From 69d1644ed158b695ea3508beaf2b6ba46d161c80 Mon Sep 17 00:00:00 2001
From: Olivier BICHLER <olivier.bichler@cea.fr>
Date: Mon, 31 Mar 2025 18:32:42 +0200
Subject: [PATCH 1/4] Wrong type for axes
---
include/aidge/operator/Squeeze.hpp | 10 +++++-----
python_binding/operator/pybind_Squeeze.cpp | 2 +-
src/operator/Squeeze.cpp | 14 +++++++-------
3 files changed, 13 insertions(+), 13 deletions(-)
diff --git a/include/aidge/operator/Squeeze.hpp b/include/aidge/operator/Squeeze.hpp
index 03db92a84..b39ecf73c 100644
--- a/include/aidge/operator/Squeeze.hpp
+++ b/include/aidge/operator/Squeeze.hpp
@@ -27,7 +27,7 @@
#define LIST_SQUEEZE_ATTR(X) \
- X(Axes, "axes", std::vector<std::int8_t>)
+ X(Axes, "axes", std::vector<std::int64_t>)
namespace Aidge {
/**
@@ -78,7 +78,7 @@ public:
Type; // name of the type of the operation (Here "Squeeze")
private:
- using Attributes_ = StaticAttributes<SqueezeAttr, std::vector<std::int8_t>>;
+ using Attributes_ = StaticAttributes<SqueezeAttr, std::vector<std::int64_t>>;
template <SqueezeAttr e> using attr = typename Attributes_::template attr<e>;
const std::shared_ptr<Attributes_> mAttributes;
@@ -87,7 +87,7 @@ public:
* @brief constructor for Squeeze op
* @param[in] axes around which perform the operation
*/
- Squeeze_Op(const std::vector<std::int8_t> &axes = {});
+ Squeeze_Op(const std::vector<std::int64_t> &axes = {});
/**
* @brief Copy-constructor. Copy the operator attributes and its output
@@ -123,7 +123,7 @@ public:
* @brief axes to squeeze, if left empty all 1 sized
* dimensions will be removed.
*/
- inline std::vector<std::int8_t> &axes() const noexcept {
+ inline std::vector<std::int64_t> &axes() const noexcept {
return mAttributes->template getAttr<SqueezeAttr::Axes>();
}
@@ -145,7 +145,7 @@ public:
// helper with C-style array instead of std::array for kernel_dims to allow
// automatic template DIM deduction
-std::shared_ptr<Node> Squeeze(const std::vector<std::int8_t> axes = {},
+std::shared_ptr<Node> Squeeze(const std::vector<std::int64_t> axes = {},
const std::string &name = "");
} // namespace Aidge
diff --git a/python_binding/operator/pybind_Squeeze.cpp b/python_binding/operator/pybind_Squeeze.cpp
index 7808c78da..d5174c98d 100644
--- a/python_binding/operator/pybind_Squeeze.cpp
+++ b/python_binding/operator/pybind_Squeeze.cpp
@@ -46,7 +46,7 @@ void init_Squeeze(py::module &m) {
.def("axes", &Squeeze_Op::axes);
declare_registrable<Squeeze_Op>(m, "SqueezeOp");
- m.def("Squeeze", &Squeeze, py::arg("axes") = std::vector<int8_t>({}),
+ m.def("Squeeze", &Squeeze, py::arg("axes") = std::vector<int64_t>({}),
py::arg("name") = "",
R"mydelimiter(
Initialize a node containing a squeeze operator.
