From 678c51cee40086b6d35fa2ef4c9dd8e248c1eadc Mon Sep 17 00:00:00 2001
From: NAUD Maxence <maxence.naud@cea.fr>
Date: Mon, 3 Jun 2024 08:48:02 +0000
Subject: [PATCH] Move MetaOp PaddedAvgPooling and LSTM definition to cpp files
---
include/aidge/operator/MetaOperatorDefs.hpp | 215 +++++++----------
src/operator/MetaOperatorDefs/LSTM.cpp | 228 ++++++++++++++++++
.../MetaOperatorDefs/PaddedAvgPooling.cpp | 91 +++++++
3 files changed, 401 insertions(+), 133 deletions(-)
create mode 100644 src/operator/MetaOperatorDefs/LSTM.cpp
create mode 100644 src/operator/MetaOperatorDefs/PaddedAvgPooling.cpp
diff --git a/include/aidge/operator/MetaOperatorDefs.hpp b/include/aidge/operator/MetaOperatorDefs.hpp
index fb3aa6384..eb57761cc 100644
--- a/include/aidge/operator/MetaOperatorDefs.hpp
+++ b/include/aidge/operator/MetaOperatorDefs.hpp
@@ -12,22 +12,26 @@
#ifndef AIDGE_CORE_OPERATOR_METAOPERATORDEFS_H_
#define AIDGE_CORE_OPERATOR_METAOPERATORDEFS_H_
+#include <array>
+#include <memory>
+#include <string>
+
+#include "aidge/graph/GraphView.hpp"
+#include "aidge/graph/Node.hpp"
+#include "aidge/graph/OpArgs.hpp" // Sequential
#include "aidge/operator/MetaOperator.hpp"
#include "aidge/operator/AvgPooling.hpp"
#include "aidge/operator/MaxPooling.hpp"
#include "aidge/operator/Conv.hpp"
#include "aidge/operator/ConvDepthWise.hpp"
#include "aidge/operator/Pad.hpp"
-#include "aidge/operator/Memorize.hpp"
-#include "aidge/operator/Add.hpp"
-#include "aidge/operator/Mul.hpp"
-#include "aidge/operator/FC.hpp"
-#include "aidge/operator/Identity.hpp"
-#include "aidge/operator/Concat.hpp"
-#include "aidge/operator/Tanh.hpp"
#include "aidge/operator/Sigmoid.hpp"
+#include "aidge/utils/ArrayHelpers.hpp"
+#include "aidge/utils/Types.h"
namespace Aidge {
+
+
template <std::array<DimSize_t, 1>::size_type DIM>
inline std::shared_ptr<Node> PaddedConv(DimSize_t in_channels,
DimSize_t out_channels,
@@ -40,7 +44,7 @@ inline std::shared_ptr<Node> PaddedConv(DimSize_t in_channels,
{
// Construct micro-graph
auto pad = Pad<DIM>(padding_dims, (!name.empty()) ? name + "_pad" : "", PadBorderType::Constant, 0.0);
- auto conv = std::make_shared<Node>(std::make_shared<Conv_Op<static_cast<DimIdx_t>(DIM)>>(in_channels, out_channels, kernel_dims, stride_dims, dilation_dims, no_bias), (!name.empty()) ? name + "_conv" : "");
+ auto conv = std::make_shared<Node>(std::make_shared<Conv_Op<static_cast<DimIdx_t>(DIM)>>(kernel_dims, stride_dims, dilation_dims, no_bias), (!name.empty()) ? name + "_conv" : "");
auto metaOp = MetaOperator("PaddedConv", Sequential({pad, conv}), name);
addProducer(metaOp, 1, append(out_channels, append(in_channels, kernel_dims)), "w");
@@ -48,6 +52,20 @@ inline std::shared_ptr<Node> PaddedConv(DimSize_t in_channels,
return metaOp;
}
+template <std::array<DimSize_t, 1>::size_type DIM>
+inline std::shared_ptr<MetaOperator_Op> PaddedConv_Op(
+ const std::array<DimSize_t, DIM> &kernel_dims,
+ const std::array<DimSize_t, DIM> &stride_dims = create_array<DimSize_t,DIM>(1),
+ const std::array<DimSize_t, 2*DIM> &padding_dims = create_array<DimSize_t,2*DIM>(0),
+ const std::array<DimSize_t, DIM> &dilation_dims = create_array<DimSize_t,DIM>(1),
+ bool no_bias = false)
+{
+ auto pad = Pad<DIM>(padding_dims, "", PadBorderType::Constant, 0.0);
+ auto conv = std::make_shared<Node>(std::make_shared<Conv_Op<static_cast<DimIdx_t>(DIM)>>(kernel_dims, stride_dims, dilation_dims, no_bias), "");
+
+ return std::make_shared<MetaOperator_Op>("PaddedConv", Sequential({pad, 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> PaddedConv(
@@ -63,6 +81,8 @@ inline std::shared_ptr<Node> PaddedConv(
return PaddedConv(in_channels, out_channels, to_array(kernel_dims), name, stride_dims, padding_dims, dilation_dims, no_bias);
}
+////////////////////////////////////////////////////////////////////////////////
