m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge pull request #13874 from dzhwinter/fix/momentum

Browse Source 
add sparse update momentum. test=develop
updatereadme
dzhwinter 7 years ago committed by GitHub
parent 6a54c3de1f 00e8791f66
commit 29382db625
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 448 additions and 105 deletions
Show Stats Download Patch File Download Diff File

53
paddle/fluid/operators/momentum_op.cc
Unescape View File

@ -24,7 +24,7 @@ class MomentumOp : public framework::OperatorWithKernel {
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext *ctx) const override {
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Param"),
"Input(param) of Momentum should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Grad"),
@ -45,12 +45,15 @@ class MomentumOp : public framework::OperatorWithKernel {
"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.");
if (ctx->GetInputsVarType("Grad")[0] ==
framework::proto::VarType::LOD_TENSOR) {
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");
@ -58,13 +61,34 @@ class MomentumOp : public framework::OperatorWithKernel {
ctx->SetOutputDim("VelocityOut", param_dim);
}
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext &ctx) const override {
auto input_data_type =
framework::ToDataType(ctx.Input<Tensor>("Param")->type());
const framework::ExecutionContext& ctx) const override {
auto input_data_type = framework::GetDataTypeOfVar(ctx.InputVar("Param"));
return framework::OpKernelType(input_data_type, ctx.GetPlace());
}
};
class MomentumOpInferVarType : 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 if (block->FindRecursiveOrCreateVar(input_var).GetType() ==
framework::proto::VarType::LOD_TENSOR) {
block->FindRecursiveOrCreateVar(out_var).SetType(
framework::proto::VarType::LOD_TENSOR);
} else {
PADDLE_THROW(
"Only support LodTensor and SelectedRows, Unexpected Input Type.");
}
}
}
};
class MomentumOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
@ -115,6 +139,9 @@ $$
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_WITHOUT_GRADIENT(momentum, ops::MomentumOp, ops::MomentumOpMaker);
REGISTER_OP_CPU_KERNEL(momentum, ops::MomentumOpKernel<float>,
ops::MomentumOpKernel<double>);
REGISTER_OPERATOR(momentum, ops::MomentumOp, ops::MomentumOpMaker,
paddle::framework::EmptyGradOpMaker,
ops::MomentumOpInferVarType);
REGISTER_OP_CPU_KERNEL(
momentum, ops::MomentumOpKernel<paddle::platform::CPUDeviceContext, float>,
ops::MomentumOpKernel<paddle::platform::CPUDeviceContext, double>);

75
paddle/fluid/operators/momentum_op.cu
Unescape View File

@ -15,76 +15,7 @@ limitations under the License. */
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/momentum_op.h"
namespace paddle {
namespace operators {
template <typename T>
__global__ void MomentumKernel(const T* p, const T* g, const T* v,
const T* learning_rate, const T mu,
const int64_t num, bool use_nesterov, T* p_out,
T* v_out) {
T lr = learning_rate[0];
if (use_nesterov) {
for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num;
i += blockDim.x * gridDim.x) {
T g_val = g[i];
T v_new = v[i] * mu + g_val;
v_out[i] = v_new;
p_out[i] = p[i] - (g_val + v_new * mu) * lr;
}
} else {
for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num;
i += blockDim.x * gridDim.x) {
T v_new = v[i] * mu + g[i];
v_out[i] = v_new;
p_out[i] = p[i] - lr * v_new;
}
}
}
template <typename T>
class MomentumOpCUDAKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
const auto* param_var = ctx.InputVar("Param");
PADDLE_ENFORCE(param_var->IsType<framework::LoDTensor>(),
"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<framework::LoDTensor>(),
"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<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");
T* p_out = param_out->mutable_data<T>(ctx.GetPlace());
T* v_out = velocity_out->mutable_data<T>(ctx.GetPlace());
T mu = static_cast<T>(ctx.Attr<float>("mu"));
bool use_nesterov = ctx.Attr<bool>("use_nesterov");
auto* p = param->data<T>();
auto* v = velocity->data<T>();
auto* g = grad->data<T>();
auto* lr = learning_rate->data<T>();
int block = 512;
int grid = (param->numel() + block - 1) / block;
MomentumKernel<T><<<grid, block, 0, ctx.cuda_device_context().stream()>>>(
p, g, v, lr, mu, param->numel(), use_nesterov, p_out, v_out);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(momentum, ops::MomentumOpCUDAKernel<float>,
ops::MomentumOpCUDAKernel<double>);
REGISTER_OP_CUDA_KERNEL(
momentum, ops::MomentumOpKernel<paddle::platform::CUDADeviceContext, float>,
ops::MomentumOpKernel<paddle::platform::CUDADeviceContext, double>);

