Forked from
Eclipse Projects / aidge / aidge_core
1868 commits behind the upstream repository.
-
Olivier BICHLER authored
Use fmt library instead of custom functions, added GraphView::getRankedNodes() and GraphView::getRankedNodesName() methods
Olivier BICHLER authoredUse fmt library instead of custom functions, added GraphView::getRankedNodes() and GraphView::getRankedNodesName() methods
Code owners
Assign users and groups as approvers for specific file changes. Learn more.
Conv.hpp 11.25 KiB
/********************************************************************************
* Copyright (c) 2023 CEA-List
*
* This program and the accompanying materials are made available under the
* terms of the Eclipse Public License 2.0 which is available at
* http://www.eclipse.org/legal/epl-2.0.
*
* SPDX-License-Identifier: EPL-2.0
*
********************************************************************************/
#ifndef AIDGE_CORE_OPERATOR_CONV_H_
#define AIDGE_CORE_OPERATOR_CONV_H_
#include <array>
#include <cmath>
#include <cstddef>
#include <numeric>
#include <vector>
#include "aidge/data/Tensor.hpp"
#include "aidge/graph/Node.hpp"
#include "aidge/operator/OperatorTensor.hpp"
#include "aidge/operator/Producer.hpp"
#include "aidge/utils/StaticAttributes.hpp"
#include "aidge/utils/Registrar.hpp"
#include "aidge/utils/Types.h"
namespace Aidge {
enum class ConvAttr { StrideDims, DilationDims, InChannels, OutChannels, KernelDims };
template <DimIdx_t DIM>
class Conv_Op : public OperatorTensor,
public Registrable<Conv_Op<DIM>, std::string, std::unique_ptr<OperatorImpl>(const Conv_Op<DIM> &)>,
public StaticAttributes<ConvAttr, std::array<DimSize_t, DIM>, std::array<DimSize_t, DIM>, DimSize_t,
DimSize_t, std::array<DimSize_t, DIM>> {
public:
static const std::string Type;
Conv_Op() = delete;
using Attributes_ = StaticAttributes<ConvAttr, std::array<DimSize_t, DIM>, std::array<DimSize_t, DIM>,
DimSize_t, DimSize_t, std::array<DimSize_t, DIM>>;
template <ConvAttr e>
using attr = typename Attributes_::template attr<e>;
constexpr Conv_Op(DimSize_t inChannels,
DimSize_t outChannels,
const std::array<DimSize_t, DIM> &kernelDims,
const std::array<DimSize_t, DIM> &strideDims = create_array<DimSize_t,DIM>(1),
const std::array<DimSize_t, DIM> &dilationDims = create_array<DimSize_t,DIM>(1))
: OperatorTensor(Type, 1, 2, 1),
Attributes_(attr<ConvAttr::StrideDims>(strideDims),
attr<ConvAttr::DilationDims>(dilationDims),
attr<ConvAttr::InChannels>(inChannels),
attr<ConvAttr::OutChannels>(outChannels),
attr<ConvAttr::KernelDims>(kernelDims)) {}
/**
* @brief Copy-constructor. Copy the operator attributes and its output tensor(s), but not its input tensors (the new operator has no input associated).
* @param op Operator to copy.
*/
Conv_Op(const Conv_Op<DIM>& op)
: OperatorTensor(op),
Attributes_(op)
{
mImpl = op.mImpl ? Registrar<Conv_Op<DIM>>::create(op.mOutputs[0]->getImpl()->backend())(*this) : nullptr;
}
/**
* @brief Clone the operator using its copy-constructor.
