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
d159a6b3
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 >>
未验证
提交
d159a6b3
编写于
10月 11, 2019
作者:
B
Bai Yifan
提交者:
GitHub
10月 11, 2019
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix PaddleSlim distillation demo configs (#3505)
* fix some configs * fix doc
上级
ea8299a9
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
72 addition
and
88 deletion
+72
-88
PaddleSlim/classification/distillation/README.md
PaddleSlim/classification/distillation/README.md
+25
-4
PaddleSlim/classification/distillation/compress.py
PaddleSlim/classification/distillation/compress.py
+44
-63
PaddleSlim/classification/distillation/configs/mobilenetv1_resnet50_distillation.yaml
...stillation/configs/mobilenetv1_resnet50_distillation.yaml
+1
-1
PaddleSlim/classification/distillation/configs/resnet34_resnet50_distillation.yaml
.../distillation/configs/resnet34_resnet50_distillation.yaml
+2
-2
PaddleSlim/classification/distillation/run.sh
PaddleSlim/classification/distillation/run.sh
+0
-18
未找到文件。
PaddleSlim/classification/distillation/README.md
浏览文件 @
d159a6b3
...
@@ -139,16 +139,30 @@ strategies:
...
@@ -139,16 +139,30 @@ strategies:
| baseline | 70.99%/89.68% |
| baseline | 70.99%/89.68% |
| 蒸馏后 | - |
| 蒸馏后 | - |
>训练超参:
#### 训练超参
-
batch size: 256
-
lr_strategy: piecewise_decay
-
step_epochs: 30, 60, 90
-
num_epochs: 120
-
l2_decay: 4e-5
-
init lr: 0.1
### MobileNetV2
### MobileNetV2
| FLOPS | top1_acc/top5_acc |
| FLOPS | top1_acc/top5_acc |
| -------- | ----------------- |
| -------- | ----------------- |
| baseline | 72.15%/90.65% |
| baseline | 72.15%/90.65% |
| 蒸馏后 |
-
|
| 蒸馏后 |
70.66%/90.42%
|
>训练超参:
#### 训练超参
-
batch size: 256
-
lr_strategy: piecewise_decay
-
step_epochs: 30, 60, 90
-
num_epochs: 120
-
l2_decay: 4e-5
-
init lr: 0.1
### ResNet34
### ResNet34
...
@@ -157,6 +171,13 @@ strategies:
...
@@ -157,6 +171,13 @@ strategies:
| baseline | 74.57%/92.14% |
| baseline | 74.57%/92.14% |
| 蒸馏后 | - |
| 蒸馏后 | - |
>训练超参:
#### 训练超参
-
batch size: 256
-
lr_strategy: piecewise_decay
-
step_epochs: 30, 60, 90
-
num_epochs: 120
-
l2_decay: 4e-5
-
init lr: 0.1
## FAQ
## FAQ
PaddleSlim/classification/distillation/compress.py
浏览文件 @
d159a6b3
...
@@ -34,7 +34,6 @@ add_arg('pretrained_model', str, None, "Whether to use pretraine
...
@@ -34,7 +34,6 @@ add_arg('pretrained_model', str, None, "Whether to use pretraine
add_arg
(
'teacher_model'
,
str
,
None
,
"Set the teacher network to use."
)
add_arg
(
'teacher_model'
,
str
,
None
,
"Set the teacher network to use."
)
add_arg
(
'teacher_pretrained_model'
,
str
,
None
,
"Whether to use pretrained model."
)
add_arg
(
'teacher_pretrained_model'
,
str
,
None
,
"Whether to use pretrained model."
)
add_arg
(
'compress_config'
,
str
,
None
,
"The config file for compression with yaml format."
)
add_arg
(
'compress_config'
,
str
,
None
,
"The config file for compression with yaml format."
)
add_arg
(
'quant_only'
,
bool
,
False
,
"Only do quantization-aware training."
)
# yapf: enable
# yapf: enable
model_list
=
[
m
for
m
in
dir
(
models
)
if
"__"
not
in
m
]
model_list
=
[
m
for
m
in
dir
(
models
)
if
"__"
not
in
m
]
...
@@ -50,45 +49,25 @@ def compress(args):
...
@@ -50,45 +49,25 @@ def compress(args):
# model definition
# model definition
model
=
models
.