+
template <std::array<DimSize_t, 1>::size_type DIM>
inline std::shared_ptr<Node> PaddedConvDepthWise(const DimSize_t nb_channels,
const std::array<DimSize_t, DIM> &kernel_dims,
@@ -74,7 +94,7 @@ inline std::shared_ptr<Node> PaddedConvDepthWise(const DimSize_t nb_channels,
{
// Construct micro-graph
auto pad = Pad<DIM>(padding_dims, (!name.empty()) ? name + "_pad" : "", PadBorderType::Constant, 0.0);
- auto conv = std::make_shared<Node>(std::make_shared<ConvDepthWise_Op<static_cast<DimIdx_t>(DIM)>>(nb_channels, kernel_dims, stride_dims, dilation_dims, no_bias), (!name.empty()) ? name + "_conv" : "");
+ auto conv = std::make_shared<Node>(std::make_shared<ConvDepthWise_Op<static_cast<DimIdx_t>(DIM)>>(kernel_dims, stride_dims, dilation_dims, no_bias), (!name.empty()) ? name + "_conv" : "");
auto metaOp = MetaOperator("PaddedConvDepthWise", Sequential({pad, conv}), name);
addProducer(metaOp, 1, append(nb_channels, append(DimSize_t(1), kernel_dims)), "w");
@@ -82,6 +102,20 @@ inline std::shared_ptr<Node> PaddedConvDepthWise(const DimSize_t nb_channels,
return metaOp;
}
+template <std::array<DimSize_t, 1>::size_type DIM>
+inline std::shared_ptr<MetaOperator_Op> PaddedConvDepthWise_Op(
+ const std::array<DimSize_t, DIM> &kernel_dims,
+ const std::array<DimSize_t, DIM> &stride_dims = create_array<DimSize_t,DIM>(1),
+ const std::array<DimSize_t, 2*DIM> &padding_dims = create_array<DimSize_t,2*DIM>(0),
+ const std::array<DimSize_t, DIM> &dilation_dims = create_array<DimSize_t,DIM>(1),
+ bool no_bias = false)
+{
+ auto pad = Pad<DIM>(padding_dims, "", PadBorderType::Constant, 0.0);
+ auto conv = std::make_shared<Node>(std::make_shared<ConvDepthWise_Op<static_cast<DimIdx_t>(DIM)>>(kernel_dims, stride_dims, dilation_dims, no_bias), "");
+
+ return std::make_shared<MetaOperator_Op>("PaddedConvDepthWise", Sequential({pad, 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> PaddedConvDepthWise(
@@ -96,30 +130,29 @@ inline std::shared_ptr<Node> PaddedConvDepthWise(
return PaddedConvDepthWise(nb_channels, to_array(kernel_dims), name, stride_dims, padding_dims, dilation_dims, no_bias);
}
+////////////////////////////////////////////////////////////////////////////////
+
+
template <std::array<DimSize_t, 1>::size_type DIM>
-inline std::shared_ptr<Node> PaddedAvgPooling(const std::array<DimSize_t, DIM> &kernel_dims,
+extern std::shared_ptr<Node> PaddedAvgPooling(const std::array<DimSize_t, DIM> &kernel_dims,
const std::string& name = "",
const std::array<DimSize_t, DIM> &stride_dims = create_array<DimSize_t,DIM>(1),
- const std::array<DimSize_t, 2*DIM> &padding_dims = create_array<DimSize_t,2*DIM>(0))
-{
- auto graph = Sequential({
- Pad<DIM>(padding_dims, (!name.empty()) ? name + "_pad" : ""),
- AvgPooling(kernel_dims, (!name.empty()) ? name + "_avgpooling" : "", stride_dims)
- });
+ const std::array<DimSize_t, 2*DIM> &padding_dims = create_array<DimSize_t,2*DIM>(0));
- return MetaOperator("PaddedAvgPooling", graph, name);
-}
+
+template <std::array<DimSize_t, 1>::size_type DIM>
+extern std::shared_ptr<MetaOperator_Op> PaddedAvgPooling_Op(const std::array<DimSize_t, DIM> &kernel_dims,
+ const std::array<DimSize_t, DIM> &stride_dims = create_array<DimSize_t,DIM>(1),
+ const std::array<DimSize_t, 2*DIM> &padding_dims = create_array<DimSize_t,2*DIM>(0));
// 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> PaddedAvgPooling(
- DimSize_t const (&kernel_dims)[DIM],
+extern std::shared_ptr<Node> PaddedAvgPooling(DimSize_t const (&kernel_dims)[DIM],
const std::string& name = "",
const std::array<DimSize_t, DIM> &stride_dims = create_array<DimSize_t,DIM>(1),
- const std::array<DimSize_t, 2*DIM> &padding_dims = create_array<DimSize_t,2*DIM>(0))
-{
- return PaddedAvgPooling(to_array(kernel_dims), name, stride_dims, padding_dims);
-}
