
bool pyopencv_to(PyObject* obj, cv::GProtoOutputArgs& value, const ArgInfo& info)
{
try
{
value = extract_proto_args<cv::GProtoOutputArgs>(obj);
return true;
}
catch (...)
{
failmsg("Can't parse cv::GProtoOutputArgs");
return false;
}
}
// extend cv.gapi methods
#define PYOPENCV_EXTRA_METHODS_GAPI \
{"kernels", CV_PY_FN_WITH_KW(pyopencv_cv_gapi_kernels), "kernels(...) -> GKernelPackage"}, \
{"__op", CV_PY_FN_WITH_KW(pyopencv_cv_gapi_op), "__op(...) -> retval\n"},
#endif // HAVE_OPENCV_GAPI
#endif // OPENCV_GAPI_PYOPENCV_GAPI_HPP