Working LSTM with any sequence length

include/aidge/backend/cpu.hpp

@@ -27,6 +27,7 @@
 #include "aidge/backend/cpu/operator/MemorizeImpl.hpp"
 #include "aidge/backend/cpu/operator/MulImpl.hpp"
 #include "aidge/backend/cpu/operator/PadImpl.hpp"
+#include "aidge/backend/cpu/operator/PopImpl.hpp"
 #include "aidge/backend/cpu/operator/PowImpl.hpp"
 #include "aidge/backend/cpu/operator/ProducerImpl.hpp"
 #include "aidge/backend/cpu/operator/ReLUImpl.hpp"

include/aidge/backend/cpu/operator/PopImpl.hpp

+/********************************************************************************
+ * 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_CPU_OPERATOR_POPIMPL_H_
+#define AIDGE_CPU_OPERATOR_POPIMPL_H_
+
+#include "aidge/backend/OperatorImpl.hpp"
+#include "aidge/operator/Pop.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/utils/Types.h"
+#include "aidge/backend/cpu/data/GetCPUPtr.h"
+#include <memory>
+#include <vector>
+
+namespace Aidge {
+// class Pop_Op;
+
+// compute kernel registry for forward and backward
+class PopImplForward_cpu
+    : public Registrable<PopImplForward_cpu, std::tuple<DataType, DataType>, void(const std::size_t, const void*, void*)> {
+};
+class PopImplBackward_cpu
+    : public Registrable<PopImplBackward_cpu, std::tuple<DataType, DataType>, void(const std::size_t, const void*, void*)> {
+};
+
+class PopImpl_cpu : public OperatorImpl {
+public:
+    PopImpl_cpu(const Pop_Op& op) : OperatorImpl(op) {}
+
+    static std::unique_ptr<PopImpl_cpu> create(const Pop_Op& op) {
+        return std::make_unique<PopImpl_cpu>(op);
+    }
+
+    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    void forward() override;
+};
+
+namespace {
+static Registrar<Pop_Op> registrarPopImpl_cpu("cpu", Aidge::PopImpl_cpu::create);
+}
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_POPIMPL_H_ */

src/operator/PopImpl.cpp

+/********************************************************************************
+ * 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 <cassert>
+#include <chrono>  // std::chrono::milliseconds
+#include <numeric> // std::accumulate
+#include <thread>  // std::this_thread::sleep_for
+#include <vector>
+
+#include "aidge/operator/Pop.hpp"
+#include "aidge/utils/Types.h"
+#include "aidge/backend/cpu/data/GetCPUPtr.h"
+
+#include "aidge/backend/cpu/operator/PopImpl.hpp"
+
+Aidge::NbElts_t Aidge::PopImpl_cpu::getNbRequiredProtected(const Aidge::IOIndex_t /*inputIdx*/) const {
+    // this implementation can be in-place
+    return 0;
+}
+
+void Aidge::PopImpl_cpu::forward() {
+    assert(std::static_pointer_cast<Tensor>(mOp.getRawInput(0)) && "missing input #0");
+
+    const Pop_Op& op = dynamic_cast<const Pop_Op&>(mOp);
+    const unsigned int forwardStep = op.template getAttr<PopAttr::ForwardStep>();
+
+    *std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))
+        = std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->extract({forwardStep});
+}

unit_tests/operator/Test_MetaOperator.cpp

@@ -14,6 +14,7 @@
 #include <cstdlib>
 #include <memory>
 
+#include "aidge/utils/TensorUtils.hpp"
 #include "aidge/backend/cpu/operator/ConvImpl.hpp"
 #include "aidge/backend/cpu/operator/PadImpl.hpp"
 #include "aidge/data/Tensor.hpp"
@@ -21,6 +22,7 @@
 #include "aidge/operator/MetaOperator.hpp"
 #include "aidge/operator/MetaOperatorDefs.hpp"
 #include "aidge/operator/Pad.hpp"
+#include "aidge/operator/Pop.hpp"
 
 using namespace Aidge;
 
@@ -197,8 +199,8 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
         auto microGraph = std::dynamic_pointer_cast<MetaOperator_Op>(op)->getMicroGraph();
         microGraph->save("lstm", false, false);
 
