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
f85f2e83
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
f85f2e83
编写于
9月 14, 2022
作者:
Z
Zhang Jun
提交者:
GitHub
9月 14, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix trt multiclass_nms3 (#45166)
* update * update * update
上级
d9fac780
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
334 addition
and
44 deletion
+334
-44
paddle/fluid/inference/tensorrt/convert/multiclass_nms3_op.cc
...le/fluid/inference/tensorrt/convert/multiclass_nms3_op.cc
+50
-19
paddle/fluid/inference/tensorrt/convert/multiclass_nms_op.cc
paddle/fluid/inference/tensorrt/convert/multiclass_nms_op.cc
+48
-17
paddle/fluid/inference/tensorrt/op_teller.cc
paddle/fluid/inference/tensorrt/op_teller.cc
+3
-2
paddle/fluid/inference/tests/infer_ut/test_ppyolo_mbv3.cc
paddle/fluid/inference/tests/infer_ut/test_ppyolo_mbv3.cc
+1
-1
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_multiclass_nms.py
...unittests/ir/inference/test_trt_convert_multiclass_nms.py
+202
-0
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_multiclass_nms3.py
...nittests/ir/inference/test_trt_convert_multiclass_nms3.py
+30
-5
未找到文件。
paddle/fluid/inference/tensorrt/convert/multiclass_nms3_op.cc
浏览文件 @
f85f2e83
...
@@ -54,18 +54,34 @@ class MultiClassNMS3OpConverter : public OpConverter {
...
@@ -54,18 +54,34 @@ class MultiClassNMS3OpConverter : public OpConverter {
PADDLE_GET_CONST
(
float
,
op_desc
.
GetAttr
(
"nms_threshold"
));
PADDLE_GET_CONST
(
float
,
op_desc
.
GetAttr
(
"nms_threshold"
));
int
keep_top_k
=
PADDLE_GET_CONST
(
int
,
op_desc
.
GetAttr
(
"keep_top_k"
));
int
keep_top_k
=
PADDLE_GET_CONST
(
int
,
op_desc
.
GetAttr
(
"keep_top_k"
));
bool
normalized
=
PADDLE_GET_CONST
(
bool
,
op_desc
.
GetAttr
(
"normalized"
));
bool
normalized
=
PADDLE_GET_CONST
(
bool
,
op_desc
.
GetAttr
(
"normalized"
));
int
num_classes
=
scores_tensor
->
getDimensions
().
d
[
0
];
int
class_index
=
engine_
->
with_dynamic_shape
()
?
1
:
0
;
int
num_classes
=
scores_tensor
->
getDimensions
().
d
[
class_index
];
auto
bboxes_dims
=
bboxes_tensor
->
getDimensions
();
auto
bboxes_dims
=
bboxes_tensor
->
getDimensions
();
nvinfer1
::
Dims3
bboxes_expand_dims
(
bboxes_dims
.
d
[
0
],
1
,
bboxes_dims
.
d
[
1
]);
nvinfer1
::
IShuffleLayer
*
bboxes_expand_layer
=
nullptr
;
auto
*
bboxes_expand_layer
=
nvinfer1
::
IShuffleLayer
*
scores_transpose_layer
=
nullptr
;
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
bboxes_tensor
);
if
(
engine_
->
with_dynamic_shape
())
{
bboxes_expand_layer
->
setReshapeDimensions
(
bboxes_expand_dims
);
nvinfer1
::
Dims4
bboxes_expand_dims
(
bboxes_dims
.
d
[
0
],
bboxes_dims
.
d
[
1
],
1
,
bboxes_dims
.