+ const std::array<DimSize_t, 2*DIM> &padding_dims = create_array<DimSize_t,2*DIM>(0));
+
+////////////////////////////////////////////////////////////////////////////////
template <std::array<DimSize_t, 1>::size_type DIM>
inline std::shared_ptr<Node> PaddedMaxPooling(const std::array<DimSize_t, DIM> &kernel_dims,
@@ -136,6 +169,20 @@ inline std::shared_ptr<Node> PaddedMaxPooling(const std::array<DimSize_t, DIM> &
return MetaOperator("PaddedMaxPooling", graph, name);
}
+template <std::array<DimSize_t, 1>::size_type DIM>
+inline std::shared_ptr<MetaOperator_Op> PaddedMaxPooling_Op(const std::array<DimSize_t, DIM> &kernel_dims,
+ const std::array<DimSize_t, DIM> &stride_dims = create_array<DimSize_t,DIM>(1),
+ const std::array<DimSize_t, 2*DIM> &padding_dims = create_array<DimSize_t,2*DIM>(0),
+ bool ceil_mode = false)
+{
+ auto graph = Sequential({
+ Pad<DIM>(padding_dims, ""),
+ MaxPooling(kernel_dims, "", stride_dims, ceil_mode)
+ });
+
+ return std::make_shared<MetaOperator_Op>("PaddedMaxPooling", graph);
+}
+
// 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> PaddedMaxPooling(
@@ -148,115 +195,17 @@ inline std::shared_ptr<Node> PaddedMaxPooling(
return PaddedMaxPooling(to_array(kernel_dims), name, stride_dims, padding_dims, ceil_mode);
}
-inline std::shared_ptr<Node> LSTM(DimSize_t in_channels,
- DimSize_t hidden_channels,
- DimSize_t seq_length,
- bool noBias = false,
- const std::string& name = "")
-{
- // Construct micro-graph
- auto input = Identity((!name.empty()) ? name + "_input" : "");
- auto hiddenState = Memorize(seq_length, (!name.empty()) ? name + "_hidden_state" : "");
- auto cellState = Memorize(seq_length, (!name.empty()) ? name + "_cell_state" : "");
- auto add = Add(2, (!name.empty()) ? name + "_add" : "");
-
- // Forget gate
- auto forgetGateX = std::make_shared<Node>(std::make_shared<FC_Op>(hidden_channels, noBias), (!name.empty()) ? name + "_forgetGateX" : "");
- input->addChild(forgetGateX, 0, 0);
- auto forgetGateH = std::make_shared<Node>(std::make_shared<FC_Op>(hidden_channels, noBias), (!name.empty()) ? name + "_forgetGateH" : "");
- hiddenState->addChild(forgetGateH, 1, 0);
- auto forgetGate = Add(2, (!name.empty()) ? name + "_forgetGate" : "");
- forgetGateX->addChild(forgetGate, 0, 0);
- forgetGateH->addChild(forgetGate, 0, 1);
- auto forgetGateAct = Sigmoid((!name.empty()) ? name + "_forgetGateAct" : "");
- auto forgetGateMul = Mul((!name.empty()) ? name + "_forgetGateMul" : "");
- forgetGate->addChild(forgetGateAct, 0, 0);
- forgetGateAct->addChild(forgetGateMul, 0, 0);
- forgetGateMul->addChild(add, 0, 0);
- cellState->addChild(forgetGateMul, 1, 1);
-
- // Input gate
- auto inputGateX = std::make_shared<Node>(std::make_shared<FC_Op>(hidden_channels, noBias), (!name.empty()) ? name + "_inputGateX" : "");
- input->addChild(inputGateX, 0, 0);
- auto inputGateH = std::make_shared<Node>(std::make_shared<FC_Op>(hidden_channels, noBias), (!name.empty()) ? name + "_inputGateH" : "");
- hiddenState->addChild(inputGateH, 1, 0);
- auto inputGate = Add(2, (!name.empty()) ? name + "_inputGate" : "");
- inputGateX->addChild(inputGate, 0, 0);
- inputGateH->addChild(inputGate, 0, 1);
- auto inputGateAct = Sigmoid((!name.empty()) ? name + "_inputGateAct" : "");
- auto inputGateMul = Mul((!name.empty()) ? name + "_inputGateMul" : "");
- inputGate->addChild(inputGateAct, 0, 0);
- inputGateAct->addChild(inputGateMul, 0, 0);
- inputGateMul->addChild(add, 0, 1);
-
- // Candidate for cell update
- auto cellCandidateX = std::make_shared<Node>(std::make_shared<FC_Op>(hidden_channels, noBias), (!name.empty()) ? name + "_cellCandidateX" : "");
- input->addChild(cellCandidateX, 0, 0);
- auto cellCandidateH = std::make_shared<Node>(std::make_shared<FC_Op>(hidden_channels, noBias), (!name.empty()) ? name + "_cellCandidateH" : "");