-        REQUIRE(myLSTM->nbInputs() == 3);
-        REQUIRE(myLSTM->nbData() == 3);
+        REQUIRE(myLSTM->nbInputs() == 3 + 8 + 8);
+        REQUIRE(myLSTM->nbData() == 1);
         REQUIRE(myLSTM->nbOutputs() == 2);
 
         std::shared_ptr<Tensor> myInput = std::make_shared<Tensor>(
@@ -207,12 +209,12 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
             Array2D<float, 1, 64>{{0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0}});
 
         op->associateInput(0, myInput);
-        op->associateInput(1, myInit);
-        op->associateInput(2, myInit);
+        op->associateInput(17, myInit);
+        op->associateInput(18, myInit);
 
         op->computeOutputDims();
+        microGraph->save("lstm_dims", true, true);
         REQUIRE(op->outputDimsForwarded());
-        microGraph->save("lstm_dims", false, false);
 
         op->setDataType(DataType::Float32);
         op->setBackend("cpu");
@@ -230,7 +232,7 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
         microGraph->save("lstm", false, false);
 
         REQUIRE(myLSTM->nbInputs() == 3 + 8 + 8);
-        REQUIRE(myLSTM->nbData() == 3);
+        REQUIRE(myLSTM->nbData() == 1);
         REQUIRE(myLSTM->nbOutputs() == 2);
 
         std::shared_ptr<Tensor> myInput = std::make_shared<Tensor>(
@@ -278,4 +280,69 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
 
         REQUIRE(approxEq<float>(*(op->getOutput(1)), *myHiddenState));
     }
+
+    SECTION("LSTM(forward_values_seq)") {
+        auto pop = Pop();
+        auto myLSTM = LSTM(2, 3, 2, true, "ltsm");
+        auto op = std::static_pointer_cast<MetaOperator_Op>(myLSTM->getOperator());
+
+        // NOTE: LSTM really need to be flatten in the graph before execution.
+        // Here, we actually don't use the meta-op as a closed black-box,
+        // because its scheduling cannot be run independently of the input.
+        // Since we use the Pop operator to generate sequential inputs, running
+        // the meta-op internal scheduler would not work because it would not 
+        // update its input!
+        // We just borrow its micro-graph into our larger myGraph graph.
+        auto myGraph = std::make_shared<GraphView>();
+        myGraph->add(pop);
+        myGraph->add(op->getMicroGraph());
+        pop->addChild(op->getMicroGraph()->getOrderedInputs()[0].first, 0, 0);
+
+        REQUIRE(myLSTM->nbInputs() == 3 + 8 + 8);
+        REQUIRE(myLSTM->nbData() == 1);
+        REQUIRE(myLSTM->nbOutputs() == 2);
+
+        std::shared_ptr<Tensor> myInput = std::make_shared<Tensor>(
+            Array3D<float, 2, 3, 2>{{{{1.0, 2.0}, {3.0, 4.0}, {5.0, 6.0}}, {{2.0, 3.0}, {4.0, 5.0}, {6.0, 7.0}}}});
+        std::shared_ptr<Tensor> myInit = std::make_shared<Tensor>(
+            Array2D<float, 3, 3>{{{0.0, 0.0, 0.0}, {0.0, 0.0, 0.0}, {0.0, 0.0, 0.0}}});
+        std::shared_ptr<Tensor> myInitW = std::make_shared<Tensor>(
+            Array2D<float, 3, 2>{{{0.1, 0.1}, {0.1, 0.1}, {0.1, 0.1}}});
+        std::shared_ptr<Tensor> myInitR = std::make_shared<Tensor>(
+            Array2D<float, 3, 3>{{{0.1, 0.1, 0.1}, {0.1, 0.1, 0.1}, {0.1, 0.1, 0.1}}});
+
+        pop->getOperator()->associateInput(0, myInput);
+        op->associateInput(17, myInit);
+        op->associateInput(18, myInit);
+
+        // Weights X
+        op->associateInput(1, myInitW);
+        op->associateInput(2, myInitW);
+        op->associateInput(3, myInitW);
+        op->associateInput(4, myInitW);
+        // Weights H
+        op->associateInput(5, myInitR);
+        op->associateInput(6, myInitR);
+        op->associateInput(7, myInitR);
+        op->associateInput(8, myInitR);
+
+        myGraph->setDataType(DataType::Float32);
+        myGraph->setBackend("cpu");
+
+        std::shared_ptr<Tensor> myHiddenState = std::make_shared<Tensor>(
+                Array2D<float, 3, 3>{{{0.24439372, 0.24439372, 0.24439372},
+                                     {0.49801484, 0.49801484, 0.49801484},
+                                     {0.67162132, 0.67162132, 0.67162132}}});
+
+        auto scheduler = SequentialScheduler(myGraph);
+        scheduler.forward();
+        scheduler.saveSchedulingDiagram("lstm_seq_schedule");
+
+        myGraph->save("lstm_seq", true, true);
+
+        op->getOutput(1)->print();
+        myHiddenState->print();
+
+        REQUIRE(approxEq<float>(*(op->getOutput(1)), *myHiddenState));
+    }
 }
\ No newline at end of file