Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 1 年 前同步成功

通知 2299
Star 20931
Fork 5422
P
Paddle
提交 84d93003 编写于 作者: S sneaxiy
浏览文件
操作

test=develop

上级 6f748a03
隐藏空白更改
内联 并排
Showing with 151 addition and 107 deletion
paddle/fluid/operators/rmsprop_op.h
浏览文件 @ 84d93003
......@@ -131,21 +131,21 @@ template <typename DeviceContext, typename T>
class RmspropOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
using Tensor = framework::LoDTensor;
using LoDTensor = framework::LoDTensor;
auto *grad_var = ctx.InputVar("Grad");
auto *param_out = ctx.Output<Tensor>("ParamOut");
auto *moment_out = ctx.Output<Tensor>("MomentOut");
auto *mean_square_out = ctx.Output<Tensor>("MeanSquareOut");
auto *param_out = ctx.Output<LoDTensor>("ParamOut");
auto *moment_out = ctx.Output<LoDTensor>("MomentOut");
auto *mean_square_out = ctx.Output<LoDTensor>("MeanSquareOut");
auto epsilon = static_cast<T>(ctx.Attr<float>("epsilon"));
auto rho = static_cast<T>(ctx.Attr<float>("decay"));
auto momentum = static_cast<T>(ctx.Attr<float>("momentum"));
bool centered = ctx.Attr<bool>("centered");
auto &p_tensor = *ctx.Input<Tensor>("Param");
auto &ms_tensor = *ctx.Input<Tensor>("MeanSquare");
auto &lr_tensor = *ctx.Input<Tensor>("LearningRate");
auto &mom_tensor = *ctx.Input<Tensor>("Moment");
auto &p_tensor = *ctx.Input<LoDTensor>("Param");
auto &ms_tensor = *ctx.Input<LoDTensor>("MeanSquare");
auto &lr_tensor = *ctx.Input<LoDTensor>("LearningRate");
auto &mom_tensor = *ctx.Input<LoDTensor>("Moment");
PADDLE_ENFORCE_EQ(&p_tensor, param_out,
"Param and ParamOut must be the same Tensor");
......@@ -157,8 +157,8 @@ class RmspropOpKernel : public framework::OpKernel<T> {
auto &dev_ctx = ctx.template device_context<DeviceContext>();
size_t limit = static_cast<size_t>(ms_tensor.numel());
if (grad_var->IsType<Tensor>()) {
auto &grad_tensor = grad_var->Get<Tensor>();
if (grad_var->IsType<LoDTensor>()) {
auto &grad_tensor = grad_var->Get<LoDTensor>();
if (std::is_same<DeviceContext, platform::CPUDeviceContext>::value) {
auto &place =
......@@ -176,9 +176,9 @@ class RmspropOpKernel : public framework::OpKernel<T> {
ms_out.device(place) = rho * ms + (1 - rho) * g * g;
if (centered) {
auto &mg_tensor = *ctx.Input<Tensor>("MeanGrad");
auto &mg_tensor = *ctx.Input<LoDTensor>("MeanGrad");
auto mg = EigenVector<T>::Flatten(mg_tensor);
auto *mean_grad_out = ctx.Output<Tensor>("MeanGradOut");
auto *mean_grad_out = ctx.Output<LoDTensor>("MeanGradOut");
PADDLE_ENFORCE(&mg_tensor, mean_grad_out,
"MeanGrad and MeanGradOut must be the same Tensor");
auto mg_out = EigenVector<T>::Flatten(*mean_grad_out);
......@@ -196,8 +196,8 @@ class RmspropOpKernel : public framework::OpKernel<T> {
DenseRmspropGradFunctor<T> grad_func(grad_tensor.data<T>());
platform::ForRange<DeviceContext> for_range(dev_ctx, limit);
if (centered) {
auto &mg_tensor = *ctx.Input<Tensor>("MeanGrad");
auto *mean_grad_out = ctx.Output<Tensor>("MeanGradOut");
auto &mg_tensor = *ctx.Input<LoDTensor>("MeanGrad");
auto *mean_grad_out = ctx.Output<LoDTensor>("MeanGradOut");
PADDLE_ENFORCE(&mg_tensor, mean_grad_out,
"MeanGrad and MeanGradOut must be the same Tensor");
for_range(CenteredRmspropFunctor<T, DenseRmspropGradFunctor<T>>(
......@@ -241,8 +241,8 @@ class RmspropOpKernel : public framework::OpKernel<T> {
row_numel, row_count);
if (centered) {
auto &mg_tensor = *ctx.Input<Tensor>("MeanGrad");
auto *mean_grad_out = ctx.Output<Tensor>("MeanGradOut");
auto &mg_tensor = *ctx.Input<LoDTensor>("MeanGrad");
auto *mean_grad_out = ctx.Output<LoDTensor>("MeanGradOut");
PADDLE_ENFORCE(&mg_tensor, mean_grad_out,
"MeanGrad and MeanGradOut must be the same Tensor");
for_range(CenteredRmspropFunctor<T, SparseRmspropGradFunctor<T>>(
......
python/paddle/fluid/tests/unittests/test_rmsprop_op.py
浏览文件 @ 84d93003
......@@ -19,29 +19,72 @@ 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
......@@ -61,118 +104,119 @@ 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)
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__":
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或