- hiddenState->addChild(cellCandidateH, 1, 0);
- auto cellCandidate = Add(2, (!name.empty()) ? name + "_cellCandidate" : "");
- cellCandidateX->addChild(cellCandidate, 0, 0);
- cellCandidateH->addChild(cellCandidate, 0, 1);
- auto cellCandidateAct = Tanh((!name.empty()) ? name + "_cellCandidateAct" : "");
- cellCandidate->addChild(cellCandidateAct, 0, 0);
- cellCandidateAct->addChild(inputGateMul, 0, 1);
-
- // Output gate
- auto outputGateX = std::make_shared<Node>(std::make_shared<FC_Op>(hidden_channels, noBias), (!name.empty()) ? name + "_outputGateX" : "");
- input->addChild(outputGateX, 0, 0);
- auto outputGateH = std::make_shared<Node>(std::make_shared<FC_Op>(hidden_channels, noBias), (!name.empty()) ? name + "_outputGateH" : "");
- hiddenState->addChild(outputGateH, 1, 0);
- auto outputGate = Add(2, (!name.empty()) ? name + "_outputGate" : "");
- outputGateX->addChild(outputGate, 0, 0);
- outputGateH->addChild(outputGate, 0, 1);
- auto outputGateAct = Sigmoid((!name.empty()) ? name + "_outputGateAct" : "");
- auto outputGateMul = Mul((!name.empty()) ? name + "_outputGateMul" : "");
- outputGate->addChild(outputGateAct, 0, 0);
- outputGateAct->addChild(outputGateMul, 0, 0);
-
- // Updated cell state to help determine new hidden state
- auto cellUpdatedAct = Tanh((!name.empty()) ? name + "_cellUpdatedAct" : "");
- add->addChild(cellUpdatedAct, 0, 0);
- cellUpdatedAct->addChild(outputGateMul, 0, 1);
- outputGateMul->addChild(hiddenState, 0, 0);
- add->addChild(cellState, 0, 0);
-
- std::shared_ptr<GraphView> microGraph = std::make_shared<GraphView>();
- microGraph->add(input);
- microGraph->add({hiddenState, cellState, add,
- forgetGateX, forgetGateH, forgetGate, forgetGateAct, forgetGateMul,
- inputGateX, inputGateH, inputGate, inputGateAct, inputGateMul,
- cellCandidateX, cellCandidateH, cellCandidate, cellCandidateAct,
- outputGateX, outputGateH, outputGate, outputGateAct, outputGateMul,
- cellUpdatedAct}, false);
-
- microGraph->setOrderedInputs({{input, 0},
- {inputGateX, 1}, {outputGateX, 1}, {forgetGateX, 1}, {cellCandidateX, 1},
- {inputGateH, 1}, {outputGateH, 1}, {forgetGateH, 1}, {cellCandidateH, 1},
- {inputGateX, 2}, {outputGateX, 2}, {forgetGateX, 2}, {cellCandidateX, 2},
- {inputGateH, 2}, {outputGateH, 2}, {forgetGateH, 2}, {cellCandidateH, 2},
- {hiddenState, 1}, {cellState, 1}});
- microGraph->setOrderedOutputs({{hiddenState, 0}, {cellState, 0}});
-
- auto metaOp = MetaOperator("LSTM", microGraph, name);
- addProducer(metaOp, 1, {hidden_channels, in_channels}, "wi");
- addProducer(metaOp, 2, {hidden_channels, in_channels}, "wo");
- addProducer(metaOp, 3, {hidden_channels, in_channels}, "wf");
- addProducer(metaOp, 4, {hidden_channels, in_channels}, "wc");
- addProducer(metaOp, 5, {hidden_channels, hidden_channels}, "ri");
- addProducer(metaOp, 6, {hidden_channels, hidden_channels}, "ro");
- addProducer(metaOp, 7, {hidden_channels, hidden_channels}, "rf");
- addProducer(metaOp, 8, {hidden_channels, hidden_channels}, "rc");
- addProducer(metaOp, 9, {(noBias ? 0 : hidden_channels)}, "wbi");
- addProducer(metaOp, 10, {(noBias ? 0 : hidden_channels)}, "wbo");
- addProducer(metaOp, 11, {(noBias ? 0 : hidden_channels)}, "wbf");
- addProducer(metaOp, 12, {(noBias ? 0 : hidden_channels)}, "wbc");
- addProducer(metaOp, 13, {(noBias ? 0 : hidden_channels)}, "rbi");
- addProducer(metaOp, 14, {(noBias ? 0 : hidden_channels)}, "rbo");
- addProducer(metaOp, 15, {(noBias ? 0 : hidden_channels)}, "rbf");
- addProducer(metaOp, 16, {(noBias ? 0 : hidden_channels)}, "rbc");
- return metaOp;
-}
+////////////////////////////////////////////////////////////////////////////////
+
+std::shared_ptr<Node> LSTM(DimSize_t in_channels,
+ DimSize_t hidden_channels,
+ DimSize_t seq_length,
+ bool noBias = false,
+ const std::string& name = "");
+