d
[
2
]);
nvinfer1
::
Permutation
permutation
{
1
,
0
};
bboxes_expand_layer
=
auto
*
scores_transpose_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
bboxes_tensor
);
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
scores_tensor
);
bboxes_expand_layer
->
setReshapeDimensions
(
bboxes_expand_dims
);
scores_transpose_layer
->
setFirstTranspose
(
permutation
);
nvinfer1
::
Permutation
permutation
{
0
,
2
,
1
};
scores_transpose_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
scores_tensor
);
scores_transpose_layer
->
setFirstTranspose
(
permutation
);
}
else
{
nvinfer1
::
Dims3
bboxes_expand_dims
(
bboxes_dims
.
d
[
0
],
1
,
bboxes_dims
.
d
[
1
]);
bboxes_expand_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
bboxes_tensor
);
bboxes_expand_layer
->
setReshapeDimensions
(
bboxes_expand_dims
);
nvinfer1
::
Permutation
permutation
{
1
,
0
};
scores_transpose_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
scores_tensor
);
scores_transpose_layer
->
setFirstTranspose
(
permutation
);
}
std
::
vector
<
nvinfer1
::
ITensor
*>
batch_nms_inputs
;
std
::
vector
<
nvinfer1
::
ITensor
*>
batch_nms_inputs
;
batch_nms_inputs
.
push_back
(
bboxes_expand_layer
->
getOutput
(
0
));
batch_nms_inputs
.
push_back
(
bboxes_expand_layer
->
getOutput
(
0
));
...
@@ -101,27 +117,41 @@ class MultiClassNMS3OpConverter : public OpConverter {
...
@@ -101,27 +117,41 @@ class MultiClassNMS3OpConverter : public OpConverter {
fields
.
size
()
*
sizeof
(
nvinfer1
::
PluginField
)));
fields
.
size
()
*
sizeof
(
nvinfer1
::
PluginField
)));
plugin_collections
->
nbFields
=
static_cast
<
int
>
(
fields
.
size
());
plugin_collections
->
nbFields
=
static_cast
<
int
>
(
fields
.
size
());
plugin_collections
->
fields
=
fields
.
data
();
plugin_collections
->
fields
=
fields
.
data
();
std
::
string
nms_plugin_name
=
"BatchedNMS_TRT"
;
auto
creator
=
GetPluginRegistry
()
->
getPluginCreator
(
"BatchedNMS_TRT"
,
"1"
);
if
(
engine_
->
with_dynamic_shape
())
{
nms_plugin_name
=
"BatchedNMSDynamic_TRT"
;
}
auto
creator
=
GetPluginRegistry
()
->
getPluginCreator
(
nms_plugin_name
.
c_str
(),
"1"
);
auto
batch_nms_plugin
=
auto
batch_nms_plugin
=
creator
->
createPlugin
(
"BatchNMSPlugin"
,
plugin_collections
);
creator
->
createPlugin
(
nms_plugin_name
.
c_str
()
,
plugin_collections
);
free
(
plugin_collections
);
free
(
plugin_collections
);
auto
batch_nms_layer
=
engine_
->
network
()
->
addPluginV2
(
auto
batch_nms_layer
=
engine_
->
network
()
->
addPluginV2
(
batch_nms_inputs
.
data
(),
batch_nms_inputs
.
size
(),
*
batch_nms_plugin
);
batch_nms_inputs
.
data
(),
batch_nms_inputs
.