__dict__
[
args
.
model
]()
model
=
models
.
__dict__
[
args
.
model
]()
if
args
.
model
is
"GoogleNet"
:
if
args
.
model
==
'ResNet34'
:
out0
,
out1
,
out2
=
model
.
net
(
input
=
image
,
class_dim
=
args
.
class_dim
)
model
.
prefix_name
=
'res34'
cost0
=
fluid
.
layers
.
cross_entropy
(
input
=
out0
,
label
=
label
)
out
=
model
.
net
(
input
=
image
,
class_dim
=
args
.
class_dim
,
fc_name
=
'fc_0'
)
cost1
=
fluid
.
layers
.
cross_entropy
(
input
=
out1
,
label
=
label
)
cost2
=
fluid
.
layers
.
cross_entropy
(
input
=
out2
,
label
=
label
)
avg_cost0
=
fluid
.
layers
.
mean
(
x
=
cost0
)
avg_cost1
=
fluid
.
layers
.
mean
(
x
=
cost1
)
avg_cost2
=
fluid
.
layers
.
mean
(
x
=
cost2
)
avg_cost
=
avg_cost0
+
0.3
*
avg_cost1
+
0.3
*
avg_cost2
acc_top1
=
fluid
.
layers
.
accuracy
(
input
=
out0
,
label
=
label
,
k
=
1
)
acc_top5
=
fluid
.
layers
.
accuracy
(
input
=
out0
,
label
=
label
,
k
=
5
)
else
:
else
:
if
args
.
model
==
'ResNet34'
:
out
=
model
.
net
(
input
=
image
,
class_dim
=
args
.
class_dim
)
model
.
prefix_name
=
'res34'
cost
=
fluid
.
layers
.
cross_entropy
(
input
=
out
,
label
=
label
)
out
=
model
.
net
(
input
=
image
,
avg_cost
=
fluid
.
layers
.
mean
(
x
=
cost
)
class_dim
=
args
.
class_dim
,
acc_top1
=
fluid
.
layers
.
accuracy
(
input
=
out
,
label
=
label
,
k
=
1
)
fc_name
=
'fc_0'
)
acc_top5
=
fluid
.
layers
.
accuracy
(
input
=
out
,
label
=
label
,
k
=
5
)
else
:
out
=
model
.
net
(
input
=
image
,
class_dim
=
args
.
class_dim
)
cost
=
fluid
.
layers
.
cross_entropy
(
input
=
out
,
label
=
label
)
avg_cost
=
fluid
.
layers
.
mean
(
x
=
cost
)
acc_top1
=
fluid
.
layers
.
accuracy
(
input
=
out
,
label
=
label
,
k
=
1
)
acc_top5
=
fluid
.
layers
.
accuracy
(
input
=
out
,
label
=
label
,
k
=
5
)
#print("="*50+"student_model_params"+"="*50)
#print("="*50+"student_model_params"+"="*50)
#for v in fluid.default_main_program().list_vars():
#for v in fluid.default_main_program().list_vars():
# print(v.name, v.shape)
# print(v.name, v.shape)
val_program
=
fluid
.
default_main_program
().
clone
()
val_program
=
fluid
.
default_main_program
().
clone
()
if
args
.
quant_only
:
boundaries
=
[
boundaries
=
[
args
.
total_images
/
args
.
batch_size
*
30
,
args
.
total_images
/
args
.
total_images
/
args
.
batch_size
*
10
,
args
.
batch_size
*
60
,
args
.
total_images
/
args
.
batch_size
*
90
args
.
total_images
/
args
.
batch_size
*
16
]
]
values
=
[
0.1
,
0.01
,
0.001
,
0.0001
]
values
=
[
1e-4
,
1e-5
,
1e-6
]
else
:
boundaries
=
[
args
.
total_images
/
args
.
batch_size
*
30
,
args
.
total_images
/
args
.
batch_size
*
60
,
args
.
total_images
/
args
.
batch_size
*
90
]
values
=
[
0.1
,
0.01
,
0.001
,
0.0001
]
opt
=
fluid
.
optimizer
.
Momentum
(
opt
=
fluid
.
optimizer
.
Momentum
(
momentum
=
0.9
,
momentum
=
0.9
,
learning_rate
=
fluid
.
layers
.
piecewise_decay
(
learning_rate
=
fluid
.
layers
.
piecewise_decay
(
...
