Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
PaddlePaddle
Paddle
提交
8b1918f5
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2298
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
Pages
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
提交
Issue看板
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
8b1918f5
编写于
5月 09, 2018
作者:
C
chengduo
提交者:
GitHub
5月 09, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #10454 from chengduoZH/fix_fetchop
Fix fetch_op_handle
上级
2bff03bc
a459764d
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
105 addition
and
22 deletion
+105
-22
paddle/fluid/framework/details/fetch_op_handle.cc
paddle/fluid/framework/details/fetch_op_handle.cc
+3
-1
paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc
paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc
+3
-1
paddle/fluid/framework/details/send_op_handle.cc
paddle/fluid/framework/details/send_op_handle.cc
+3
-1
python/paddle/fluid/tests/unittests/test_parallel_executor.py
...on/paddle/fluid/tests/unittests/test_parallel_executor.py
+96
-19
未找到文件。
paddle/fluid/framework/details/fetch_op_handle.cc
浏览文件 @
8b1918f5
...
...
@@ -49,7 +49,9 @@ void FetchOpHandle::RunImpl() {
platform
::
DeviceContextPool
::
Instance
().
Get
(
platform
::
CPUPlace
());
for
(
auto
*
input
:
inputs_
)
{
auto
*
var
=
static_cast
<
VarHandle
*>
(
input
);
var
->
generated_op_
->
Wait
(
cpu_ctx
);
if
(
var
->
generated_op_
)
{
var
->
generated_op_
->
Wait
(
cpu_ctx
);
}
}
tensors_
.
resize
(
inputs_
.
size
());
auto
*
var_handle
=
static_cast
<
VarHandle
*>
(
inputs_
[
0
]);
...
...
paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc
浏览文件 @
8b1918f5
...
...
@@ -36,7 +36,9 @@ void NCCLAllReduceOpHandle::RunImpl() {
// Wait input done
for
(
auto
*
in
:
inputs_
)
{
auto
&
p
=
static_cast
<
VarHandle
*>
(
in
)
->
place_
;
in
->
generated_op_
->
Wait
(
dev_ctxes_
[
p
]);
if
(
in
->
generated_op_
)
{
in
->
generated_op_
->
Wait
(
dev_ctxes_
[
p
]);
}
}
auto
&
var_name
=
static_cast
<
VarHandle
*>
(
this
->
inputs_
[
0
])
->
name_
;
...
...
paddle/fluid/framework/details/send_op_handle.cc
浏览文件 @
8b1918f5
...
...
@@ -32,7 +32,9 @@ void SendOpHandle::RunImpl() {
if
(
in
->
DebugString
()
==
"dummy"
)
{
// HACK
continue
;
}
in
->
generated_op_
->
Wait
(
dev_ctxes_
[
p
]);
if
(
in
->
generated_op_
)
{
in
->
generated_op_
->
Wait
(
dev_ctxes_
[
p
]);
}
}
auto
&
tmp_scope
=
local_scope_
->
FindVar
(
kLocalExecScopeName
)
->
Get
<
Scope
*>
();
// FIXME(wuyi): can not use RunAndRecordEvent here, for it will cause dead
...
...
python/paddle/fluid/tests/unittests/test_parallel_executor.py
浏览文件 @
8b1918f5
...
...
@@ -12,7 +12,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import
numpy
import
numpy
as
np
import
unittest
import
paddle.fluid
as
fluid
...
...
@@ -243,7 +243,7 @@ class TestParallelExecutorBase(unittest.TestCase):
begin
=
time
.
time
()
first_loss
,
=
run_executor
(
exe
=
exe
,
feed
=
feed_dict
,
fetch_list
=
[
loss
.
name
])
first_loss
=
n
umpy
.
array
(
first_loss
)
first_loss
=
n
p
.
array
(
first_loss
)
for
i
in
xrange
(
iter
):
run_executor
(
exe
=
exe
,
feed
=
feed_dict
,
fetch_list
=
[])
...
...
@@ -256,7 +256,7 @@ class TestParallelExecutorBase(unittest.TestCase):
print
"%.4f Instance per second"
%
(
(
batch_size
*
iter
+
2
)
/
(
end
-
begin
))
last_loss
=
n
umpy
.
array
(
last_loss
)
last_loss
=
n
p
.
array
(
last_loss
)
print
first_loss
,
last_loss
# self.assertGreater(first_loss[0], last_loss[0])
...
...