size
(),
*
batch_nms_plugin
);
// static shape: [keep_topk, 4], [keep_topk], [keep_topk]
// dynamic shape: [bs, keep_topk, 4], [bs, keep_topk], [bs, keep_topk]
auto
nmsed_boxes
=
batch_nms_layer
->
getOutput
(
1
);
auto
nmsed_boxes
=
batch_nms_layer
->
getOutput
(
1
);
auto
nmsed_scores
=
batch_nms_layer
->
getOutput
(
2
);
auto
nmsed_scores
=
batch_nms_layer
->
getOutput
(
2
);
auto
nmsed_classes
=
batch_nms_layer
->
getOutput
(
3
);
auto
nmsed_classes
=
batch_nms_layer
->
getOutput
(
3
);
auto
nmsed_scores_transpose_layer
=
auto
nmsed_scores_transpose_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
nmsed_scores
);
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
nmsed_scores
);
nmsed_scores_transpose_layer
->
setReshapeDimensions
(
nvinfer1
::
Dims2
(
keep_top_k
,
1
));
auto
nmsed_classes_reshape_layer
=
auto
nmsed_classes_reshape_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
nmsed_classes
);
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
nmsed_classes
);
nmsed_classes_reshape_layer
->
setReshapeDimensions
(
if
(
engine_
->
with_dynamic_shape
())
{
nvinfer1
::
Dims2
(
keep_top_k
,
1
));
nmsed_scores_transpose_layer
->
setReshapeDimensions
(
nvinfer1
::
Dims3
(
bboxes_dims
.
d
[
0
],
keep_top_k
,
1
));
nmsed_classes_reshape_layer
->
setReshapeDimensions
(
nvinfer1
::
Dims3
(
bboxes_dims
.
d
[
0
],
keep_top_k
,
1
));
}
else
{
nmsed_scores_transpose_layer
->
setReshapeDimensions
(
nvinfer1
::
Dims2
(
keep_top_k
,
1
));
nmsed_classes_reshape_layer
->
setReshapeDimensions
(
nvinfer1
::
Dims2
(
keep_top_k
,
1
));
}
std
::
vector
<
nvinfer1
::
ITensor
*>
concat_inputs
;
std
::
vector
<
nvinfer1
::
ITensor
*>
concat_inputs
;
concat_inputs
.
push_back
(
nmsed_classes_reshape_layer
->
getOutput
(
0
));
concat_inputs
.
push_back
(
nmsed_classes_reshape_layer
->
getOutput
(
0
));
concat_inputs
.
push_back
(
nmsed_scores_transpose_layer
->
getOutput
(
0
));
concat_inputs
.
push_back
(
nmsed_scores_transpose_layer
->
getOutput
(
0
));
...
@@ -129,7 +159,8 @@ class MultiClassNMS3OpConverter : public OpConverter {
...
@@ -129,7 +159,8 @@ class MultiClassNMS3OpConverter : public OpConverter {
auto
nms_concat_layer
=
TRT_ENGINE_ADD_LAYER
(
auto
nms_concat_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Concatenation
,
concat_inputs
.
data
(),
concat_inputs
.
size
());
engine_
,
Concatenation
,
concat_inputs
.
data
(),
concat_inputs
.
size
());
nms_concat_layer
->
setAxis
(
1
);
int
axis_index
=
engine_
->
with_dynamic_shape
()
?
1
:
0
;
nms_concat_layer
->
setAxis
(
axis_index
+
1
);
// add fake index as output to be consistent with the outputs of
// add fake index as output to be consistent with the outputs of
// multiclass_nms3
// multiclass_nms3
...
...
paddle/fluid/inference/tensorrt/convert/multiclass_nms_op.cc
浏览文件 @
f85f2e83
...
@@ -52,18 +52,34 @@ class MultiClassNMSOpConverter : public OpConverter {
...
@@ -52,18 +52,34 @@ class MultiClassNMSOpConverter : public OpConverter {
PADDLE_GET_CONST
(
float
,
op_desc
.
GetAttr
(
"nms_threshold"
));
PADDLE_GET_CONST
(
float
,
op_desc
.
GetAttr
(
"nms_threshold"
));
int
keep_top_k
=
PADDLE_GET_CONST
(
int
,
op_desc
.
GetAttr
(
"keep_top_k"
));
int
keep_top_k
=
PADDLE_GET_CONST
(
int
,
op_desc
.
GetAttr
(
"keep_top_k"
));
bool
normalized
=
PADDLE_GET_CONST
(
bool
,
op_desc
.
GetAttr
(
"normalized"
));
bool
normalized
=
PADDLE_GET_CONST
(
bool
,
op_desc
.
