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
models
提交
523d28c2
M
models
项目概览
PaddlePaddle
/
models
大约 1 年 前同步成功
通知
222
Star
6828
Fork
2962
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
602
列表
看板
标记
里程碑
合并请求
255
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
M
models
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
602
Issue
602
列表
看板
标记
里程碑
合并请求
255
合并请求
255
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 >>
提交
523d28c2
编写于
12月 03, 2018
作者:
L
liuhongyu
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add cudnn lm
上级
545ae923
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
36 addition
and
14 deletion
+36
-14
fluid/PaddleNLP/language_model/lstm/.run_ce.sh
fluid/PaddleNLP/language_model/lstm/.run_ce.sh
+1
-0
fluid/PaddleNLP/language_model/lstm/args.py
fluid/PaddleNLP/language_model/lstm/args.py
+6
-1
fluid/PaddleNLP/language_model/lstm/lm_model.py
fluid/PaddleNLP/language_model/lstm/lm_model.py
+19
-5
fluid/PaddleNLP/language_model/lstm/train.py
fluid/PaddleNLP/language_model/lstm/train.py
+10
-8
未找到文件。
fluid/PaddleNLP/language_model/lstm/.run_ce.sh
浏览文件 @
523d28c2
...
...
@@ -7,5 +7,6 @@ python train.py \
--data_path
data/simple-examples/data/
\
--model_type
small
\
--use_gpu
True
\
--rnn_model
static
\
--enable_ce
| python _ce.py
fluid/PaddleNLP/language_model/lstm/args.py
浏览文件 @
523d28c2
...
...
@@ -26,7 +26,12 @@ def parse_args():
"--model_type"
,
type
=
str
,
default
=
"small"
,
help
=
"model_type [test|small|med|big]"
)
help
=
"model_type [test|small|medium|large]"
)
parser
.
add_argument
(
"--rnn_model"
,
type
=
str
,
default
=
"static"
,
help
=
"model_type [static|padding|cudnn]"
)
parser
.
add_argument
(
"--data_path"
,
type
=
str
,
help
=
"all the data for train,valid,test"
)
parser
.
add_argument
(
'--para_init'
,
action
=
'store_true'
)
...
...
fluid/PaddleNLP/language_model/lstm/lm_model.py
浏览文件 @
523d28c2
...
...
@@ -28,7 +28,8 @@ def lm_model(hidden_size,
num_layers
=
2
,
num_steps
=
20
,
init_scale
=
0.1
,
dropout
=
None
):
dropout
=
None
,
rnn_model
=
'static'
):
def
padding_rnn
(
input_embedding
,
len
=
3
,
init_hidden
=
None
,
init_cell
=
None
):
weight_1_arr
=
[]
weight_2_arr
=
[]
...
...
@@ -243,7 +244,7 @@ def lm_model(hidden_size,
input
=
x
,
size
=
[
vocab_size
,
hidden_size
],
dtype
=
'float32'
,
is_sparse
=
Tru
e
,
is_sparse
=
Fals
e
,
param_attr
=
fluid
.
ParamAttr
(
name
=
'embedding_para'
,
initializer
=
fluid
.
initializer
.
UniformInitializer
(
...
...
@@ -255,9 +256,22 @@ def lm_model(hidden_size,
x_emb
,
dropout_prob
=
dropout
,
dropout_implementation
=
'upscale_in_train'
)
rnn_out
,
last_hidden
,
last_cell
=
padding_rnn
(
x_emb
,
len
=
num_steps
,
init_hidden
=
init_hidden
,
init_cell
=
init_cell
)
if
rnn_model
==
"padding"
:
rnn_out
,
last_hidden
,
last_cell
=
padding_rnn
(
x_emb
,
len
=
num_steps
,
init_hidden
=
init_hidden
,
init_cell
=
init_cell
)
elif
rnn_model
==
"static"
:
rnn_out
,
last_hidden
,
last_cell
=
encoder_static
(
x_emb
,
len
=
num_steps
,
init_hidden
=
init_hidden
,
init_cell
=
init_cell
)
elif
rnn_model
==
"cudnn"
:
x_emb
=
layers
.
transpose
(
x_emb
,
perm
=
[
1
,
0
,
2
])
rnn_out
,
last_hidden
,
last_cell
=
layers
.
lstm
(
x_emb
,
init_hidden
,
init_cell
,
num_steps
,
hidden_size
,
num_layers
,
\
is_bidirec
=
False
,
\
default_initializer
=
fluid
.
initializer
.