@@ -117,37 +96,39 @@ def compress(args):
...
@@ -117,37 +96,39 @@ def compress(args):
teacher_programs
=
[]
teacher_programs
=
[]
distiller_optimizer
=
None
distiller_optimizer
=
None
if
args
.
teacher_model
:
teacher_model
=
models
.
__dict__
[
args
.
teacher_model
](
prefix_name
=
'res50'
)
# define teacher program
teacher_program
=
fluid
.
Program
()
startup_program
=
fluid
.
Program
()
with
fluid
.
program_guard
(
teacher_program
,
startup_program
):
img
=
teacher_program
.
global_block
().
_clone_variable
(
image
,
force_persistable
=
False
)
predict
=
teacher_model
.
net
(
img
,
class_dim
=
args
.
class_dim
,
fc_name
=
'fc_0'
)
#print("="*50+"teacher_model_params"+"="*50)
#for v in teacher_program.list_vars():
# print(v.name, v.shape)
exe
.
run
(
startup_program
)
assert
args
.
teacher_pretrained_model
and
os
.
path
.
exists
(
args
.
teacher_pretrained_model
),
"teacher_pretrained_model should be set when teacher_model is not None."
def
if_exist
(
var
):
teacher_model
=
models
.
__dict__
[
args
.
teacher_model
](
prefix_name
=
'res50'
)
return
os
.
path
.
exists
(
# define teacher program
os
.
path
.
join
(
args
.
teacher_pretrained_model
,
var
.
name
))
teacher_program
=
fluid
.
Program
()
startup_program
=
fluid
.
Program
()
with
fluid
.
program_guard
(
teacher_program
,
startup_program
):
img
=
teacher_program
.
global_block
().
_clone_variable
(
image
,
force_persistable
=
False
)
predict
=
teacher_model
.
net
(
img
,
class_dim
=
args
.
class_dim
,
fc_name
=
'fc_0'
)
#print("="*50+"teacher_model_params"+"="*50)
#for v in teacher_program.list_vars():
# print(v.name, v.shape)
#return
exe
.
run
(
startup_program
)
assert
args
.
teacher_pretrained_model
and
os
.
path
.
exists
(
args
.
teacher_pretrained_model
),
"teacher_pretrained_model should be set when teacher_model is not None."
def
if_exist
(
var
):
return
os
.
path
.
exists
(
os
.
path
.
join
(
args
.
teacher_pretrained_model
,
var
.
name
))
fluid
.
io
.
load_vars
(
fluid
.
io
.
load_vars
(
exe
,
exe
,
args
.
teacher_pretrained_model
,
args
.
teacher_pretrained_model
,
main_program
=
teacher_program
,
main_program
=
teacher_program
,
predicate
=
if_exist
)
predicate
=
if_exist
)
distiller_optimizer
=
opt
distiller_optimizer
=
opt
teacher_programs
.
append
(
teacher_program
.
clone
(
for_test
=
True
))
teacher_programs
.
append
(
teacher_program
.
clone
(
for_test
=
True
))
com_pass
=
Compressor
(
com_pass
=
Compressor
(
place
,
place
,
...
...
PaddleSlim/classification/distillation/configs/mobilenetv1_resnet50_distillation.yaml
浏览文件 @
d159a6b3
...
@@ -3,7 +3,7 @@ distillers:
...
@@ -3,7 +3,7 @@ distillers:
fsp_distiller
:
fsp_distiller
:
class
:
'
FSPDistiller'
class
:
'
FSPDistiller'
teacher_pairs
:
[[
'
res50_res2a_branch2a.conv2d.output.1.tmp_0'
,
'
res50_res3a_branch2a.conv2d.output.1.tmp_0'
]]
teacher_pairs
:
[[
'
res50_res2a_branch2a.conv2d.output.1.tmp_0'
,
'
res50_res3a_branch2a.conv2d.output.1.tmp_0'
]]
student_pairs
:
[[
'
depthwise_conv2d_1.tmp_0'
,
'
conv2d_3
.tmp_0'
]]
student_pairs
:
[[
'
depthwise_conv2d_1.tmp_0'
,
'
depthwise_conv2d_2
.tmp_0'
]]
distillation_loss_weight
:
1
distillation_loss_weight
:
1
l2_distiller
:
l2_distiller
:
class
:
'
L2Distiller'
class
:
'
L2Distiller'
...