diff --git a/src/operator/Squeeze.cpp b/src/operator/Squeeze.cpp
index ea3452878..e1c99f88f 100644
--- a/src/operator/Squeeze.cpp
+++ b/src/operator/Squeeze.cpp
@@ -28,7 +28,7 @@
namespace Aidge {
const std::string Squeeze_Op::Type = "Squeeze";
-Squeeze_Op::Squeeze_Op(const std::vector<std::int8_t> &axes)
+Squeeze_Op::Squeeze_Op(const std::vector<std::int64_t> &axes)
: OperatorTensor(
Type,
{InputCategory::Data, InputCategory::OptionalData},
@@ -84,7 +84,7 @@ bool Squeeze_Op::forwardDims(bool allowDataDependency) {
this->axes().clear(); // If both are provided input would override attrs
this->axes().reserve(getInput(1)->size());
const auto &axes =
- getInput(1)->refCastFrom(fallback, NativeType_v<std::int8_t>, "cpu");
+ getInput(1)->refCastFrom(fallback, NativeType_v<std::int64_t>, "cpu");
if (axes.nbDims() == 0) {
this->axes().clear();
} else {
@@ -92,7 +92,7 @@ bool Squeeze_Op::forwardDims(bool allowDataDependency) {
axes.nbDims() == 1,
"Axes input tensor should be of size 1. Received {} dimensions : {}",
axes.nbDims(), axes.dims());
- std::copy_n(static_cast<std::int8_t *>(axes.getImpl()->hostPtr()), axes.size(),
+ std::copy_n(static_cast<std::int64_t *>(axes.getImpl()->hostPtr()), axes.size(),
std::back_inserter(this->axes()));
}
}
@@ -107,9 +107,9 @@ bool Squeeze_Op::forwardDims(bool allowDataDependency) {
}
}
} else {
- for (std::int8_t axis : this->axes()) {
- axis = axis >= 0 ? axis : axis + static_cast<std::int8_t>(inputDims.size());
- if ((axis < 0) || (axis >= static_cast<std::int8_t>(inputDims.size()))) {
+ for (std::int64_t axis : this->axes()) {
+ axis = axis >= 0 ? axis : axis + static_cast<std::int64_t>(inputDims.size());
+ if ((axis < 0) || (axis >= static_cast<std::int64_t>(inputDims.size()))) {
Log::error("{} : Axis index OutOfBounds error, expected value "
"within size limits of input tensor : "
"[-{},{}], got {}.",
@@ -152,7 +152,7 @@ std::set<std::string> Aidge::Squeeze_Op::getAvailableBackends() const {
////////////////////////////////////////////////////////////////////////////////
-std::shared_ptr<Node> Squeeze(const std::vector<std::int8_t> axes,
+std::shared_ptr<Node> Squeeze(const std::vector<std::int64_t> axes,
const std::string &name) {
return std::make_shared<Node>(std::make_shared<Squeeze_Op>(axes), name);
}
--
GitLab
From 56166eee204696887ce1dac9d1bc4d11de0112d6 Mon Sep 17 00:00:00 2001
From: Noam ZERAH <noam.zerah@cea.fr>
Date: Tue, 1 Apr 2025 09:38:50 +0000
Subject: [PATCH 2/4] Fixing an issue in the ForwardDims of avgPooling when
kernel size > input dim size
---
src/operator/AvgPooling.cpp | 20 +++++++++++++++-----
1 file changed, 15 insertions(+), 5 deletions(-)
diff --git a/src/operator/AvgPooling.cpp b/src/operator/AvgPooling.cpp
index 7561297ba..fe7083c34 100644
--- a/src/operator/AvgPooling.cpp
+++ b/src/operator/AvgPooling.cpp
@@ -74,11 +74,21 @@ bool Aidge::AvgPooling_Op<DIM>::forwardDims(bool /*allowDataDependency*/) {
const auto kernelDim = mAttributes->template getAttr<AvgPoolingAttr::KernelDims>()[dim];
const auto strideDim = mAttributes->template getAttr<AvgPoolingAttr::StrideDims>()[dim];
const auto dilationDim = mAttributes->template getAttr<AvgPoolingAttr::Dilations>()[dim];
-
- outputDims[dim+2] = 1 + static_cast<DimSize_t>(
- roundingFunction(static_cast<float>(inputDims[dim+2] -