GetAttr
(
"normalized"
));
int
num_classes
=
scores_tensor
->
getDimensions
().
d
[
0
];
int
class_index
=
engine_
->
with_dynamic_shape
()
?
1
:
0
;
int
num_classes
=
scores_tensor
->
getDimensions
().
d
[
class_index
];
auto
bboxes_dims
=
bboxes_tensor
->
getDimensions
();
auto
bboxes_dims
=
bboxes_tensor
->
getDimensions
();
nvinfer1
::
Dims3
bboxes_expand_dims
(
bboxes_dims
.
d
[
0
],
1
,
bboxes_dims
.
d
[
1
]);
nvinfer1
::
IShuffleLayer
*
bboxes_expand_layer
=
nullptr
;
auto
*
bboxes_expand_layer
=
nvinfer1
::
IShuffleLayer
*
scores_transpose_layer
=
nullptr
;
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
bboxes_tensor
);
if
(
engine_
->
with_dynamic_shape
())
{
bboxes_expand_layer
->
setReshapeDimensions
(
bboxes_expand_dims
);
nvinfer1
::
Dims4
bboxes_expand_dims
(
bboxes_dims
.
d
[
0
],
bboxes_dims
.
d
[
1
],
1
,
bboxes_dims
.
d
[
2
]);
nvinfer1
::
Permutation
permutation
{
1
,
0
};
bboxes_expand_layer
=
auto
*
scores_transpose_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
bboxes_tensor
);
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
scores_tensor
);
bboxes_expand_layer
->
setReshapeDimensions
(
bboxes_expand_dims
);
scores_transpose_layer
->
setFirstTranspose
(
permutation
);
nvinfer1
::
Permutation
permutation
{
0
,
2
,
1
};
scores_transpose_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
scores_tensor
);
scores_transpose_layer
->
setFirstTranspose
(
permutation
);
}
else
{
nvinfer1
::
Dims3
bboxes_expand_dims
(
bboxes_dims
.
d
[
0
],
1
,
bboxes_dims
.
d
[
1
]);
bboxes_expand_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
bboxes_tensor
);
bboxes_expand_layer
->
setReshapeDimensions
(
bboxes_expand_dims
);
nvinfer1
::
Permutation
permutation
{
1
,
0
};
scores_transpose_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
scores_tensor
);
scores_transpose_layer
->
setFirstTranspose
(
permutation
);
}
std
::
vector
<
nvinfer1
::
ITensor
*>
batch_nms_inputs
;
std
::
vector
<
nvinfer1
::
ITensor
*>
batch_nms_inputs
;
batch_nms_inputs
.
push_back
(
bboxes_expand_layer
->
getOutput
(
0
));
batch_nms_inputs
.
push_back
(
bboxes_expand_layer
->
getOutput
(
0
));
...
@@ -100,9 +116,14 @@ class MultiClassNMSOpConverter : public OpConverter {
...
@@ -100,9 +116,14 @@ class MultiClassNMSOpConverter : public OpConverter {
plugin_collections
->
nbFields
=
static_cast
<
int
>
(
fields
.
size
());
plugin_collections
->
nbFields
=
static_cast
<
int
>
(
fields
.
size
());
plugin_collections
->
fields
=
fields
.
data
();
plugin_collections
->
fields
=
fields
.
data
();
auto
creator
=
GetPluginRegistry
()
->
getPluginCreator
(
"BatchedNMS_TRT"
,
"1"
);
std
::
string
nms_plugin_name
=
"BatchedNMS_TRT"
;
if
(
engine_
->
with_dynamic_shape
())
{
nms_plugin_name
=
"BatchedNMSDynamic_TRT"
;
}
auto
creator
=
GetPluginRegistry
()
->
getPluginCreator
(
nms_plugin_name
.
c_str
(),
"1"
);
auto
batch_nms_plugin
=
auto
batch_nms_plugin
=
creator
->
createPlugin
(
"BatchNMSPlugin"
,
plugin_collections
);
creator
->
createPlugin
(
nms_plugin_name
.
c_str
()
,
plugin_collections
);
free
(
plugin_collections
);
free
(
plugin_collections
);
auto
batch_nms_layer
=
engine_
->
network
()
->
addPluginV2
(
auto
batch_nms_layer
=
engine_
->
network
()
->
addPluginV2
(
...