@@ -284,8 +284,8 @@ class TestMNIST(TestParallelExecutorBase):
self
.
check_network_convergence
(
simple_fc_net
)
self
.
check_network_convergence
(
simple_fc_net
,
allow_op_delay
=
True
)
img
=
n
umpy
.
zeros
(
shape
=
[
32
,
784
],
dtype
=
'float32'
)
label
=
n
umpy
.
ones
(
shape
=
[
32
,
1
],
dtype
=
'int64'
)
img
=
n
p
.
zeros
(
shape
=
[
32
,
784
],
dtype
=
'float32'
)
label
=
n
p
.
ones
(
shape
=
[
32
,
1
],
dtype
=
'int64'
)
self
.
check_network_convergence
(
simple_fc_net
,
feed_dict
=
{
"image"
:
img
,
"label"
:
label
})
...
...
@@ -294,8 +294,8 @@ class TestMNIST(TestParallelExecutorBase):
self
.
check_simple_fc_convergence
()
def
check_simple_fc_parallel_accuracy
(
self
):
img
=
n
umpy
.
zeros
(
shape
=
[
32
,
784
],
dtype
=
'float32'
)
label
=
n
umpy
.
ones
(
shape
=
[
32
,
1
],
dtype
=
'int64'
)
img
=
n
p
.
zeros
(
shape
=
[
32
,
784
],
dtype
=
'float32'
)
label
=
n
p
.
ones
(
shape
=
[
32
,
1
],
dtype
=
'int64'
)
single_first_loss
,
single_last_loss
=
self
.
check_network_convergence
(
method
=
simple_fc_net
,
seed
=
1000
,
...
...
@@ -319,8 +319,8 @@ class TestMNIST(TestParallelExecutorBase):
def
check_batchnorm_fc_convergence
(
self
):
self
.
check_network_convergence
(
fc_with_batchnorm
)
img
=
n
umpy
.
zeros
(
shape
=
[
32
,
784
],
dtype
=
'float32'
)
label
=
n
umpy
.
ones
(
shape
=
[
32
,
1
],
dtype
=
'int64'
)
img
=
n
p
.
zeros
(
shape
=
[
32
,
784
],
dtype
=
'float32'
)
label
=
n
p
.
ones
(
shape
=
[
32
,
1
],
dtype
=
'int64'
)
self
.
check_network_convergence
(
fc_with_batchnorm
,
feed_dict
=
{
"image"
:
img
,
"label"
:
label
})
...
...
@@ -404,9 +404,6 @@ class ModelHyperParams(object):
dropout
=
0.1
import
numpy
as
np
def
prepare_batch_input
(
insts
,
src_pad_idx
,
trg_pad_idx
,
n_head
):
"""
Pad the instances to the max sequence length in batch, and generate the
...
...
@@ -533,9 +530,8 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase):
opt
.
minimize
(
loss
)
batch_size
=
32
image
=
numpy
.
random
.
normal
(
size
=
(
batch_size
,
784
)).
astype
(
'float32'
)
label
=
numpy
.
random
.
randint
(
0
,
10
,
(
batch_size
,
1
),
dtype
=
"int64"
)
image
=
np
.
random
.
normal
(
size
=
(
batch_size
,
784
)).
astype
(
'float32'
)
label
=
np
.
random
.
randint
(
0
,
10
,
(
batch_size
,
1
),
dtype
=
"int64"
)
place
=
fluid
.
CUDAPlace
(
0
)
exe
=
fluid
.
Executor
(
place
)
...
...
@@ -552,12 +548,12 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase):
for
i
in
xrange
(
5
):
test_loss
,
=
test_exe
.
run
([
loss
.
name
],
feed
=
feed_dict
)
test_loss
=
n
umpy
.
array
(
test_loss
)
test_loss
=
n
p
.
array
(
test_loss
)
train_loss
,
=
train_exe
.
run
([
loss
.
name
],
feed
=
feed_dict
)
train_loss
=
n
umpy
.
array
(
train_loss
)
train_loss
=
n
p
.
array
(
train_loss
)
self
.
assertTrue
(
n
umpy
.
allclose
(
n
p
.
allclose
(
train_loss
,
test_loss
,
atol
=
1e-8
),
"Train loss: "
+
str
(
train_loss
)
+
"
\n
Test loss:"
+
str
(
test_loss
))
...
...