- (kernelDim - 1) * dilationDim - 1) /
- static_cast<float>(strideDim)));
+ const auto inputDim = inputDims[dim + 2];
+ const auto effective_kernel_size = (kernelDim - 1) * dilationDim + 1;
+
+ if (effective_kernel_size >= inputDim) {
+ //If the kernel is greater than the input we set it to 1
+ outputDims[dim + 2] = 1;
+ }
+ else {
+ outputDims[dim + 2] = 1 + static_cast<DimSize_t>(
+ roundingFunction(
+ static_cast<float>(inputDim - effective_kernel_size) /
+ static_cast<float>(strideDim)
+ )
+ );
+ }
}
outputDims[1] = inputDims[1];
outputDims[0] = inputDims[0];
--
GitLab
From c815fccc65d31d7cfc1af5aefead0392596a20c5 Mon Sep 17 00:00:00 2001
From: Noam ZERAH <noam.zerah@cea.fr>
Date: Tue, 1 Apr 2025 09:46:27 +0000
Subject: [PATCH 3/4] Rebase on dev version for squeeze
---
include/aidge/operator/Squeeze.hpp | 10 +++++-----
python_binding/operator/pybind_Squeeze.cpp | 2 +-
src/operator/Squeeze.cpp | 14 +++++++-------
3 files changed, 13 insertions(+), 13 deletions(-)
diff --git a/include/aidge/operator/Squeeze.hpp b/include/aidge/operator/Squeeze.hpp
index b39ecf73c..03db92a84 100644
--- a/include/aidge/operator/Squeeze.hpp
+++ b/include/aidge/operator/Squeeze.hpp
@@ -27,7 +27,7 @@
#define LIST_SQUEEZE_ATTR(X) \
- X(Axes, "axes", std::vector<std::int64_t>)
+ X(Axes, "axes", std::vector<std::int8_t>)
namespace Aidge {
/**
@@ -78,7 +78,7 @@ public:
Type; // name of the type of the operation (Here "Squeeze")
private:
- using Attributes_ = StaticAttributes<SqueezeAttr, std::vector<std::int64_t>>;
+ using Attributes_ = StaticAttributes<SqueezeAttr, std::vector<std::int8_t>>;
template <SqueezeAttr e> using attr = typename Attributes_::template attr<e>;
const std::shared_ptr<Attributes_> mAttributes;
@@ -87,7 +87,7 @@ public:
* @brief constructor for Squeeze op
* @param[in] axes around which perform the operation
*/
- Squeeze_Op(const std::vector<std::int64_t> &axes = {});
+ Squeeze_Op(const std::vector<std::int8_t> &axes = {});
/**
* @brief Copy-constructor. Copy the operator attributes and its output
@@ -123,7 +123,7 @@ public:
* @brief axes to squeeze, if left empty all 1 sized
* dimensions will be removed.
*/
- inline std::vector<std::int64_t> &axes() const noexcept {
+ inline std::vector<std::int8_t> &axes() const noexcept {
return mAttributes->template getAttr<SqueezeAttr::Axes>();
}
@@ -145,7 +145,7 @@ public:
// helper with C-style array instead of std::array for kernel_dims to allow
// automatic template DIM deduction
-std::shared_ptr<Node> Squeeze(const std::vector<std::int64_t> axes = {},
+std::shared_ptr<Node> Squeeze(const std::vector<std::int8_t> axes = {},
const std::string &name = "");
} // namespace Aidge
diff --git a/python_binding/operator/pybind_Squeeze.cpp b/python_binding/operator/pybind_Squeeze.cpp
index d5174c98d..7808c78da 100644
--- a/python_binding/operator/pybind_Squeeze.cpp
+++ b/python_binding/operator/pybind_Squeeze.cpp
@@ -46,7 +46,7 @@ void init_Squeeze(py::module &m) {
.def("axes", &Squeeze_Op::axes);
declare_registrable<Squeeze_Op>(m, "SqueezeOp");
- m.def("Squeeze", &Squeeze, py::arg("axes") = std::vector<int64_t>({}),
+ m.def("Squeeze", &Squeeze, py::arg("axes") = std::vector<int8_t>({}),
py::arg("name") = "",
R"mydelimiter(
Initialize a node containing a squeeze operator.