@@ -113,12 +134,21 @@ class MultiClassNMSOpConverter : public OpConverter {
...
@@ -113,12 +134,21 @@ class MultiClassNMSOpConverter : public OpConverter {
auto
nmsed_scores_transpose_layer
=
auto
nmsed_scores_transpose_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
nmsed_scores
);
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
nmsed_scores
);
nmsed_scores_transpose_layer
->
setReshapeDimensions
(
nvinfer1
::
Dims2
(
keep_top_k
,
1
));
auto
nmsed_classes_reshape_layer
=
auto
nmsed_classes_reshape_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
nmsed_classes
);
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
nmsed_classes
);
nmsed_classes_reshape_layer
->
setReshapeDimensions
(
if
(
engine_
->
with_dynamic_shape
())
{
nvinfer1
::
Dims2
(
keep_top_k
,
1
));
nmsed_scores_transpose_layer
->
setReshapeDimensions
(
nvinfer1
::
Dims3
(
bboxes_dims
.
d
[
0
],
keep_top_k
,
1
));
nmsed_classes_reshape_layer
->
setReshapeDimensions
(
nvinfer1
::
Dims3
(
bboxes_dims
.
d
[
0
],
keep_top_k
,
1
));
}
else
{
nmsed_scores_transpose_layer
->
setReshapeDimensions
(
nvinfer1
::
Dims2
(
keep_top_k
,
1
));
nmsed_classes_reshape_layer
->
setReshapeDimensions
(
nvinfer1
::
Dims2
(
keep_top_k
,
1
));
}
std
::
vector
<
nvinfer1
::
ITensor
*>
concat_inputs
;
std
::
vector
<
nvinfer1
::
ITensor
*>
concat_inputs
;
concat_inputs
.
push_back
(
nmsed_classes_reshape_layer
->
getOutput
(
0
));
concat_inputs
.
push_back
(
nmsed_classes_reshape_layer
->
getOutput
(
0
));
...
@@ -127,7 +157,8 @@ class MultiClassNMSOpConverter : public OpConverter {
...
@@ -127,7 +157,8 @@ class MultiClassNMSOpConverter : public OpConverter {
auto
nms_concat_layer
=
TRT_ENGINE_ADD_LAYER
(
auto
nms_concat_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Concatenation
,
concat_inputs
.
data
(),
concat_inputs
.
size
());
engine_
,
Concatenation
,
concat_inputs
.
data
(),
concat_inputs
.
size
());
nms_concat_layer
->
setAxis
(
1
);
int
axis_index
=
engine_
->
with_dynamic_shape
()
?
1
:
0
;
nms_concat_layer
->
setAxis
(
axis_index
+
1
);
RreplenishLayerAndOutput
(
RreplenishLayerAndOutput
(
nms_concat_layer
,
"multiclass_nms"
,
{
output_name
},
test_mode
);
nms_concat_layer
,
"multiclass_nms"
,
{
output_name
},
test_mode
);
...
...
paddle/fluid/inference/tensorrt/op_teller.cc
浏览文件 @
f85f2e83
...
@@ -33,7 +33,10 @@ namespace tensorrt {
...
