From 229d4bbba1b212ff6fae79e5d88e40def3a8da1e Mon Sep 17 00:00:00 2001 From: Zeng Jinle <32832641+sneaxiy@users.noreply.github.com> Date: Tue, 16 Oct 2018 23:50:12 +0800 Subject: [PATCH] cherry-pick sparse rmsprop to release/1.0.0 (#13907) * test=release/1.0.0 * fix sparse rmsprop * test=develop * add check for opt op * test=release/1.0.0 --- paddle/fluid/operators/adadelta_op.cc | 12 + paddle/fluid/operators/adadelta_op.h | 11 + paddle/fluid/operators/adagrad_op.h | 33 ++- paddle/fluid/operators/adam_op.h | 25 +- paddle/fluid/operators/adamax_op.cc | 10 + paddle/fluid/operators/adamax_op.h | 11 + paddle/fluid/operators/decayed_adagrad_op.cc | 10 + paddle/fluid/operators/decayed_adagrad_op.h | 11 + paddle/fluid/operators/ftrl_op.cc | 10 + paddle/fluid/operators/ftrl_op.h | 11 + paddle/fluid/operators/math/algorithm.h | 44 +++ paddle/fluid/operators/momentum_op.cc | 5 + paddle/fluid/operators/momentum_op.cu | 11 + paddle/fluid/operators/momentum_op.h | 6 + paddle/fluid/operators/rmsprop_op.cc | 5 + paddle/fluid/operators/rmsprop_op.h | 275 +++++++++++++++--- paddle/fluid/operators/sgd_op.cc | 29 +- paddle/fluid/operators/sgd_op.cu | 6 + python/paddle/fluid/optimizer.py | 12 + .../fluid/tests/unittests/test_rmsprop_op.py | 231 +++++++++------ 20 files changed, 593 insertions(+), 175 deletions(-) create mode 100644 paddle/fluid/operators/math/algorithm.h diff --git a/paddle/fluid/operators/adadelta_op.cc b/paddle/fluid/operators/adadelta_op.cc index d1970515f5..89a7a49e0f 100644 --- a/paddle/fluid/operators/adadelta_op.cc +++ b/paddle/fluid/operators/adadelta_op.cc @@ -18,6 +18,7 @@ namespace paddle { namespace operators { using Tensor = framework::Tensor; + class AdadeltaOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -31,6 +32,16 @@ class AdadeltaOp : public framework::OperatorWithKernel { "Input(AvgSquaredGrad) of AdadeltaOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("AvgSquaredUpdate"), "Input(AvgSquaredUpdate) of AdadeltaOp should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Grad").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(ParamOut) of AdadeltaOp should not be null."); @@ -56,6 +67,7 @@ class AdadeltaOp : public framework::OperatorWithKernel { ctx->SetOutputDim("AvgSquaredGradOut", param_dim); ctx->SetOutputDim("AvgSquaredUpdateOut", param_dim); } + framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext &ctx) const override { auto input_data_type = diff --git a/paddle/fluid/operators/adadelta_op.h b/paddle/fluid/operators/adadelta_op.h index 822458daf6..6c616aa03d 100644 --- a/paddle/fluid/operators/adadelta_op.h +++ b/paddle/fluid/operators/adadelta_op.h @@ -23,6 +23,17 @@ template class AdadeltaOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + const auto* grad_var = ctx.InputVar("Grad"); + PADDLE_ENFORCE(grad_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Grad").front(), grad_var->Type().name()); + auto param_out_tensor = ctx.Output("ParamOut"); auto avg_squared_grad_out_tensor = ctx.Output("AvgSquaredGradOut"); diff --git a/paddle/fluid/operators/adagrad_op.h b/paddle/fluid/operators/adagrad_op.h index df520fcc89..0a16ce00f7 100644 --- a/paddle/fluid/operators/adagrad_op.h +++ b/paddle/fluid/operators/adagrad_op.h @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once + #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" @@ -21,25 +22,31 @@ namespace operators { template struct SparseAdagradFunctor { - void operator()(const DeviceContext& context, - const framework::SelectedRows& grad, - const framework::Tensor& learning_rate, T epsilon, - framework::Tensor* moment, framework::Tensor* param); + void operator()(const DeviceContext &context, + const framework::SelectedRows &grad, + const framework::Tensor &learning_rate, T epsilon, + framework::Tensor *moment, framework::Tensor *param); }; template class AdagradOpKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto* param_out_tensor = ctx.Output("ParamOut"); - auto* moment_out_tensor = ctx.Output("MomentOut"); + void Compute(const framework::ExecutionContext &ctx) const override { + const auto *param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + + auto *param_out_tensor = ctx.Output("ParamOut"); + auto *moment_out_tensor = ctx.Output("MomentOut"); param_out_tensor->mutable_data(ctx.GetPlace()); moment_out_tensor->mutable_data(ctx.GetPlace()); T epsilon = static_cast(ctx.Attr("epsilon")); - auto* grad_var = ctx.InputVar("Grad"); + auto *grad_var = ctx.InputVar("Grad"); if (grad_var->IsType()) { auto param = framework::EigenVector::Flatten( *ctx.Input("Param")); @@ -47,16 +54,16 @@ class AdagradOpKernel : public framework::OpKernel { *ctx.Input("Grad")); auto moment = framework::EigenVector::Flatten( *ctx.Input("Moment")); - auto* learning_rate = ctx.Input("LearningRate"); + auto *learning_rate = ctx.Input("LearningRate"); auto param_out = framework::EigenVector::Flatten(*param_out_tensor); auto moment_out = framework::EigenVector::Flatten(*moment_out_tensor); - auto* place = ctx.template device_context().eigen_device(); + auto *place = ctx.template device_context().eigen_device(); moment_out.device(*place) = moment + grad * grad; Eigen::DSizes m_dsize(moment_out_tensor->numel()); if (platform::is_cpu_place(ctx.GetPlace())) { - auto* lr = learning_rate->data(); + auto *lr = learning_rate->data(); param_out.device(*place) = param - lr[0] * grad / (moment_out.sqrt() + epsilon); } else { @@ -66,10 +73,10 @@ class AdagradOpKernel : public framework::OpKernel { lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon); } } else if (grad_var->IsType()) { - auto* param_tensor = ctx.Input("Param"); + auto *param_tensor = ctx.Input("Param"); PADDLE_ENFORCE_EQ(param_tensor, param_out_tensor); - auto* moment_tensor = ctx.Input("Moment"); + auto *moment_tensor = ctx.Input("Moment"); PADDLE_ENFORCE_EQ(moment_tensor, moment_out_tensor); SparseAdagradFunctor functor; diff --git a/paddle/fluid/operators/adam_op.h b/paddle/fluid/operators/adam_op.h index 4cb1f3a80e..3455d1ee54 100644 --- a/paddle/fluid/operators/adam_op.h +++ b/paddle/fluid/operators/adam_op.h @@ -18,6 +18,7 @@ limitations under the License. */ #include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detail/safe_ref.h" +#include "paddle/fluid/operators/math/algorithm.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" #include "paddle/fluid/platform/for_range.h" @@ -199,23 +200,9 @@ struct SparseAdamFunctor { row_numel_(row_numel), row_count_(row_count) {} - inline HOSTDEVICE int64_t BinarySearchInRows(int64_t row) const { - int64_t beg = 0, end = row_count_ - 1; - while (beg <= end) { - auto mid = ((beg + end) >> 1); - if (rows_[mid] == row) - return mid; - else if (rows_[mid] < row) - beg = mid + 1; - else - end = mid - 1; - } - return -1; - } - inline HOSTDEVICE void operator()(size_t i) const { - int64_t row = i / row_numel_; - auto row_idx = BinarySearchInRows(row); + auto row_idx = + math::BinarySearch(rows_, row_count_, i / row_numel_); T g = row_idx >= 0 ? grad_[row_idx * row_numel_ + i % row_numel_] : 0; // The following code is the same as dense @@ -244,6 +231,12 @@ template class AdamOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + using paddle::framework::LoDTensor; using paddle::operators::detail::Ref; diff --git a/paddle/fluid/operators/adamax_op.cc b/paddle/fluid/operators/adamax_op.cc index 32062574bc..d4aa4d338a 100644 --- a/paddle/fluid/operators/adamax_op.cc +++ b/paddle/fluid/operators/adamax_op.cc @@ -35,6 +35,16 @@ class AdamaxOp : public framework::OperatorWithKernel { "Input(LearningRate) of AdamaxOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("Beta1Pow"), "Input(Beta1Pow) of AdamaxOp should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Grad").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(ParamOut) of AdamaxOp should not be null."); diff --git a/paddle/fluid/operators/adamax_op.h b/paddle/fluid/operators/adamax_op.h index de644676fd..7137fbd965 100644 --- a/paddle/fluid/operators/adamax_op.h +++ b/paddle/fluid/operators/adamax_op.h @@ -23,6 +23,17 @@ template class AdamaxOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + const auto* grad_var = ctx.InputVar("Grad"); + PADDLE_ENFORCE(grad_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Grad").front(), grad_var->Type().name()); + auto param_out_tensor = ctx.Output("ParamOut"); auto moment_out_tensor = ctx.Output("MomentOut"); auto inf_norm_out_tensor = ctx.Output("InfNormOut"); diff --git a/paddle/fluid/operators/decayed_adagrad_op.cc b/paddle/fluid/operators/decayed_adagrad_op.cc index c0f2b49a04..d73ae9e272 100644 --- a/paddle/fluid/operators/decayed_adagrad_op.cc +++ b/paddle/fluid/operators/decayed_adagrad_op.cc @@ -32,6 +32,16 @@ class DecayedAdagradOp : public framework::OperatorWithKernel { PADDLE_ENFORCE( ctx->HasInput("LearningRate"), "Input(LearningRate) of DecayedAdagradOp should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Grad").