diff --git a/src/operator/Squeeze.cpp b/src/operator/Squeeze.cpp
index e1c99f88f..ea3452878 100644
--- a/src/operator/Squeeze.cpp
+++ b/src/operator/Squeeze.cpp
@@ -28,7 +28,7 @@
namespace Aidge {
const std::string Squeeze_Op::Type = "Squeeze";
-Squeeze_Op::Squeeze_Op(const std::vector<std::int64_t> &axes)
+Squeeze_Op::Squeeze_Op(const std::vector<std::int8_t> &axes)
: OperatorTensor(
Type,
{InputCategory::Data, InputCategory::OptionalData},
@@ -84,7 +84,7 @@ bool Squeeze_Op::forwardDims(bool allowDataDependency) {
this->axes().clear(); // If both are provided input would override attrs
this->axes().reserve(getInput(1)->size());
const auto &axes =
- getInput(1)->refCastFrom(fallback, NativeType_v<std::int64_t>, "cpu");
+ getInput(1)->refCastFrom(fallback, NativeType_v<std::int8_t>, "cpu");
if (axes.nbDims() == 0) {
this->axes().clear();
} else {
@@ -92,7 +92,7 @@ bool Squeeze_Op::forwardDims(bool allowDataDependency) {
axes.nbDims() == 1,
"Axes input tensor should be of size 1. Received {} dimensions : {}",
axes.nbDims(), axes.dims());
- std::copy_n(static_cast<std::int64_t *>(axes.getImpl()->hostPtr()), axes.size(),
+ std::copy_n(static_cast<std::int8_t *>(axes.getImpl()->hostPtr()), axes.size(),
std::back_inserter(this->axes()));
}
}
@@ -107,9 +107,9 @@ bool Squeeze_Op::forwardDims(bool allowDataDependency) {
}
}
} else {
- for (std::int64_t axis : this->axes()) {
- axis = axis >= 0 ? axis : axis + static_cast<std::int64_t>(inputDims.size());
- if ((axis < 0) || (axis >= static_cast<std::int64_t>(inputDims.size()))) {
+ for (std::int8_t axis : this->axes()) {
+ axis = axis >= 0 ? axis : axis + static_cast<std::int8_t>(inputDims.size());
+ if ((axis < 0) || (axis >= static_cast<std::int8_t>(inputDims.size()))) {
Log::error("{} : Axis index OutOfBounds error, expected value "
"within size limits of input tensor : "
"[-{},{}], got {}.",
@@ -152,7 +152,7 @@ std::set<std::string> Aidge::Squeeze_Op::getAvailableBackends() const {
////////////////////////////////////////////////////////////////////////////////
-std::shared_ptr<Node> Squeeze(const std::vector<std::int64_t> axes,
+std::shared_ptr<Node> Squeeze(const std::vector<std::int8_t> axes,
const std::string &name) {
return std::make_shared<Node>(std::make_shared<Squeeze_Op>(axes), name);
}
--
GitLab
From a7eb20537e51252e3cab94df6e5518f25778a9ac Mon Sep 17 00:00:00 2001
From: Noam ZERAH <noam.zerah@cea.fr>
Date: Tue, 1 Apr 2025 11:56:00 +0000
Subject: [PATCH 4/4] Refactoring the condition in forward dim (using std::max)
---
src/operator/AvgPooling.cpp | 19 +++++--------------
1 file changed, 5 insertions(+), 14 deletions(-)
diff --git a/src/operator/AvgPooling.cpp b/src/operator/AvgPooling.cpp
index fe7083c34..966063cd0 100644
--- a/src/operator/AvgPooling.cpp
+++ b/src/operator/AvgPooling.cpp
@@ -74,22 +74,13 @@ bool Aidge::AvgPooling_Op<DIM>::forwardDims(bool /*allowDataDependency*/) {
const auto kernelDim = mAttributes->template getAttr<AvgPoolingAttr::KernelDims>()[dim];
const auto strideDim = mAttributes->template getAttr<AvgPoolingAttr::StrideDims>()[dim];
const auto dilationDim = mAttributes->template getAttr<AvgPoolingAttr::Dilations>()[dim];
- const auto inputDim = inputDims[dim + 2];
- const auto effective_kernel_size = (kernelDim - 1) * dilationDim + 1;
+
+ const float effective_size = static_cast<float>(inputDims[dim+2] - (kernelDim - 1) * dilationDim - 1);
+ const float rounded_val = roundingFunction(effective_size / static_cast<float>(strideDim));
- if (effective_kernel_size >= inputDim) {
- //If the kernel is greater than the input we set it to 1
- outputDims[dim + 2] = 1;
- }
- else {
- outputDims[dim + 2] = 1 + static_cast<DimSize_t>(
- roundingFunction(
- static_cast<float>(inputDim - effective_kernel_size) /
- static_cast<float>(strideDim)
- )
- );
- }
+ outputDims[dim+2] = 1 + std::max(0, static_cast<int>(rounded_val));
}
+
outputDims[1] = inputDims[1];
outputDims[0] = inputDims[0];
getOutput(0)->resize(outputDims);
--
GitLab