UniformInitializer
(
low
=-
init_scale
,
high
=
init_scale
)
)
rnn_out
=
layers
.
transpose
(
rnn_out
,
perm
=
[
1
,
0
,
2
])
else
:
print
(
"type not support"
)
return
rnn_out
=
layers
.
reshape
(
rnn_out
,
shape
=
[
-
1
,
num_steps
,
hidden_size
])
...
...
fluid/PaddleNLP/language_model/lstm/train.py
浏览文件 @
523d28c2
...
...
@@ -77,6 +77,7 @@ def save_para_npz(train_prog, train_exe):
def
train
():
args
=
parse_args
()
model_type
=
args
.
model_type
rnn_model
=
args
.
rnn_model
logger
=
logging
.
getLogger
(
"lm"
)
logger
.
setLevel
(
logging
.
INFO
)
formatter
=
logging
.
Formatter
(
...
...
@@ -157,7 +158,8 @@ def train():
num_layers
=
num_layers
,
num_steps
=
num_steps
,
init_scale
=
init_scale
,
dropout
=
dropout
)
dropout
=
dropout
,
rnn_model
=
rnn_model
)
# clone from default main program and use it as the validation program
main_program
=
fluid
.
default_main_program
()
inference_program
=
fluid
.
default_main_program
().
clone
(
for_test
=
True
)
...
...
@@ -206,18 +208,19 @@ def train():
def
eval
(
data
):
# when eval the batch_size set to 1
eval_data_iter
=
reader
.
get_data_iter
(
data
,
1
,
num_steps
)
eval_data_iter
=
reader
.
get_data_iter
(
data
,
batch_size
,
num_steps
)
total_loss
=
0.0
iters
=
0
init_hidden
=
np
.
zeros
((
num_layers
,
1
,
hidden_size
),
dtype
=
'float32'
)
init_cell
=
np
.
zeros
((
num_layers
,
1
,
hidden_size
),
dtype
=
'float32'
)
init_hidden
=
np
.
zeros
((
num_layers
,
batch_size
,
hidden_size
),
dtype
=
'float32'
)
init_cell
=
np
.
zeros
((
num_layers
,
batch_size
,
hidden_size
),
dtype
=
'float32'
)
for
batch_id
,
batch
in
enumerate
(
eval_data_iter
):
input_data_feed
=
prepare_input
(
batch
,
init_hidden
,
init_cell
,
epoch_id
,
with_lr
=
False
)
fetch_outs
=
exe
.
run
(
inference_program
,
feed
=
input_data_feed
,
fetch_list
=
[
loss
.
name
,
last_hidden
.
name
,
last_cell
.
name
])
fetch_list
=
[
loss
.
name
,
last_hidden
.
name
,
last_cell
.
name
],
use_program_cache
=
True
)
cost_train
=
np
.
array
(
fetch_outs
[
0
])
init_hidden
=
np
.
array
(
fetch_outs
[
1
])
...
...
@@ -283,9 +286,8 @@ def train():
print
(
"train ppl"
,
ppl
[
0
])
if
epoch_id
==
max_epoch
-
1
and
args
.
enable_ce
:
print
(
"ptblm
\t
lstm_language_model_duration
\t
%s"
%
(
total_time
/
max_epoch
))
print
(
"ptblm
\t
lstm_language_model_loss
\t
%s"
%
ppl
[
0
])
print
(
"lstm_language_model_duration
\t
%s"
%
(
total_time
/
max_epoch
))
print
(
"lstm_language_model_loss
\t
%s"
%
ppl
[
0
])
model_path
=
os
.
path
.
join
(
"model_new/"
,
str
(
epoch_id
))
if
not
os
.
path
.
isdir
(
model_path
):
...
...
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