...
PaddleSlim/classification/distillation/configs/resnet34_resnet50_distillation.yaml
浏览文件 @
d159a6b3
...
@@ -2,8 +2,8 @@ version: 1.0
...
@@ -2,8 +2,8 @@ version: 1.0
distillers
:
distillers
:
fsp_distiller
:
fsp_distiller
:
class
:
'
FSPDistiller'
class
:
'
FSPDistiller'
teacher_pairs
:
[[
'
res50_res2a_branch2a.conv2d.output.1.tmp_0'
,
'
res50_res2
a_branch2c.conv2d.output.1.tmp_0'
],
[
'
res50_res3b_branch2a.conv2d.output.1.tmp_0'
,
'
res50_res3b_branch2c
.conv2d.output.1.tmp_0'
]]
teacher_pairs
:
[[
'
res50_res2a_branch2a.conv2d.output.1.tmp_0'
,
'
res50_res2
b_branch2a.conv2d.output.1.tmp_0'
],
[
'
res50_res4b_branch2a.conv2d.output.1.tmp_0'
,
'
res50_res4c_branch2a
.conv2d.output.1.tmp_0'
]]
student_pairs
:
[[
'
res34_res2a_branch2a.conv2d.output.1.tmp_0'
,
'
res34_res2a_branch
2c.conv2d.output.1.tmp_0'
],
[
'
res34_res3b_branch2a.conv2d.output.1.tmp_0'
,
'
res34_res3b_branch2c
.conv2d.output.1.tmp_0'
]]
student_pairs
:
[[
'
res34_res2a_branch2a.conv2d.output.1.tmp_0'
,
'
res34_res2a_branch
1.conv2d.output.1.tmp_0'
],
[
'
res34_res4b_branch2a.conv2d.output.1.tmp_0'
,
'
res34_res4c_branch2a
.conv2d.output.1.tmp_0'
]]
distillation_loss_weight
:
1
distillation_loss_weight
:
1
l2_distiller
:
l2_distiller
:
class
:
'
L2Distiller'
class
:
'
L2Distiller'
...
...
PaddleSlim/classification/distillation/run.sh
浏览文件 @
d159a6b3
...
@@ -44,12 +44,6 @@ python -u compress.py \
...
@@ -44,12 +44,6 @@ python -u compress.py \
>
mobilenet_v1.log 2>&1 &
>
mobilenet_v1.log 2>&1 &
tailf mobilenet_v1.log
tailf mobilenet_v1.log
cd
${
pretrain_dir
}
/ResNet50_pretrained
for
files
in
$(
ls
res50_
*
)
do
mv
$files
${
files
#*_
}
done
cd
-
## for mobilenet_v2 distillation
## for mobilenet_v2 distillation
#cd ${pretrain_dir}/ResNet50_pretrained
#cd ${pretrain_dir}/ResNet50_pretrained
#for files in $(ls res50_*)
#for files in $(ls res50_*)
...
@@ -67,12 +61,6 @@ cd -
...
@@ -67,12 +61,6 @@ cd -
#--compress_config ./configs/mobilenetv2_resnet50_distillation.yaml\
#--compress_config ./configs/mobilenetv2_resnet50_distillation.yaml\
#> mobilenet_v2.log 2>&1 &
#> mobilenet_v2.log 2>&1 &
#tailf mobilenet_v2.log
#tailf mobilenet_v2.log
#
#cd ${pretrain_dir}/ResNet50_pretrained
#for files in $(ls res50_*)
# do mv $files ${files#*_}
#done
#cd -
## for resnet34 distillation
## for resnet34 distillation
#cd ${pretrain_dir}/ResNet50_pretrained
#cd ${pretrain_dir}/ResNet50_pretrained
...
@@ -91,9 +79,3 @@ cd -
...
@@ -91,9 +79,3 @@ cd -
#--compress_config ./configs/resnet34_resnet50_distillation.yaml \
#--compress_config ./configs/resnet34_resnet50_distillation.yaml \
#> resnet34.log 2>&1 &
#> resnet34.log 2>&1 &
#tailf resnet34.log
#tailf resnet34.log
#
#cd ${pretrain_dir}/ResNet50_pretrained
#for files in $(ls res50_*)
# do mv $files ${files#*_}
#done
#cd -
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