Merge pull request #4571 from sidgoyal78/momentum_op

Add momentum operator

paddle/operators/momentum_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/momentum_op.h"
+
+namespace paddle {
+namespace operators {
+
+class MomentumOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContextBase *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Param"),
+                   "Input(param) of Momentum should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Grad"),
+                   "Input(grad) of Momentum should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Velocity"),
+                   "Input(velocity) of Momentum should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("LearningRate"),
+                   "Input(LearningRate) of Momentum should not be null.");
+
+    PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
+                   "Output(ParamOut) of Momentum should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("VelocityOut"),
+                   "Output(VelocityOut) of Momentum should not be null.");
+
+    auto param_dim = ctx->GetInputDim("Param");
+    PADDLE_ENFORCE_EQ(
+        param_dim, ctx->GetInputDim("Grad"),
+        "Param and Grad input of MomentumOp should have the same dimension.");
+    PADDLE_ENFORCE_EQ(
+        param_dim, ctx->GetInputDim("Velocity"),
+        "Param and Velocity of MomentumOp should have the same dimension.");
+    PADDLE_ENFORCE_EQ(framework::product(ctx->GetInputDim("LearningRate")), 1,
+                      "Learning_rate should be a scalar");
+
+    ctx->SetOutputDim("ParamOut", param_dim);
+    ctx->SetOutputDim("VelocityOut", param_dim);
+  }
+};
+
+class MomentumOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  MomentumOpMaker(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("Grad",
+             "(Tensor, default Tensor<float>) "
+             "Input gradient of the parameter");
+    AddInput("Velocity",
+             "(Tensor, default Tensor<float>) "
+             "Input velocity (corresponding to the parameter) "
+             "that has to be updated");
+    AddInput("LearningRate",
+             "(Tensor, default Tensor<float>) "
+             "Input learning rate");
+
+    AddOutput("ParamOut", "(Tensor) Output updated parameter");
+    AddOutput("VelocityOut", "(Tensor) Output updated velocity");
+
+    AddAttr<float>("mu", "(float) Momentum coefficient");
+    AddComment(R"DOC(
+
+Momentum Algorithm (momentum).
+
+velocity = mu * velocity + gradient
+param = param - learning_rate * velocity
+
+)DOC");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(momentum, ops::MomentumOp, ops::MomentumOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    momentum, ops::MomentumOpKernel<paddle::platform::CPUPlace, float>);

paddle/operators/momentum_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/momentum_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(
+    momentum, ops::MomentumOpKernel<paddle::platform::GPUPlace, float>);

paddle/operators/momentum_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 {
+
+template <typename Place, typename T>
+class MomentumOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto param_out = ctx.Output<framework::Tensor>("ParamOut");
+    auto velocity_out = ctx.Output<framework::Tensor>("VelocityOut");
+    auto param = ctx.Input<framework::Tensor>("Param");
+    auto velocity = ctx.Input<framework::Tensor>("Velocity");
+    auto grad = ctx.Input<framework::Tensor>("Grad");
+    auto learning_rate = ctx.Input<framework::Tensor>("LearningRate");
+
+    param_out->mutable_data<T>(ctx.GetPlace());
+    velocity_out->mutable_data<T>(ctx.GetPlace());
+
+    float mu = ctx.Attr<float>("mu");
+
+    auto p_out = framework::EigenVector<T>::Flatten(*param_out);
+    auto v_out = framework::EigenVector<T>::Flatten(*velocity_out);
+
+    auto p = framework::EigenVector<T>::Flatten(*param);
+    auto v = framework::EigenVector<T>::Flatten(*velocity);
+    auto g = framework::EigenVector<T>::Flatten(*grad);
+    auto lr = framework::EigenVector<T>::Flatten(*learning_rate);
+
+    auto place = ctx.GetEigenDevice<Place>();
+
+    Eigen::DSizes<int, 1> grad_dsize(grad->numel());
+    v_out.device(place) = v * mu + g;
+    p_out.device(place) = p - lr.broadcast(grad_dsize) * v_out;
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

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

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestMomentumOp(OpTest):
+    def setUp(self):
+        self.op_type = "momentum"
+
+        param = np.random.random((123, 321)).astype("float32")
+        grad = np.random.random((123, 321)).astype("float32")
+        velocity = np.zeros((123, 321)).astype("float32")
+        learning_rate = np.array([0.001]).astype("float32")
+        mu = 0.0001
+
+        self.inputs = {
+            'Param': param,
+            'Grad': grad,
+            'Velocity': velocity,
+            'LearningRate': learning_rate
+        }
+
+        self.attrs = {'mu': mu}
+
+        velocity_out = mu * velocity + grad
+        param_out = param - learning_rate * velocity_out
+
+        self.outputs = {'ParamOut': param_out, 'VelocityOut': velocity_out}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == "__main__":
+    unittest.main()