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(ParamOut) of DecayedAdagradOp should not be null."); diff --git a/paddle/fluid/operators/decayed_adagrad_op.h b/paddle/fluid/operators/decayed_adagrad_op.h index a46af078e0..5df43d33ef 100644 --- a/paddle/fluid/operators/decayed_adagrad_op.h +++ b/paddle/fluid/operators/decayed_adagrad_op.h @@ -23,6 +23,17 @@ template class DecayedAdagradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + const auto* grad_var = ctx.InputVar("Grad"); + PADDLE_ENFORCE(grad_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Grad").front(), grad_var->Type().name()); + auto param_out_tensor = ctx.Output("ParamOut"); auto moment_out_tensor = ctx.Output("MomentOut"); diff --git a/paddle/fluid/operators/ftrl_op.cc b/paddle/fluid/operators/ftrl_op.cc index 70ba25c213..b77e12d650 100644 --- a/paddle/fluid/operators/ftrl_op.cc +++ b/paddle/fluid/operators/ftrl_op.cc @@ -34,6 +34,16 @@ class FTRLOp : public framework::OperatorWithKernel { "Input(Grad) of FTRL should not be null."); PADDLE_ENFORCE(ctx->HasInput("LearningRate"), "Input(LearningRate) of FTRL should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Grad").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(ParamOut) of FTRL should not be null."); diff --git a/paddle/fluid/operators/ftrl_op.h b/paddle/fluid/operators/ftrl_op.h index 6f821e7e99..8f812c9a03 100644 --- a/paddle/fluid/operators/ftrl_op.h +++ b/paddle/fluid/operators/ftrl_op.h @@ -28,6 +28,17 @@ template class FTRLOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + const auto* grad_var = ctx.InputVar("Grad"); + PADDLE_ENFORCE(grad_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Grad").front(), grad_var->Type().name()); + auto* param_out = ctx.Output("ParamOut"); auto* sq_accum_out = ctx.Output("SquaredAccumOut"); auto* lin_accum_out = ctx.Output("LinearAccumOut"); diff --git a/paddle/fluid/operators/math/algorithm.h b/paddle/fluid/operators/math/algorithm.h new file mode 100644 index 0000000000..262469beea --- /dev/null +++ b/paddle/fluid/operators/math/algorithm.h @@ -0,0 +1,44 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// 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 +#include // for int64_t +#include + +#include "paddle/fluid/platform/hostdevice.h" + +namespace paddle { +namespace operators { +namespace math { + +template +HOSTDEVICE inline int64_t BinarySearch(const T *x, int64_t num, const T &val) { + int64_t beg = 0, end = num - 1; + while (beg <= end) { + auto mid = ((beg + end) >> 1); + if (x[mid] == val) + return mid; + else if (x[mid] < val) + beg = mid + 1; + else + end = mid - 1; + } + return -1; +} + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/momentum_op.cc b/paddle/fluid/operators/momentum_op.cc index 5f43c58108..c8079a99fb 100644 --- a/paddle/fluid/operators/momentum_op.cc +++ b/paddle/fluid/operators/momentum_op.cc @@ -33,6 +33,11 @@ class MomentumOp : public framework::OperatorWithKernel { "Input(velocity) of Momentum should not be null."); PADDLE_ENFORCE(ctx->HasInput("LearningRate"), "Input(LearningRate) of Momentum should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(ParamOut) of Momentum should not be null."); diff --git a/paddle/fluid/operators/momentum_op.cu b/paddle/fluid/operators/momentum_op.cu index a3932db1f3..5dc920c709 100644 --- a/paddle/fluid/operators/momentum_op.cu +++ b/paddle/fluid/operators/momentum_op.cu @@ -46,6 +46,17 @@ template class MomentumOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + const auto* grad_var = ctx.InputVar("Grad"); + PADDLE_ENFORCE(grad_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Grad").front(), grad_var->Type().name()); + auto param_out = ctx.Output("ParamOut"); auto velocity_out = ctx.Output("VelocityOut"); auto param = ctx.Input("Param"); diff --git a/paddle/fluid/operators/momentum_op.h b/paddle/fluid/operators/momentum_op.h index 264726040f..40073d21b7 100644 --- a/paddle/fluid/operators/momentum_op.h +++ b/paddle/fluid/operators/momentum_op.h @@ -23,6 +23,12 @@ template class MomentumOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + auto param_out = ctx.Output("ParamOut"); auto velocity_out = ctx.Output("VelocityOut"); auto