+std::shared_ptr<MetaOperator_Op> LSTM_Op(DimSize_t seq_length,
+ bool noBias = false);
+
} // namespace Aidge
#endif /* AIDGE_CORE_OPERATOR_METAOPERATORDEFS_H_ */
diff --git a/src/operator/MetaOperatorDefs/LSTM.cpp b/src/operator/MetaOperatorDefs/LSTM.cpp
new file mode 100644
index 000000000..cd993f9e5
--- /dev/null
+++ b/src/operator/MetaOperatorDefs/LSTM.cpp
@@ -0,0 +1,228 @@
+/********************************************************************************
+ * 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
+ *
+ ********************************************************************************/
+
+#include "aidge/operator/MetaOperatorDefs.hpp"
+
+#include <array>
+#include <memory>
+#include <string>
+
+#include "aidge/operator/Memorize.hpp"
+#include "aidge/operator/Add.hpp"
+#include "aidge/operator/Mul.hpp"
+#include "aidge/operator/FC.hpp"
+#include "aidge/operator/Identity.hpp"
+#include "aidge/operator/Concat.hpp"
+#include "aidge/operator/Tanh.hpp"
+
+namespace Aidge {
+
+std::shared_ptr<Node> LSTM(const DimSize_t inChannel,
+ const DimSize_t hiddenChannel,
+ const DimSize_t seqLength,
+ bool noBias,
+ const std::string& name)
+{
+ // Construct micro-graph
+ auto input = Identity((!name.empty()) ? name + "_input" : "");
+ auto hiddenState = Memorize(seqLength, (!name.empty()) ? name + "_hidden_state" : "");
+ auto cellState = Memorize(seqLength, (!name.empty()) ? name + "_cell_state" : "");
+ auto add = Add(2, (!name.empty()) ? name + "_add" : "");
+
+ // Forget gate
+ auto forgetGateX = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), (!name.empty()) ? name + "_forgetGateX" : "");
+ input->addChild(forgetGateX, 0, 0);
+ auto forgetGateH = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), (!name.empty()) ? name + "_forgetGateH" : "");
+ hiddenState->addChild(forgetGateH, 1, 0);
+ auto forgetGate = Add(2, (!name.empty()) ? name + "_forgetGate" : "");
+ forgetGateX->addChild(forgetGate, 0, 0);
+ forgetGateH->addChild(forgetGate, 0, 1);
+ auto forgetGateAct = Sigmoid((!name.empty()) ? name + "_forgetGateAct" : "");
+ auto forgetGateMul = Mul((!name.empty()) ? name + "_forgetGateMul" : "");
+ forgetGate->addChild(forgetGateAct, 0, 0);
+ forgetGateAct->addChild(forgetGateMul, 0, 0);
+ forgetGateMul->addChild(add, 0, 0);
+ cellState->addChild(forgetGateMul, 1, 1);
+
+ // Input gate
+ auto inputGateX = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), (!name.empty()) ? name + "_inputGateX" : "");
+ input->addChild(inputGateX, 0, 0);
+ auto inputGateH = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), (!name.empty()) ? name + "_inputGateH" : "");
+ hiddenState->addChild(inputGateH, 1, 0);
+ auto inputGate = Add(2, (!name.empty()) ? name + "_inputGate" : "");
+ inputGateX->addChild(inputGate, 0, 0);
+ inputGateH->addChild(inputGate, 0, 1);
+ auto inputGateAct = Sigmoid((!name.empty()) ? name + "_inputGateAct" : "");
+ auto inputGateMul = Mul((!name.empty()) ? name + "_inputGateMul" : "");
+ inputGate->addChild(inputGateAct, 0, 0);
+ inputGateAct->addChild(inputGateMul, 0, 0);
+ inputGateMul->addChild(add, 0, 1);
+
+ // Candidate for cell update
+ auto cellCandidateX = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), (!name.empty()) ? name + "_cellCandidateX" : "");
+ input->addChild(cellCandidateX, 0, 0);
+ auto cellCandidateH = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), (!name.empty()) ? name + "_cellCandidateH" : "");
+ hiddenState->addChild(cellCandidateH, 1, 0);
+ auto cellCandidate = Add(2, (!name.empty()) ? name + "_cellCandidate" : "");
+ cellCandidateX->addChild(cellCandidate, 0, 0);
+ cellCandidateH->addChild(cellCandidate, 0, 1);
+ auto cellCandidateAct = Tanh((!name.empty()) ? name + "_cellCandidateAct" : "");
+ cellCandidate->addChild(cellCandidateAct, 0, 0);
+ cellCandidateAct->addChild(inputGateMul, 0, 1);
