version 0.1.2

include/aidge/learning/optimizer/Adam.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_CORE_OPTIMIZER_ADAM_H_
+#define AIDGE_CORE_OPTIMIZER_ADAM_H_
+
+#include <functional>
+#include <memory>
+#include <vector>
+#include <cmath>  // std::sqrt, std::pow
+
+#include "aidge/data/Tensor.hpp"
+#include "aidge/learning/optimizer/Optimizer.hpp"
+#include "aidge/utils/StaticAttributes.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/utils/TensorUtils.hpp"
+
+namespace Aidge {
+
+enum class AdamAttr {
+    Beta1,
+    Beta2,
+    Epsilon
+};
+
+class Adam: public Optimizer, public StaticAttributes<AdamAttr, float, float, float> {
+private:
+    std::vector<Tensor> mMomentum1;
+    std::vector<Tensor> mMomentum2;
+    Tensor mLR{std::vector<std::size_t>({1})};
+    Tensor mBeta1{std::vector<std::size_t>({1})};
+    Tensor mReversedBeta1{std::vector<std::size_t>({1})};
+    Tensor mBeta2{std::vector<std::size_t>({1})};
+    Tensor mReversedBeta2{std::vector<std::size_t>({1})};
+    Tensor mEpsilon{std::vector<std::size_t>({1})};
+
+public:
+    using Attributes_ = StaticAttributes<AdamAttr, float, float, float>;
+    template <AdamAttr e>
+    using attr = typename Attributes_::template attr<e>;
+
+    Adam(const float beta1 = 0.9f, const float beta2 = 0.999f, const float epsilon = 1.0e-8f)
+        : Optimizer(),
+          Attributes_(attr<AdamAttr::Beta1>(beta1),
+                      attr<AdamAttr::Beta2>(beta2),
+                      attr<AdamAttr::Epsilon>(epsilon))
+    {
+        mBeta1.setBackend("cpu");
+        mBeta1.set<float>(0, beta1);
+        mReversedBeta1.setBackend("cpu");
+        mReversedBeta1.set<float>(0, 1.0f - beta1);
+		
+        mBeta2.setBackend("cpu");
+        mBeta2.set<float>(0, beta2);
+        mReversedBeta2.setBackend("cpu");
+        mReversedBeta2.set<float>(0, 1.0f - beta2);
+		
+        mEpsilon.setBackend("cpu");
+        mEpsilon.set<float>(0, epsilon);
+    }
+
+    void update() override final {		
+        mLR.setBackend(mParameters[0]->getImpl()->backend());
+        mLR.set<float>(0, learningRate());
+        if (mParameters[0]->getImpl()->backend() != mBeta1.getImpl()->backend()) {
+            mBeta1.setBackend(mParameters[0]->getImpl()->backend());
+            mReversedBeta1.setBackend(mParameters[0]->getImpl()->backend());
+            mBeta2.setBackend(mParameters[0]->getImpl()->backend());
+            mReversedBeta2.setBackend(mParameters[0]->getImpl()->backend());
+        }
+		
+        Tensor alpha{std::vector<std::size_t>({1})};
+        alpha.setBackend(mParameters[0]->getImpl()->backend());
+        alpha.set<float>(0, learningRate() * std::sqrt(1.0f - std::pow(mBeta2.get<float>(0), mLRScheduler.step() + 1))
+                                           / (1.0f - std::pow(mBeta1.get<float>(0), mLRScheduler.step() + 1)));
+
+        Tensor epsilon{std::vector<std::size_t>({1})};
+        epsilon.setBackend(mParameters[0]->getImpl()->backend());
+        epsilon.set<float>(0, mEpsilon.get<float>(0) * std::sqrt(1.0f - std::pow(mBeta2.get<float>(0), mLRScheduler.step() + 1)));
+		
+        if (mLRScheduler.step() == 0) {
+            for (std::size_t i = 0; i < mParameters.size(); ++i) {
+                mMomentum1[i].setBackend(mParameters[i]->getImpl()->backend());
+                mMomentum1[i].setDataType(mParameters[i]->grad()->dataType());
+                mMomentum1[i].zeros();
+                mMomentum2[i].setBackend(mParameters[i]->getImpl()->backend());
+                mMomentum2[i].setDataType(mParameters[i]->grad()->dataType());
+                mMomentum2[i].zeros();
+            }
+        }
+		
+        for (std::size_t i = 0; i < mParameters.size(); ++i) {
+            mMomentum1[i] = mBeta1 * mMomentum1[i] + mReversedBeta1 * (*mParameters[i]->grad());
+            mMomentum2[i] = mBeta2 * mMomentum2[i] + mReversedBeta2 * (*mParameters[i]->grad()) * (*mParameters[i]->grad());
+            *mParameters[i] = *mParameters[i] - alpha * mMomentum1[i] / (mMomentum2[i].sqrt() +  epsilon);
+        }
+        
+        mLRScheduler.update();
+    }
+
+    void setParameters(const std::vector<std::shared_ptr<Tensor>>& parameters) override final {
+        Optimizer::setParameters(parameters);
+        mMomentum1 = std::vector<Tensor>(parameters.size());
+        mMomentum2 = std::vector<Tensor>(parameters.size());
+        for (std::size_t i = 0; i < parameters.size(); ++i) {
+            mMomentum1[i] = Tensor(parameters[i]->dims());
+            mMomentum2[i] = Tensor(parameters[i]->dims());
+        }
+    }
+};
+
+} // namespace Aidge
+
+
+namespace {
+template <>
+const char *const EnumStrings<Aidge::AdamAttr>::data[] = {
+    "Beta1",
+    "Beta2",
+    "Epsilon"
+};
+}
+#endif // AIDGE_CORE_OPTIMIZER_ADAM_H_