* @see Operator::Conv_Op
*/
std::shared_ptr<Operator> clone() const override {
return std::make_shared<Conv_Op<DIM>>(*this);
}
// Data operator[](const char* inputName) override final {
// std::shared_ptr<Tensor> in = (strcmp(inputName, "data")) ? getInput(0) :
// (strcmp(inputName, "weight") ? getInput(1) :
// (strcmp(inputName, "bias") ? getInput(2) :
// nullptr));
// assert((in!=nullptr) && "No such parameter");
// return *in;
// }
// std::shared_ptr<Conv_Op> clone() const override final {
// }
void computeOutputDims() override final {
// check inputs have been associated
bool associated = true;
for (IOIndex_t i = 0; i < 3; ++i) {
if (!getInput(i)) {
AIDGE_THROW_OR_ABORT(std::runtime_error, "Every input should be associated with a Tensor");
}
associated &= !(getInput(i)->empty());
}
if (associated) {
std::array<DimSize_t, DIM + 2> outputDims{};
const std::array<DimSize_t, DIM + 2> inputDims(getInput(0)->template dims<DIM+2>());
for (std::size_t dim = 0; dim < this->template getAttr<ConvAttr::KernelDims>().size() ; ++dim) {
const DimSize_t kernelExtent = this->template getAttr<ConvAttr::DilationDims>()[dim] *
(this->template getAttr<ConvAttr::KernelDims>()[dim] - 1) +
1;
outputDims[dim+2] = 1 + static_cast<DimSize_t>(
floor(static_cast<float>(inputDims[dim+2] - kernelExtent) /
static_cast<float>(this->template getAttr<ConvAttr::StrideDims>()[dim])));
}
outputDims[1] = this->template getAttr<ConvAttr::OutChannels>();
outputDims[0] = inputDims[0];
mOutputs[0]->resize(outputDims);
}
}
std::vector<std::pair<std::vector<Aidge::DimSize_t>, std::vector<DimSize_t>>> computeReceptiveField(const std::vector<DimSize_t>& firstEltDims, const std::vector<DimSize_t>& outputDims, const IOIndex_t outputIdx = 0) const override {
if (outputIdx != 0) {
AIDGE_THROW_OR_ABORT(std::runtime_error, "Conv_Op Operator has got only one output Tensor.");
}
if (firstEltDims.size() != outputDims.size()) {
AIDGE_THROW_OR_ABORT(std::runtime_error, "outputDims and firstEltDims should have the size of the output Tensor dimensions.");
}
if ((outputDims.size() == (DIM+2)) && outputDimsForwarded()) {
// Offset
auto inputIdxDims = firstEltDims; // batch idx is the same
inputIdxDims[1] = 0; // each channel is used so start with the first one
for (DimIdx_t i = 0; i < (DIM+2); ++i) {
if (((outputDims[i] + firstEltDims[i]) > mOutputs[0]->template dims<DIM+2>()[i]) || (outputDims[i] == 0)) {
AIDGE_THROW_OR_ABORT(std::runtime_error, "Given outputDim out of range for dimension {} ({} + {})", static_cast<std::size_t>(i), firstEltDims[i], outputDims[i]);
}
}
// padding is not a parameter of Conv_Op. It is handled in Pad_Op Operator // Input
// same batch value, every input channel is used
std::vector<DimSize_t> inputDims{outputDims[0], getInput(0)->dims()[1]};
for (DimIdx_t i = 0; i < DIM; ++i) {
inputDims.push_back((outputDims[2+static_cast<std::size_t>(i)] - 1)
* this->template getAttr<ConvAttr::StrideDims>()[static_cast<std::size_t>(i)]
+ 1
+ (this->template getAttr<ConvAttr::KernelDims>()[static_cast<std::size_t>(i)] - 1)
* this->template getAttr<ConvAttr::DilationDims>()[static_cast<std::size_t>(i)]);
inputIdxDims[2+i] *= this->template getAttr<ConvAttr::StrideDims>()[static_cast<std::size_t>(i)];
}
// Weight
// same output value, every input channel is used
std::vector<DimSize_t> weightDims{outputDims[1], getInput(0)->dims()[1]};
for (std::size_t i = 0; i < DIM; ++i) {
weightDims.push_back(this->template getAttr<ConvAttr::KernelDims>()[i]);
}
std::vector<DimSize_t> weightIdxDims = std::vector<DimSize_t>(DIM+2, 0);
weightIdxDims[0] = firstEltDims[1];