+
+ // Output gate
+ auto outputGateX = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), (!name.empty()) ? name + "_outputGateX" : "");
+ input->addChild(outputGateX, 0, 0);
+ auto outputGateH = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), (!name.empty()) ? name + "_outputGateH" : "");
+ hiddenState->addChild(outputGateH, 1, 0);
+ auto outputGate = Add(2, (!name.empty()) ? name + "_outputGate" : "");
+ outputGateX->addChild(outputGate, 0, 0);
+ outputGateH->addChild(outputGate, 0, 1);
+ auto outputGateAct = Sigmoid((!name.empty()) ? name + "_outputGateAct" : "");
+ auto outputGateMul = Mul((!name.empty()) ? name + "_outputGateMul" : "");
+ outputGate->addChild(outputGateAct, 0, 0);
+ outputGateAct->addChild(outputGateMul, 0, 0);
+
+ // Updated cell state to help determine new hidden state
+ auto cellUpdatedAct = Tanh((!name.empty()) ? name + "_cellUpdatedAct" : "");
+ add->addChild(cellUpdatedAct, 0, 0);
+ cellUpdatedAct->addChild(outputGateMul, 0, 1);
+ outputGateMul->addChild(hiddenState, 0, 0);
+ add->addChild(cellState, 0, 0);
+
+ std::shared_ptr<GraphView> microGraph = std::make_shared<GraphView>();
+ microGraph->add(input);
+ microGraph->add({hiddenState, cellState, add,
+ forgetGateX, forgetGateH, forgetGate, forgetGateAct, forgetGateMul,
+ inputGateX, inputGateH, inputGate, inputGateAct, inputGateMul,
+ cellCandidateX, cellCandidateH, cellCandidate, cellCandidateAct,
+ outputGateX, outputGateH, outputGate, outputGateAct, outputGateMul,
+ cellUpdatedAct}, false);
+
+ microGraph->setOrderedInputs({{input, 0},
+ {inputGateX, 1}, {outputGateX, 1}, {forgetGateX, 1}, {cellCandidateX, 1},
+ {inputGateH, 1}, {outputGateH, 1}, {forgetGateH, 1}, {cellCandidateH, 1},
+ {inputGateX, 2}, {outputGateX, 2}, {forgetGateX, 2}, {cellCandidateX, 2},
+ {inputGateH, 2}, {outputGateH, 2}, {forgetGateH, 2}, {cellCandidateH, 2},
+ {hiddenState, 1}, {cellState, 1}});
+ microGraph->setOrderedOutputs({{hiddenState, 0}, {cellState, 0}});
+
+ auto metaOp = MetaOperator("LSTM", microGraph, name);
+ addProducer(metaOp, 1, {hiddenChannel, inChannel}, "wi");
+ addProducer(metaOp, 2, {hiddenChannel, inChannel}, "wo");
+ addProducer(metaOp, 3, {hiddenChannel, inChannel}, "wf");
+ addProducer(metaOp, 4, {hiddenChannel, inChannel}, "wc");
+ addProducer(metaOp, 5, {hiddenChannel, hiddenChannel}, "ri");
+ addProducer(metaOp, 6, {hiddenChannel, hiddenChannel}, "ro");
+ addProducer(metaOp, 7, {hiddenChannel, hiddenChannel}, "rf");
+ addProducer(metaOp, 8, {hiddenChannel, hiddenChannel}, "rc");
+ addProducer(metaOp, 9, {(noBias ? 0 : hiddenChannel)}, "wbi");
+ addProducer(metaOp, 10, {(noBias ? 0 : hiddenChannel)}, "wbo");
+ addProducer(metaOp, 11, {(noBias ? 0 : hiddenChannel)}, "wbf");
+ addProducer(metaOp, 12, {(noBias ? 0 : hiddenChannel)}, "wbc");
+ addProducer(metaOp, 13, {(noBias ? 0 : hiddenChannel)}, "rbi");
+ addProducer(metaOp, 14, {(noBias ? 0 : hiddenChannel)}, "rbo");
+ addProducer(metaOp, 15, {(noBias ? 0 : hiddenChannel)}, "rbf");
+ addProducer(metaOp, 16, {(noBias ? 0 : hiddenChannel)}, "rbc");
+ return metaOp;
+}
+
+std::shared_ptr<MetaOperator_Op> LSTM_Op(const DimSize_t seqLength,
+ bool noBias)
+{
+ // Construct micro-graph
+ auto input = Identity("");
+ auto hiddenState = Memorize(seqLength, "");
+ auto cellState = Memorize(seqLength, "");
+ auto add = Add(2, "");
+
+ // Forget gate
+ auto forgetGateX = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), "");
+ input->addChild(forgetGateX, 0, 0);
+ auto forgetGateH = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), "");
+ hiddenState->addChild(forgetGateH, 1, 0);
+ auto forgetGate = Add(2, "");
+ forgetGateX->addChild(forgetGate, 0, 0);
+ forgetGateH->addChild(forgetGate, 0, 1);
+ auto forgetGateAct = Sigmoid("");
+ auto forgetGateMul = Mul("");
+ forgetGate->addChild(forgetGateAct, 0, 0);