param = ctx.Input("Param"); diff --git a/paddle/fluid/operators/rmsprop_op.cc b/paddle/fluid/operators/rmsprop_op.cc index 2f773f222e..f06f87e61d 100644 --- a/paddle/fluid/operators/rmsprop_op.cc +++ b/paddle/fluid/operators/rmsprop_op.cc @@ -32,6 +32,11 @@ class RmspropOp : public framework::OperatorWithKernel { "Input(Grad) of RmspropOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("Moment"), "Input(Moment) of RmspropOp should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(param_out) of RmspropOp should not be null."); diff --git a/paddle/fluid/operators/rmsprop_op.h b/paddle/fluid/operators/rmsprop_op.h index 25ed32c5eb..2a1527a3d9 100644 --- a/paddle/fluid/operators/rmsprop_op.h +++ b/paddle/fluid/operators/rmsprop_op.h @@ -13,66 +13,259 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/algorithm.h" +#include "paddle/fluid/operators/math/selected_rows_functor.h" +#include "paddle/fluid/platform/for_range.h" namespace paddle { namespace operators { -using Tensor = framework::Tensor; template using EigenVector = framework::EigenVector; +template +struct DenseRmspropGradFunctor { + inline explicit DenseRmspropGradFunctor(const T *grad) : grad_(grad) {} + + HOSTDEVICE inline T operator()(int64_t idx) const { return grad_[idx]; } + + const T *grad_; +}; + +template +struct SparseRmspropGradFunctor { + inline SparseRmspropGradFunctor(const T *grad, const int64_t *rows, + int64_t row_numel, int64_t row_count) + : grad_(grad), + rows_(rows), + row_numel_(row_numel), + row_count_(row_count) {} + + HOSTDEVICE inline T operator()(int64_t idx) const { + auto row_idx = math::BinarySearch(rows_, row_count_, idx / row_numel_); + return row_idx >= 0 ? grad_[row_idx * row_numel_ + idx % row_numel_] : 0; + } + + const T *grad_; + const int64_t *rows_; + int64_t row_numel_; + int64_t row_count_; +}; + +template +struct UncenteredRmspropFunctor { + UncenteredRmspropFunctor(T *param, T *ms, T *mom, const T *lr, T rho, + T epsilon, T momentum, + const GradFunctor &grad_functor) + : param_(param), + ms_(ms), + mom_(mom), + lr_(lr), + rho_(rho), + epsilon_(epsilon), + momentum_(momentum), + grad_functor_(grad_functor) {} + + HOSTDEVICE inline void operator()(int64_t idx) const { + T g = grad_functor_(idx); + T ms_out = rho_ * ms_[idx] + (1 - rho_) * g * g; + T mom_out = momentum_ * mom_[idx] + lr_[0] * g / sqrt(ms_out + epsilon_); + param_[idx] -= mom_out; + ms_[idx] = ms_out; + mom_[idx] = mom_out; + } + + T *param_; + T *ms_; + T *mom_; + const T *lr_; + T rho_; + T epsilon_; + T momentum_; + GradFunctor grad_functor_; +}; + +template +struct CenteredRmspropFunctor { + CenteredRmspropFunctor(T *param, T *ms, T *mom, T *mean_grad, const T *lr, + T rho, T epsilon, T momentum, + const GradFunctor &grad_functor) + : param_(param), + ms_(ms), + mom_(mom), + mean_grad_(mean_grad), + lr_(lr), + rho_(rho), + epsilon_(epsilon), + momentum_(momentum), + grad_functor_(grad_functor) {} + + HOSTDEVICE inline void operator()(int64_t idx) const { + T g = grad_functor_(idx); + T ms_out = rho_ * ms_[idx] + (1 - rho_) * g * g; + T mg_out = rho_ * mean_grad_[idx] + (1 - rho_) * g; + T mom_out = momentum_ * mom_[idx] + + lr_[0] * g / sqrt(ms_out - mg_out * mg_out + epsilon_); + param_[idx] -= mom_out; + ms_[idx] = ms_out; + mom_[idx] = mom_out; + mean_grad_[idx] = mg_out; + } + + T *param_; + T *ms_; + T *mom_; + T *mean_grad_; + const T *lr_; + T rho_; + T epsilon_; + T momentum_; + GradFunctor grad_functor_; +}; + template class RmspropOpKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto* param_out = ctx.Output("ParamOut"); - auto* moment_out = ctx.Output("MomentOut"); - auto* mean_square_out = ctx.Output("MeanSquareOut"); + void Compute(const framework::ExecutionContext &ctx) const override { + using LoDTensor = framework::LoDTensor; + const auto *param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + auto *grad_var = ctx.InputVar("Grad"); + auto *param_out = ctx.Output("ParamOut"); + auto *moment_out = ctx.Output("MomentOut"); + auto *mean_square_out = ctx.Output("MeanSquareOut"); - auto grad = ctx.Input("Grad"); + auto epsilon = static_cast(ctx.Attr("epsilon")); + auto rho = static_cast(ctx.Attr("decay")); + auto momentum = static_cast(ctx.Attr("momentum")); + bool centered = ctx.Attr("centered"); - param_out->mutable_data(ctx.GetPlace()); - moment_out->mutable_data(ctx.GetPlace()); - mean_square_out->mutable_data(ctx.GetPlace()); + auto &p_tensor = *ctx.Input("Param"); + auto &ms_tensor = *ctx.Input("MeanSquare"); + auto &lr_tensor = *ctx.Input("LearningRate"); + auto &mom_tensor = *ctx.Input("Moment"); - float epsilon = ctx.Attr("epsilon"); - float rho = ctx.Attr("decay"); - float momentum = ctx.Attr("momentum"); - bool centered = ctx.Attr("centered"); + PADDLE_ENFORCE_EQ(&p_tensor, param_out, + "Param and ParamOut must be the same Tensor"); + PADDLE_ENFORCE_EQ(&mom_tensor, moment_out, + "Moment and MomentOut must be the same Tensor"); + PADDLE_ENFORCE_EQ(&ms_tensor, mean_square_out, + "MeanSquare and MeanSquareOut must be the same Tensor"); + + auto &dev_ctx = ctx.template device_context(); + size_t limit = static_cast(ms_tensor.numel()); + + if (grad_var->IsType()) { + auto &grad_tensor = grad_var->Get(); + + if (std::is_same::value) { + auto &place = + *ctx.template device_context().eigen_device(); + auto lr_value = lr_tensor.data()[0]; + + auto p = EigenVector::Flatten(p_tensor); + auto ms = EigenVector::Flatten(ms_tensor); + auto g = EigenVector::Flatten(grad_tensor); + auto mom = EigenVector::Flatten(mom_tensor); + + auto p_out = EigenVector::Flatten(*param_out); + auto mom_out = EigenVector::Flatten(*moment_out); + auto ms_out = EigenVector::Flatten(*mean_square_out); + + ms_out.device(place) = rho * ms + (1 - rho) * g * g; + if (centered) { + auto &mg_tensor = *ctx.Input("MeanGrad"); + auto mg = EigenVector::Flatten(mg_tensor); + auto *mean_grad_out = ctx.Output("MeanGradOut"); + PADDLE_ENFORCE(&mg_tensor, mean_grad_out, + "MeanGrad and MeanGradOut must be the same Tensor"); + auto mg_out = EigenVector::Flatten(*mean_grad_out); + + mg_out.device(place) = rho * mg + (1 - rho) * g; + mom_out.device(place) = + momentum * mom + + lr_value * g / (ms_out - mg_out.square() + epsilon).sqrt(); + } else { + mom_out.device(place) = + momentum * mom + lr_value * g / (ms_out + epsilon).sqrt(); + } + p_out.device(place) = p - mom_out; + } else { + DenseRmspropGradFunctor grad_func(grad_tensor.data()); + platform::ForRange for_range(dev_ctx, limit); + if (centered) { + auto &mg_tensor = *ctx.Input("MeanGrad"); + auto *mean_grad_out = ctx.Output("MeanGradOut"); + PADDLE_ENFORCE(&mg_tensor, mean_grad_out, + "MeanGrad and MeanGradOut must be the same Tensor"); + for_range(CenteredRmspropFunctor>( + param_out->mutable_data(ctx.GetPlace()), + mean_square_out->mutable_data(ctx.GetPlace()), + moment_out->mutable_data(ctx.GetPlace()), + mean_grad_out->mutable_data(ctx.GetPlace()), + lr_tensor.data(), rho, epsilon, momentum, grad_func)); + } else { + for_range(UncenteredRmspropFunctor>( + param_out->mutable_data(ctx.GetPlace()), + mean_square_out->mutable_data(ctx.GetPlace()), + moment_out->mutable_data(ctx.GetPlace()), lr_tensor.data(), + rho, epsilon, momentum, grad_func)); + } + } + } else if (grad_var->IsType()) { + auto &grad = grad_var->Get(); + auto *merged_grad = const_cast(ctx.scope()) + .Var() + ->GetMutable(); + + math::scatter::MergeAdd merge_func; + merge_func(dev_ctx, grad, merged_grad); + + platform::ForRange for_range(dev_ctx, limit); + const int64_t *rows; +#ifdef PADDLE_WITH_CUDA + if (platform::is_gpu_place(ctx.GetPlace())) { + rows = merged_grad->rows().CUDAData(ctx.GetPlace()); + } else { +#endif + rows = merged_grad->rows().data(); +#ifdef PADDLE_WITH_CUDA + } +#endif + auto &merged_tensor = merged_grad->value(); + int64_t row_count = merged_grad->rows().size(); + int64_t row_numel = merged_tensor.numel() / row_count; + SparseRmspropGradFunctor grad_func(merged_tensor.data(), rows, + row_numel, row_count); - auto p = EigenVector::Flatten(*ctx.Input("Param")); - auto ms = EigenVector::Flatten(*ctx.Input("MeanSquare")); - auto lr = EigenVector::Flatten(*ctx.Input("LearningRate")); - auto g = EigenVector::Flatten(*grad); - auto mom = EigenVector::Flatten(*ctx.Input("Moment")); - - auto p_out = EigenVector::Flatten(*param_out); - auto mom_out = EigenVector::Flatten(*moment_out); - auto ms_out = EigenVector::Flatten(*mean_square_out); - auto& place = *ctx.template device_context().eigen_device(); - - Eigen::DSizes grad_dsize(static_cast(grad->numel())); - - ms_out.device(place) = rho * ms + (1 - rho) * g * g; - if (centered) { - auto mg = EigenVector::Flatten(*ctx.Input("MeanGrad")); - auto* mean_grad_out = ctx.Output("MeanGradOut"); - mean_grad_out->mutable_data(ctx.GetPlace()); - auto mg_out = EigenVector::Flatten(*mean_grad_out); - - mg_out.device(place) = rho * mg + (1 - rho) * g; - mom_out.device(place) = momentum * mom + - lr.broadcast(grad_dsize) * g / - (ms_out - mg_out.square() + epsilon).sqrt(); + if (centered) { + auto &mg_tensor = *ctx.Input("MeanGrad"); + auto *mean_grad_out = ctx.Output("MeanGradOut"); + PADDLE_ENFORCE(&mg_tensor, mean_grad_out, + "MeanGrad and MeanGradOut must be the same Tensor"); + for_range(CenteredRmspropFunctor>( + param_out->mutable_data(ctx.GetPlace()), + mean_square_out->mutable_data(ctx.GetPlace()), + moment_out->mutable_data(ctx.GetPlace()), + mean_grad_out->mutable_data(ctx.GetPlace()), lr_tensor.data(), + rho, epsilon, momentum, grad_func)); + } else { + for_range(UncenteredRmspropFunctor>( + param_out->mutable_data(ctx.GetPlace()), + mean_square_out->mutable_data(ctx.GetPlace()), + moment_out->mutable_data(ctx.GetPlace()), lr_tensor.data(), + rho, epsilon, momentum, grad_func)); + } } else { - mom_out.device(place) = - momentum * mom + - lr.broadcast(grad_dsize) * g / (ms_out + epsilon).sqrt(); + PADDLE_THROW("RMSProp only supports LoDTensor or SelectedRows gradient"); } - p_out.device(place) = p - mom_out; } }; diff --git a/paddle/fluid/operators/sgd_op.cc b/paddle/fluid/operators/sgd_op.cc index fef230e42d..411a126bc8 100644 --- a/paddle/fluid/operators/sgd_op.cc +++ b/paddle/fluid/operators/sgd_op.cc @@ -21,7 +21,7 @@ class SGDOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - void InferShape(framework::InferShapeContext* ctx) const override { + void InferShape(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasInput("Param"), "Input(Param) of SGDOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("Grad"), @@ -42,7 +42,7 @@ class SGDOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext& ctx) const override { + const framework::ExecutionContext &ctx) const override { auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("Param")); return framework::OpKernelType(data_type, ctx.device_context()); } @@ -50,17 +50,20 @@ class SGDOp : public framework::OperatorWithKernel { class SGDOpInferVarType : public framework::VarTypeInference { public: - void operator()(const framework::OpDesc& op_desc, - framework::BlockDesc* block) const override { - auto input_var = op_desc.Input("Param")[0]; - for (auto& out_var : op_desc.Output("ParamOut")) { - if (block->FindRecursiveOrCreateVar(input_var).GetType() == - framework::proto::VarType::SELECTED_ROWS) { - block->FindRecursiveOrCreateVar(out_var).SetType( - framework::proto::VarType::SELECTED_ROWS); - } else { - block->FindRecursiveOrCreateVar(out_var).SetType( - framework::proto::VarType::LOD_TENSOR); + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + auto input_var_n = op_desc.Input("Param")[0]; + auto in_var_type = block->FindRecursiveOrCreateVar(input_var_n).GetType(); + PADDLE_ENFORCE(in_var_type == framework::proto::VarType::SELECTED_ROWS || + in_var_type == framework::proto::VarType::LOD_TENSOR, + "The input Var's type should be LoDtensor or SelectedRows," + " but the received var(%s)'s type is %s", + input_var_n, in_var_type); + + for (auto &out_var_n : op_desc.Output("ParamOut")) { + auto &out_var = block->FindRecursiveOrCreateVar(out_var_n); + if (out_var.GetType() != in_var_type) { + out_var.SetType(in_var_type); } } } diff --git a/paddle/fluid/operators/sgd_op.cu b/paddle/fluid/operators/sgd_op.cu index 4722be7a66..5ddbac4c81 100644 --- a/paddle/fluid/operators/sgd_op.cu +++ b/paddle/fluid/operators/sgd_op.cu @@ -57,6 +57,12 @@ template class SGDOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + auto* param = ctx.Input("Param"); auto* param_out = ctx.Output("ParamOut"); auto* learning_rate = ctx.Input("LearningRate"); diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py index 1b9571f6d3..73e3a43af4 100644 --- a/python/paddle/fluid/optimizer.py +++ b/python/paddle/fluid/optimizer.py @@ -659,6 +659,9 @@ class AdamaxOptimizer(Optimizer): optimizer = fluid.optimizer.Adamax(learning_rate=0.2) optimizer.minimize(cost) + + Notes: + Currently, AdamaxOptimizer doesn't support sparse gradient. """ _moment_acc_str = "moment" _inf_norm_acc_str = "inf_norm" @@ -778,6 +781,9 @@ class DecayedAdagradOptimizer(Optimizer): optimizer = fluid.optimizer.DecayedAdagrad(learning_rate=0.2) optimizer.minimize(cost) + + Notes: + Currently, DecayedAdagradOptimizer doesn't support sparse gradient. """ _moment_acc_str = "moment" @@ -858,6 +864,9 @@ class AdadeltaOptimizer(Optimizer): optimizer = fluid.optimizer.Adadelta( learning_rate=0.0003, epsilon=1.0e-6, rho=0.95) _, params_grads = optimizer.minimize(cost) + + Notes: + Currently, AdadeltaOptimizer doesn't support sparse gradient. """ _avg_squared_grad_acc_str = "_avg_squared_grad" @@ -1126,6 +1135,9 @@ class FtrlOptimizer(Optimizer): optimizer = fluid.optimizer.Ftrl(0.0001) _, params_grads = optimizer.minimize(cost) + + Notes: + Currently, FtrlOptimizer doesn't support sparse gradient. """ _squared_acc_str = "squared" diff --git a/python/paddle/fluid/tests/unittests/test_rmsprop_op.py b/python/paddle/fluid/tests/unittests/test_rmsprop_op.py index 70848e4e22..eb12bc7417 100644 --- a/python/paddle/fluid/tests/unittests/test_rmsprop_op.py +++ b/python/paddle/fluid/tests/unittests/test_rmsprop_op.py @@ -19,33 +19,76 @@ import unittest import numpy as np import paddle.fluid.core as core from paddle.fluid.op import Operator +import paddle.fluid as fluid + + +def create_selected_rows_and_tensor(scope, place, height, row_num, + embedding_size): + sr = scope.var("@selected_rows@").get_selected_rows() + tensor = scope.var("grad").get_tensor() + + rows = np.random.random_integers( + low=0, high=height - 1, size=[row_num, ]).astype('int64') + sr_val = np.random.random(size=[row_num, embedding_size]).astype('float32') + + sr.set_height(height) + sr.set_rows(rows) + sr.get_tensor().set(sr_val, place) + + tensor_val = np.zeros(shape=[height, embedding_size], dtype='float32') + for i in range(row_num): + row = rows[i] + tensor_val[row, :] = tensor_val[row, :] + sr_val[i, :] + + tensor.set(tensor_val, place) + return tensor_val, sr_val class TestBase(unittest.TestCase): - def setup(self, centered, epsilon=1e-6): + def setup(self, + place, + is_sparse, + centered, + size, + row_num=None, + epsilon=1e-6): np.random.seed(5) # fix seed + self.scope = fluid.global_scope() + self.place = place + self.param_name = "param" - self.param = np.random.random((123, 321)).astype("float32") + self.param = np.random.random(size).astype("float32") self.mean_square_name = "mean_square" - self.mean_square = np.random.random((123, 321)).astype("float32") + self.mean_square = np.random.uniform( + low=1, high=2, size=size).astype("float32") self.mean_grad_name = "mean_grad" - self.mean_grad = np.random.random((123, 321)).astype("float32") + self.mean_grad = np.random.random(size).astype("float32") self.lr_name = "lr" self.learning_rate = np.array([0.01]).astype("float32") self.grad_name = "grad" - self.grad = np.random.random((123, 321)).astype("float32") + + self.is_sparse = is_sparse + if self.is_sparse: + self.grad_sr_name = "@selected_rows@" + self.grad, self.grad_sr = create_selected_rows_and_tensor( + self.scope, place, size[0], row_num, size[1]) + else: + self.grad = np.random.random(size).astype("float32") + grad_tensor = self.scope.var(self.grad_name).get_tensor() + grad_tensor.set(self.grad, place) self.moment_name = "moment" - self.moment = np.zeros((123, 321)).astype("float32") + self.moment = np.random.uniform( + low=0, high=1, size=size).astype("float32") self.epsilon = epsilon self.decay = 0.9 - self.momentum = 0.0 + self.momentum = 0.1 self.centered = centered self.ms_out = self.decay * self.mean_square + (1 - self.decay @@ -61,118 +104,122 @@ class TestBase(unittest.TestCase): self.param_out = self.param - self.moment_out - def check(self, - actual_t, - expect_t, - place, - out_name, - atol=1e-5, - equal_nan=False): - self.assertTrue( - np.allclose( - actual_t, expect_t, atol=atol, equal_nan=equal_nan), - "Output (" + out_name + ") has diff at " + str(place) + "\nExpect " - + str(expect_t) + "\n" + "But Got" + str(actual_t)) - - -class TestRmspropOp(TestBase): - def check_with_place(self, place, centered, epsilon): - self.setup(centered, epsilon) - scope = core.Scope() - # create and initialize Param Variable - param = scope.var(self.param_name).get_tensor() - param.set(self.param, place) + self.param_tensor = self.scope.var(self.param_name).get_tensor() + self.param_tensor.set(self.param, place) - mean_square = scope.var(self.mean_square_name).get_tensor() - mean_square.set(self.mean_square, place) + self.mean_square_tensor = self.scope.var( + self.mean_square_name).get_tensor() + self.mean_square_tensor.set(self.mean_square, place) - lr = scope.var(self.lr_name).get_tensor() + lr = self.scope.var(self.lr_name).get_tensor() lr.set(self.learning_rate, place) - grad = scope.var(self.grad_name).get_tensor() - grad.set(self.grad, place) + self.moment_tensor = self.scope.var(self.moment_name).get_tensor() + self.moment_tensor.set(self.moment, place) - moment = scope.var(self.moment_name).get_tensor() - moment.set(self.moment, place) + if self.centered: + self.mean_grad_tensor = self.scope.var( + self.mean_grad_name).get_tensor() + self.mean_grad_tensor.set(self.mean_grad, place) - # create and run sgd operator + def check(self, actual_t, expect_t, place, out_name, atol=1e-5): + self.assertTrue( + np.allclose( + actual_t, expect_t, atol=atol), + "Output (" + out_name + ") has diff at " + str(place) + "\nExpect " + + str(expect_t) + "\n" + "But Got" + str(actual_t)) - if self.centered: - mean_grad = scope.var(self.mean_grad_name).get_tensor() - mean_grad.set(self.mean_grad, place) - - rmsprop_op = Operator( - "rmsprop", - Param=self.param_name, - Grad=self.grad_name, - MeanSquare=self.mean_square_name, - MeanGrad=self.mean_grad_name, - Moment=self.moment_name, - LearningRate=self.lr_name, - ParamOut=self.param_name, - MeanSquareOut=self.mean_square_name, - MomentOut=self.moment_name, - MeanGradOut=self.mean_grad_name, - epsilon=self.epsilon, - decay=self.decay, - momentum=self.momentum, - centered=True) - else: - rmsprop_op = Operator( - "rmsprop", - Param=self.param_name, - Grad=self.grad_name, - MeanSquare=self.mean_square_name, - Moment=self.moment_name, - LearningRate=self.lr_name, - ParamOut=self.param_name, - MeanSquareOut=self.mean_square_name, - MomentOut=self.moment_name, - epsilon=self.epsilon, - decay=self.decay, - momentum=self.momentum, - centered=False) - - rmsprop_op.run(scope, place) - - atol = 1e-5 - equal_nan = False + +class TestRmspropOp(TestBase): + def check_with_place(self, + place, + is_sparse, + centered, + size, + row_num=None, + epsilon=1e-6): + self.setup(place, is_sparse, centered, size, row_num, epsilon) + self.run_and_check() + + def run_and_check(self): + grad_name = self.grad_sr_name if self.is_sparse else self.grad_name + + kwargs = { + 'Param': self.param_name, + 'Grad': grad_name, + 'MeanSquare': self.mean_square_name, + 'Moment': self.moment_name, + 'LearningRate': self.lr_name, + 'ParamOut': self.param_name, + 'MeanSquareOut': self.mean_square_name, + 'MomentOut': self.moment_name, + 'epsilon': self.epsilon, + 'decay': self.decay, + 'momentum': self.momentum, + 'centered': self.centered + } if self.centered: - atol = 1e-3 - equal_nan = True + kwargs['MeanGrad'] = self.mean_grad_name + kwargs['MeanGradOut'] = self.mean_grad_name + + rmsprop_op = Operator('rmsprop', **kwargs) + atol = 1e-6 + + rmsprop_op.run(self.scope, self.place) self.check( - np.array(mean_square), self.ms_out, place, self.mean_square_name) + np.array(self.mean_square_tensor), + self.ms_out, + self.place, + self.mean_square_name, + atol=atol) self.check( - np.array(moment), + np.array(self.moment_tensor), self.moment_out, - place, + self.place, self.moment_name, - atol=atol, - equal_nan=equal_nan) + atol=atol) self.check( - np.array(param), + np.array(self.param_tensor), self.param_out, - place, + self.place, self.param_name, - atol=atol, - equal_nan=equal_nan) + atol=atol) if self.centered: self.check( - np.array(mean_grad), self.mg_out, place, self.mean_grad_name) + np.array(self.mean_grad_tensor), self.mg_out, self.place, + self.mean_grad_name) def test_rmsprop(self): places = [core.CPUPlace()] if core.is_compiled_with_cuda(): places.append(core.CUDAPlace(0)) + + size = (128, 320) for place in places: - self.check_with_place(place, False, 1e-6) - self.check_with_place(place, False, 1e-10) - self.check_with_place(place, True, 1e-6) - self.check_with_place(place, True, 1e-10) + for centered in [False, True]: + with fluid.scope_guard(core.Scope()): + self.check_with_place( + place, is_sparse=False, centered=centered, size=size) + + with fluid.scope_guard(core.Scope()): + self.check_with_place( + place, + is_sparse=True, + centered=centered, + row_num=512, + size=size) + + with fluid.scope_guard(core.Scope()): + self.check_with_place( + place, + is_sparse=True, + centered=centered, + row_num=60, + size=size) if __name__ == "__main__": -- GitLab