// Bias
const std::vector<DimSize_t> biasDims{outputDims[1]}; // the number of output channel
const std::vector<DimSize_t> biasIdxDims{firstEltDims[1]};
// Result
std::vector<std::pair<std::vector<Aidge::DimSize_t>, std::vector<DimSize_t>>> res;
res.push_back(std::pair<std::vector<Aidge::DimSize_t>, std::vector<DimSize_t>>(inputIdxDims, inputDims));
res.push_back(std::pair<std::vector<Aidge::DimSize_t>, std::vector<DimSize_t>>(weightIdxDims, weightDims));
res.push_back(std::pair<std::vector<Aidge::DimSize_t>, std::vector<DimSize_t>>(biasIdxDims, biasDims));
return res;
}
AIDGE_THROW_OR_ABORT(std::runtime_error, "Given outputDim out of range or output dim not forwarded yet.");
}
void setBackend(const std::string &name, DeviceIdx_t device = 0) override {
mImpl = Registrar<Conv_Op<DIM>>::create(name)(*this);
mOutputs[0]->setBackend(name, device);
// By default, automatically set backend for weight and bias inputs
getInput(1)->setBackend(name, device);
getInput(2)->setBackend(name, device);
}
static const std::vector<std::string> getInputsName(){
return {"data_input", "weight", "bias"};
}
static const std::vector<std::string> getOutputsName(){
return {"data_output"};
}
};
template <DimIdx_t DIM>
const std::string Conv_Op<DIM>::Type = "Conv";
/**
* @brief Perform a convolution on the input Tensor.
*
* @tparam DIM Number of dimensions for the feature map.
* @param inChannels Number of input channels.
* @param outChannels Number of output channels.
* @param kernelDims Dimensions of the kernel. Must be the same number of dimensions as the feature map.
* @param name Name of the operator.
* @param strideDims Dimensions of the stride attribute. Must be the same number of dimensions as the feature map.
* @param dilationDims Dimensions of the dilation attribute. Must be the same number of dimensions as the feature map.
* @return std::shared_ptr<Node> A Node containing the operator.
*/
template <std::array<DimSize_t, 1>::size_type DIM>
inline std::shared_ptr<Node> Conv(DimSize_t inChannels,
DimSize_t outChannels, const std::array<DimSize_t, DIM> &kernelDims,
const std::string& name = "",
const std::array<DimSize_t, DIM> &strideDims = create_array<DimSize_t,DIM>(1),
const std::array<DimSize_t, DIM> &dilationDims = create_array<DimSize_t,DIM>(1)) {
// FIXME: properly handle default w&b initialization in every cases
static_assert(DIM<=MaxDim,"Too many kernel dimensions required by Conv, not supported");
auto conv = std::make_shared<Node>(std::make_shared<Conv_Op<static_cast<DimIdx_t>(DIM)>>(inChannels, outChannels, kernelDims, strideDims, dilationDims), name);
addProducer(conv, 1, append(outChannels, append(inChannels, kernelDims)), "w");
addProducer(conv, 2, {outChannels}, "b");
return conv;
}
// helper with C-style array instead of std::array for kernel_dims to allow automatic template DIM deduction
template <DimSize_t DIM>
inline std::shared_ptr<Node> Conv(
DimSize_t inChannels,
DimSize_t outChannels,
DimSize_t const (&kernelDims)[DIM],
const std::string& name = "",
const std::array<DimSize_t, DIM> &strideDims = create_array<DimSize_t,DIM>(1),
const std::array<DimSize_t, DIM> &dilationDims = create_array<DimSize_t,DIM>(1)) {
static_assert(DIM<=MaxDim,"Too many kernel dimensions required by Conv, not supported");
return Conv(inChannels, outChannels, to_array(kernelDims), name, strideDims, dilationDims);
}
} // namespace Aidge
namespace {
template <>
const char *const EnumStrings<Aidge::ConvAttr>::data[] = {
"StrideDims",
"DilationDims",
"InChannels",
"OutChannels",
"KernelDims"
};
}
#endif /* AIDGE_CORE_OPERATOR_CONV_H_ */