+ forgetGateAct->addChild(forgetGateMul, 0, 0);
+ forgetGateMul->addChild(add, 0, 0);
+ cellState->addChild(forgetGateMul, 1, 1);
+
+ // Input gate
+ auto inputGateX = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), "");
+ input->addChild(inputGateX, 0, 0);
+ auto inputGateH = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), "");
+ hiddenState->addChild(inputGateH, 1, 0);
+ auto inputGate = Add(2, "");
+ inputGateX->addChild(inputGate, 0, 0);
+ inputGateH->addChild(inputGate, 0, 1);
+ auto inputGateAct = Sigmoid("");
+ auto inputGateMul = Mul("");
+ inputGate->addChild(inputGateAct, 0, 0);
+ inputGateAct->addChild(inputGateMul, 0, 0);
+ inputGateMul->addChild(add, 0, 1);
+
+ // Candidate for cell update
+ auto cellCandidateX = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), "");
+ input->addChild(cellCandidateX, 0, 0);
+ auto cellCandidateH = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), "");
+ hiddenState->addChild(cellCandidateH, 1, 0);
+ auto cellCandidate = Add(2, "");
+ cellCandidateX->addChild(cellCandidate, 0, 0);
+ cellCandidateH->addChild(cellCandidate, 0, 1);
+ auto cellCandidateAct = Tanh("");
+ cellCandidate->addChild(cellCandidateAct, 0, 0);
+ cellCandidateAct->addChild(inputGateMul, 0, 1);
+
+ // Output gate
+ auto outputGateX = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), "");
+ input->addChild(outputGateX, 0, 0);
+ auto outputGateH = std::make_shared<Node>(std::make_shared<FC_Op>(noBias), "");
+ hiddenState->addChild(outputGateH, 1, 0);
+ auto outputGate = Add(2,"");
+ outputGateX->addChild(outputGate, 0, 0);
+ outputGateH->addChild(outputGate, 0, 1);
+ auto outputGateAct = Sigmoid("");
+ auto outputGateMul = Mul("");
+ outputGate->addChild(outputGateAct, 0, 0);
+ outputGateAct->addChild(outputGateMul, 0, 0);
+
+ // Updated cell state to help determine new hidden state
+ auto cellUpdatedAct = Tanh("");
+ add->addChild(cellUpdatedAct, 0, 0);
+ cellUpdatedAct->addChild(outputGateMul, 0, 1);
+ outputGateMul->addChild(hiddenState, 0, 0);
+ add->addChild(cellState, 0, 0);
+
+ std::shared_ptr<GraphView> microGraph = std::make_shared<GraphView>();
+ microGraph->add(input);
+ microGraph->add({hiddenState, cellState, add,
+ forgetGateX, forgetGateH, forgetGate, forgetGateAct, forgetGateMul,
+ inputGateX, inputGateH, inputGate, inputGateAct, inputGateMul,
+ cellCandidateX, cellCandidateH, cellCandidate, cellCandidateAct,
+ outputGateX, outputGateH, outputGate, outputGateAct, outputGateMul,
+ cellUpdatedAct}, false);
+
+ microGraph->setOrderedInputs({{input, 0},
+ {inputGateX, 1}, {outputGateX, 1}, {forgetGateX, 1}, {cellCandidateX, 1},
+ {inputGateH, 1}, {outputGateH, 1}, {forgetGateH, 1}, {cellCandidateH, 1},
+ {inputGateX, 2}, {outputGateX, 2}, {forgetGateX, 2}, {cellCandidateX, 2},
+ {inputGateH, 2}, {outputGateH, 2}, {forgetGateH, 2}, {cellCandidateH, 2},
+ {hiddenState, 1}, {cellState, 1}});
+ microGraph->setOrderedOutputs({{hiddenState, 0}, {cellState, 0}});
+
+ return std::make_shared<MetaOperator_Op>("LSTM", microGraph);
+}
+
+} // namespace Aidge
diff --git a/src/operator/MetaOperatorDefs/PaddedAvgPooling.cpp b/src/operator/MetaOperatorDefs/PaddedAvgPooling.cpp
new file mode 100644
index 000000000..ad300cd4f
--- /dev/null
+++ b/src/operator/MetaOperatorDefs/PaddedAvgPooling.cpp
@@ -0,0 +1,91 @@
+/********************************************************************************
+ * 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
+ *
+ ********************************************************************************/
+
+#include "aidge/operator/MetaOperatorDefs.hpp"
+
+#include <array>
+#include <memory>
+#include <string>
+
+#include "aidge/graph/GraphView.hpp"
+#include "aidge/graph/Node.hpp"
+#include "aidge/graph/OpArgs.hpp"
+#include "aidge/operator/AvgPooling.hpp"