@@ -712,7 +708,7 @@ class TestCRFModel(unittest.TestCase):
data
=
train_data
()
for
i
in
xrange
(
10
):
cur_batch
=
next
(
data
)
print
map
(
n
umpy
.
array
,
print
map
(
n
p
.
array
,
pe
.
run
(
feed
=
feeder
.
feed
(
cur_batch
),
fetch_list
=
[
avg_cost
.
name
]))[
0
]
...
...
@@ -721,3 +717,84 @@ class TestCRFModel(unittest.TestCase):
def
test_update_dense_parameter
(
self
):
self
.
check_network_convergence
(
is_sparse
=
False
)
# test fetch all the variables of global_block
import
paddle.dataset.flowers
as
flowers
import
math
def
Lenet
(
data
,
class_dim
):
conv1
=
fluid
.
layers
.
conv2d
(
data
,
32
,
5
,
1
,
act
=
None
)
bn1
=
fluid
.
layers
.
batch_norm
(
conv1
,
act
=
'relu'
)
pool1
=
fluid
.
layers
.
pool2d
(
bn1
,
2
,
'max'
,
2
)
conv2
=
fluid
.
layers
.
conv2d
(
pool1
,
50
,
5
,
1
,
act
=
None
)
bn2
=
fluid
.
layers
.
batch_norm
(
conv2
,
act
=
'relu'
)
pool2
=
fluid
.
layers
.
pool2d
(
bn2
,
2
,
'max'
,
2
)
fc1
=
fluid
.
layers
.
fc
(
pool2
,
size
=
500
,
act
=
'relu'
)
fc2
=
fluid
.
layers
.
fc
(
fc1
,
size
=
class_dim
,
act
=
'softmax'
)
return
fc2
class
TestFetchOp
(
unittest
.
TestCase
):
def
parallel_exe
(
self
,
train_inputs
,
seed
):
main
=
fluid
.
Program
()
startup
=
fluid
.
Program
()
startup
.
random_seed
=
seed
with
fluid
.
program_guard
(
main
,
startup
):
data
=
fluid
.
layers
.
data
(
name
=
'image'
,
shape
=
[
3
,
224
,
224
],
dtype
=
'float32'
)
label
=
fluid
.
layers
.
data
(
name
=
'label'
,
shape
=
[
1
],
dtype
=
'int64'
)
out
=
Lenet
(
data
,
class_dim
=
102
)
loss
=
fluid
.
layers
.
cross_entropy
(
input
=
out
,
label
=
label
)
loss
=
fluid
.
layers
.
mean
(
loss
)
opt
=
fluid
.
optimizer
.
Momentum
(
learning_rate
=
0.1
,
momentum
=
0.9
,
regularization
=
fluid
.
regularizer
.
L2Decay
(
1e-4
))
opt
.
minimize
(
loss
)
# TODO(zcd): I found that onece the memory optimizer is open,
# parallel_exe doesn't fetch some variable, such as conv2d_0.b_0@GRAD,
# conv2d_1.b_0@GRAD. Those variables should not be pruned.
# fluid.memory_optimize(main)
place
=
fluid
.
CUDAPlace
(
0
)
exe
=
fluid
.
Executor
(
place
)
exe
.
run
(
startup
)
feeder
=
fluid
.
DataFeeder
(
place
=
place
,
feed_list
=
[
data
,
label
])
pe
=
fluid
.
ParallelExecutor
(
use_cuda
=
True
,
loss_name
=
loss
.
name
,
main_program
=
main
)
fetch_list
=
[]
all_vars
=
main
.
global_block
().
vars
for
k
,
v
in
all_vars
.
iteritems
():
if
'tmp'
not
in
k
and
k
[
0
]
is
not
'_'
or
v
.
persistable
:
fetch_list
.
append
(
k
)
for
data
in
train_inputs
:
ret
=
pe
.
run
(
fetch_list
,
feed
=
feeder
.
feed
(
data
))
for
i
in
range
(
len
(
fetch_list
)):
assert
not
math
.
isnan
(
np
.
sum
(
ret
[
i
]))
and
\
not
math
.
isinf
(
np
.
sum
(
ret
[
i
]))
def
test_update_sparse_parameter
(
self
):
tst_reader
=
paddle
.
batch
(
flowers
.
test
(
use_xmap
=
False
),
batch_size
=
16
)
tst_reader_iter
=
tst_reader
()
iters
=
3
train_inputs
=
[]
for
i
in
range
(
iters
):
train_inputs
.
append
(
tst_reader_iter
.
next
())
self
.
parallel_exe
(
train_inputs
,
seed
=
1
)
if
__name__
==
'__main__'
:
unittest
.
main
()
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录