Add proximal adagrad optimizer (#5128)

paddle/operators/proximal_adagrad_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/proximal_adagrad_op.h"
+
+namespace paddle {
+namespace operators {
+
+class ProximalAdagradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Param"),
+                   "Input(Param) of ProximalAdagradOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Moment"),
+                   "Input(Moment) of ProximalAdagradOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Grad"),
+                   "Input(Grad) of ProximalAdagradOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasInput("LearningRate"),
+        "Input(LearningRate) of ProximalAdagradOp should not be null.");
+
+    PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
+                   "Output(ParamOut) of ProximalAdagradOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("MomentOut"),
+        "Output(MomentOut) of ProximalAdagradOp should not be null.");
+
+    auto param_dim = ctx->GetInputDim("Param");
+    PADDLE_ENFORCE_EQ(
+        param_dim, ctx->GetInputDim("Grad"),
+        "Param and Grad of ProximalAdagrad Op must have same dimension.");
+
+    PADDLE_ENFORCE_EQ(
+        param_dim, ctx->GetInputDim("Moment"),
+        "Param and Moment of ProximalAdagrad Op must have same dimension.");
+
+    auto lr_dim = ctx->GetInputDim("LearningRate");
+    PADDLE_ENFORCE_EQ(framework::product(lr_dim), 1,
+                      "Learning Rate should be a scalar.");
+
+    ctx->SetOutputDim("ParamOut", param_dim);
+    ctx->SetOutputDim("MomentOut", param_dim);
+  }
+};
+
+class ProximalAdagradOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ProximalAdagradOpMaker(framework::OpProto *proto,
+                         framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Param",
+             "(Tensor, default Tensor<float>) "
+             "Input parameter that has to be updated.");
+    AddInput("Moment",
+             "(Tensor, default Tensor<float>) "
+             "Moment parameter that has to be updated.");
+    AddInput("Grad",
+             "(Tensor, default Tensor<float>) "
+             "Input gradient of the parameter.");
+    AddInput("LearningRate",
+             "(Tensor, default Tensor<float>) "
+             "The learning rate should be a tensor of size 1.");
+
+    AddOutput("ParamOut", "(Tensor) Output updated parameter value.");
+    AddOutput("MomentOut", "(Tensor) Output updated moment value.");
+
+    AddAttr<float>("l1",
+                   "(float, default 0.0) "
+                   "L1 regularization strength.")
+        .SetDefault(0.0f);
+    AddAttr<float>("l2",
+                   "(float, default 0.0)"
+                   "L2 regularization strength.")
+        .SetDefault(0.0f);
+    AddComment(R"DOC(
+
+Optimizer that implements the proximal adagrad algorithm.
+
+moment = moment + grad * grad
+prox_param = param - learning_rate * grad * (1 / sqrt(moment))
+param = sign(prox_param) / (1 + learning_rate * l2) *
+        max { |prox_param| - learning_rate * l1 , 0 }
+
+The paper that proposed Proximal GD: 
+(http://papers.nips.cc/paper/3793-efficient-learning-using-forward-backward-splitting.pdf)
+Here, we use the adagrad learning rate as specified here: 
+(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
+)DOC");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(proximal_adagrad, ops::ProximalAdagradOp,
+                             ops::ProximalAdagradOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    proximal_adagrad,
+    ops::ProximalAdagradOpKernel<paddle::platform::CPUPlace, float>);

paddle/operators/proximal_adagrad_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+You may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed
+under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/proximal_adagrad_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(
+    proximal_adagrad,
+    ops::ProximalAdagradOpKernel<paddle::platform::GPUPlace, float>);

paddle/operators/proximal_adagrad_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+
+template <typename Place, typename T>
+class ProximalAdagradOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* param_out = ctx.Output<Tensor>("ParamOut");
+    auto* moment_out = ctx.Output<Tensor>("MomentOut");
+
+    param_out->mutable_data<T>(ctx.GetPlace());
+    moment_out->mutable_data<T>(ctx.GetPlace());
+
+    auto l1 = static_cast<T>(ctx.Attr<float>("l1"));
+    auto l2 = static_cast<T>(ctx.Attr<float>("l2"));
+
+    auto grad = ctx.Input<Tensor>("Grad");
+    auto p = EigenVector<T>::Flatten(*ctx.Input<Tensor>("Param"));
+    auto m = EigenVector<T>::Flatten(*ctx.Input<Tensor>("Moment"));
+    auto g = EigenVector<T>::Flatten(*grad);
+    auto lr = EigenVector<T>::Flatten(*ctx.Input<Tensor>("LearningRate"));
+
+    auto p_out = EigenVector<T>::Flatten(*param_out);
+    auto m_out = EigenVector<T>::Flatten(*moment_out);
+    auto place = ctx.GetEigenDevice<Place>();
+
+    Eigen::DSizes<int, 1> grad_dsize(grad->numel());
+
+    m_out.device(place) = m + g * g;
+    auto prox_param = p - lr.broadcast(grad_dsize) * g / m_out.sqrt();
+    if (l1 > static_cast<T>(0)) {
+      p_out.device(place) =
+          prox_param.sign() *
+          (((prox_param.abs() - (lr * l1).broadcast(grad_dsize))
+                .cwiseMax(static_cast<T>(0.0))) /
+           (static_cast<T>(1.0) + (lr * l2).broadcast(grad_dsize)));
+    } else {
+      p_out.device(place) =
+          prox_param / (static_cast<T>(1.0) + (lr * l2).broadcast(grad_dsize));
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

python/paddle/v2/framework/tests/test_proximal_adagrad_op.py

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestProximalAdagradOp(OpTest):
+    def setUp(self):
+        self.op_type = "proximal_adagrad"
+        w = np.random.random((102, 105)).astype("float32")
+        m = np.random.random((102, 105)).astype("float32")
+        g = np.random.random((102, 105)).astype("float32")
+        lr = np.array([0.1]).astype("float32")
+        l1 = 0.1
+        l2 = 0.2
+
+        self.inputs = {'Param': w, 'Grad': g, 'Moment': m, 'LearningRate': lr}
+        self.attrs = {'l1': l1, 'l2': l2}
+        param_out = 0.0
+
+        moment_out = m + g * g
+        prox_param = w - lr * g / np.sqrt(moment_out)
+        if l1 > 0.0:
+            x = np.abs(prox_param) - lr * l1
+            x[x < 0] = 0
+            param_out = np.sign(prox_param) * (x / (1.0 + lr * l2))
+        else:
+            param_out = prox_param / (1.0 + lr * l2)
+
+        self.outputs = {'ParamOut': param_out, 'MomentOut': moment_out}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == "__main__":
+    unittest.main()