+#include "aidge/operator/MetaOperator.hpp"
+#include "aidge/operator/Pad.hpp"
+#include "aidge/utils/ArrayHelpers.hpp"
+#include "aidge/utils/Types.h"
+
+namespace Aidge {
+
+//////////////////////////////////
+// Node functions
+//////////////////////////////////
+
+template <std::array<DimSize_t, 1>::size_type DIM>
+std::shared_ptr<Node> PaddedAvgPooling(const std::array<DimSize_t, DIM> &kernel_dims,
+ const std::string& name,
+ const std::array<DimSize_t, DIM> &stride_dims,
+ const std::array<DimSize_t, 2*DIM> &padding_dims)
+{
+ auto graph = Sequential({
+ Pad<DIM>(padding_dims, (!name.empty()) ? name + "_pad" : ""),
+ AvgPooling(kernel_dims, (!name.empty()) ? name + "_avgpooling" : "", stride_dims)
+ });
+
+ return MetaOperator("PaddedAvgPooling", graph, name);
+}
+
+template std::shared_ptr<Node> PaddedAvgPooling<1>(const std::array<DimSize_t,1>&, const std::string&, const std::array<DimSize_t,1>&, const std::array<DimSize_t,2>&);
+template std::shared_ptr<Node> PaddedAvgPooling<2>(const std::array<DimSize_t,2>&, const std::string&, const std::array<DimSize_t,2>&, const std::array<DimSize_t,4>&);
+template std::shared_ptr<Node> PaddedAvgPooling<3>(const std::array<DimSize_t,3>&, const std::string&, const std::array<DimSize_t,3>&, const std::array<DimSize_t,6>&);
+template std::shared_ptr<Node> PaddedAvgPooling<4>(const std::array<DimSize_t,4>&, const std::string&, const std::array<DimSize_t,4>&, const std::array<DimSize_t,8>&);
+
+// helper with C-style array instead of std::array for kernel_dims to allow automatic template DIM deduction
+template <DimSize_t DIM>
+std::shared_ptr<Node> PaddedAvgPooling(const DimSize_t (&kernel_dims)[DIM],
+ const std::string& name,
+ const std::array<DimSize_t, DIM> &stride_dims,
+ const std::array<DimSize_t, 2*DIM> &padding_dims)
+{
+ return PaddedAvgPooling(to_array(kernel_dims), name, stride_dims, padding_dims);
+}
+
+template std::shared_ptr<Node> PaddedAvgPooling<1>(const DimSize_t (&kernel_dims)[1], const std::string&, const std::array<DimSize_t,1>&, const std::array<DimSize_t,2>&);
+template std::shared_ptr<Node> PaddedAvgPooling<2>(const DimSize_t (&kernel_dims)[2], const std::string&, const std::array<DimSize_t,2>&, const std::array<DimSize_t,4>&);
+template std::shared_ptr<Node> PaddedAvgPooling<3>(const DimSize_t (&kernel_dims)[3], const std::string&, const std::array<DimSize_t,3>&, const std::array<DimSize_t,6>&);
+template std::shared_ptr<Node> PaddedAvgPooling<4>(const DimSize_t (&kernel_dims)[4], const std::string&, const std::array<DimSize_t,4>&, const std::array<DimSize_t,8>&);
+
+
+//////////////////////////////////
+// Operator functions
+//////////////////////////////////
+
+template <std::array<DimSize_t, 1>::size_type DIM>
+inline std::shared_ptr<MetaOperator_Op> PaddedAvgPooling_Op(const std::array<DimSize_t, DIM> &kernel_dims,
+ const std::array<DimSize_t, DIM> &stride_dims,
+ const std::array<DimSize_t, 2*DIM> &padding_dims)
+{
+ auto graph = Sequential({
+ Pad<DIM>(padding_dims, ""),
+ AvgPooling(kernel_dims, "", stride_dims)
+ });
+
+ return std::make_shared<MetaOperator_Op>("PaddedAvgPooling", graph);
+}
+
+template std::shared_ptr<MetaOperator_Op> PaddedAvgPooling_Op<1>(const std::array<DimSize_t,1>&, const std::array<DimSize_t,1>&, const std::array<DimSize_t,2>&);
+template std::shared_ptr<MetaOperator_Op> PaddedAvgPooling_Op<2>(const std::array<DimSize_t,2>&, const std::array<DimSize_t,2>&, const std::array<DimSize_t,4>&);
+template std::shared_ptr<MetaOperator_Op> PaddedAvgPooling_Op<3>(const std::array<DimSize_t,3>&, const std::array<DimSize_t,3>&, const std::array<DimSize_t,6>&);
+template std::shared_ptr<MetaOperator_Op> PaddedAvgPooling_Op<4>(const std::array<DimSize_t,4>&, const std::array<DimSize_t,4>&, const std::array<DimSize_t,8>&);
+
+
+} // namespace Aidge
--
GitLab