331
paddle/fluid/operators/momentum_op.h
View File

File diff suppressed because it is too large Load Diff

94
python/paddle/fluid/tests/unittests/test_momentum_op.py
Unescape View File

@ -16,6 +16,8 @@ from __future__ import print_function
import unittest
import numpy as np
import paddle.fluid.core as core
from paddle.fluid.op import Operator
from op_test import OpTest
@ -88,5 +90,97 @@ class TestMomentumOp2(OpTest):
self.check_output()
class TestSparseMomentumOp(unittest.TestCase):
def setUp(self):
self.use_nesterov = False
def check_with_place(self, place):
self.init_kernel()
scope = core.Scope()
# create and initialize Grad Variable
height = 10
rows = [0, 4, 7]
row_numel = 12
mu = 1.0
use_nesterov = self.use_nesterov
# create and initialize Param Variable
param = scope.var('Param').get_tensor()
param_array = np.full((height, row_numel), 5.0).astype("float32")
param.set(param_array, place)
param_out = scope.var("ParamOut").get_tensor()
param_out_array = np.full((height, row_numel), 0.0).astype("float32")
param_out.set(param_out_array, place)
grad_selected_rows = scope.var('Grad').get_selected_rows()
grad_selected_rows.set_height(height)
grad_selected_rows.set_rows(rows)
grad_np_array = np.ones((len(rows), row_numel)).astype("float32")
grad_np_array[0, 0] = 2.0
grad_np_array[2, 8] = 4.0
grad_tensor = grad_selected_rows.get_tensor()
grad_tensor.set(grad_np_array, place)
velocity = scope.var('Velocity').get_tensor()
velocity_np_array = np.ones((height, row_numel)).astype("float32")
velocity.set(velocity_np_array, place)
velocity_out = scope.var('VelocityOut').get_tensor()
velocity_out_np_array = np.full((height, row_numel),
0.0).astype("float32")
velocity_out.set(velocity_out_np_array, place)
# create and initialize LeraningRate Variable
lr = scope.var('LearningRate').get_tensor()
lr_array = np.full((1), 2.0).astype("float32")
lr.set(lr_array, place)
# create and run operator
op = Operator(
"momentum",
Param='Param',
Grad='Grad',
Velocity='Velocity',
ParamOut='ParamOut',
VelocityOut='VelocityOut',
LearningRate='LearningRate',
mu=mu,
use_nesterov=use_nesterov)
op.run(scope, place)
# get and compare result
param_out_np_array = np.array(param_out)
velocity_out_np_array = np.array(velocity_out)
# TODO(dzh): add a more suitable general numpy interface
# for sparse update.
_grad_np_array = np.full((height, row_numel), 0.0).astype("float32")
for i in range(len(rows)):
_grad_np_array[rows[i]] = grad_np_array[i]
_velocity_out = mu * velocity_np_array + _grad_np_array
_param = param_array
if use_nesterov:
_param_out = _param - (_grad_np_array + _velocity_out * mu
) * lr_array
else:
_param_out = _param - lr_array * _velocity_out
self.assertTrue((_velocity_out == velocity_out_np_array).all())
self.assertTrue((_param_out == param_out_np_array).all())
def init_kernel(self):
pass
def test_sparse_momentum(self):
places = [core.CPUPlace()]
if core.is_compiled_with_cuda():
places.append(core.CUDAPlace(0))
for place in places:
self.check_with_place(place)
class TestSparseMomentumOp2(TestSparseMomentumOp):
def init_kernel(self):
self.use_nesterov = True
if __name__ == "__main__":
unittest.main()

Write Preview
Loading…
Cancel
Save