@@ -33,7 +33,10 @@ namespace tensorrt {
struct
SimpleOpTypeSetTeller
:
public
Teller
{
struct
SimpleOpTypeSetTeller
:
public
Teller
{
SimpleOpTypeSetTeller
()
{
SimpleOpTypeSetTeller
()
{
#if IS_TRT_VERSION_GE(7130)
#if IS_TRT_VERSION_GE(7130)
// use TensorRT plugin
teller_set
.
insert
(
"group_norm"
);
teller_set
.
insert
(
"group_norm"
);
teller_set
.
insert
(
"multiclass_nms3"
);
teller_set
.
insert
(
"multiclass_nms"
);
#endif
#endif
#if IS_TRT_VERSION_GE(7000)
#if IS_TRT_VERSION_GE(7000)
teller_set
.
insert
(
"tile"
);
teller_set
.
insert
(
"tile"
);
...
@@ -278,7 +281,6 @@ struct SimpleOpTypeSetTeller : public Teller {
...
@@ -278,7 +281,6 @@ struct SimpleOpTypeSetTeller : public Teller {
"c_allreduce_prod"
,
"c_allreduce_prod"
,
"roll"
,
"roll"
,
"cast"
,
"cast"
,
"multiclass_nms3"
,
"transformer_input_convert"
,
"transformer_input_convert"
,
"recover_padding"
,
"recover_padding"
,
"remove_padding"
,
"remove_padding"
,
...
@@ -847,7 +849,6 @@ bool OpTeller::Tell(const framework::ir::Node* node,
...
@@ -847,7 +849,6 @@ bool OpTeller::Tell(const framework::ir::Node* node,
}
}
if
(
op_type
==
"multiclass_nms"
||
op_type
==
"multiclass_nms3"
)
{
if
(
op_type
==
"multiclass_nms"
||
op_type
==
"multiclass_nms3"
)
{
if
(
with_dynamic_shape
)
return
false
;
auto
*
block
=
desc
.
Block
();
auto
*
block
=
desc
.
Block
();
if
(
block
==
nullptr
)
{
if
(
block
==
nullptr
)
{
VLOG
(
3
)
<<
"The block desc is nullptr, we can't continue to analyze. "
VLOG
(
3
)
<<
"The block desc is nullptr, we can't continue to analyze. "
...
...
paddle/fluid/inference/tests/infer_ut/test_ppyolo_mbv3.cc
浏览文件 @
f85f2e83
...
@@ -73,7 +73,7 @@ TEST(tensorrt_tester_ppyolo_mbv3, multi_thread4_trt_fp32_bz2) {
...
@@ -73,7 +73,7 @@ TEST(tensorrt_tester_ppyolo_mbv3, multi_thread4_trt_fp32_bz2) {
FLAGS_modeldir
+
"/model.pdiparams"
);
FLAGS_modeldir
+
"/model.pdiparams"
);
config
.
EnableUseGpu
(
100
,
0
);
config
.
EnableUseGpu
(
100
,
0
);
config
.
EnableTensorRtEngine
(
config
.
EnableTensorRtEngine
(
1
<<
2
0
,
2
,
3
,
paddle_infer
::
PrecisionType
::
kFloat32
,
false
,
false
);
1
<<
2
5
,
2
,
3
,
paddle_infer
::
PrecisionType
::
kFloat32
,
false
,
false
);
LOG
(
INFO
)
<<
config
.
Summary
();
LOG
(
INFO
)
<<
config
.
Summary
();
// get groudtruth by disbale ir
// get groudtruth by disbale ir
paddle_infer
::
services
::
PredictorPool
pred_pool_no_ir
(
config_no_ir
,
1
);
paddle_infer
::
services
::
PredictorPool
pred_pool_no_ir
(
config_no_ir
,
1
);
...
...
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_multiclass_nms.py
0 → 100644
浏览文件 @
f85f2e83
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from
trt_layer_auto_scan_test
import
TrtLayerAutoScanTest
,
SkipReasons
from
program_config
import
TensorConfig
,
ProgramConfig
import
numpy
as
np
import
paddle.inference
as
paddle_infer
from
functools
import
partial
from
typing
import
Optional
,
List
,
Callable
,
Dict
,
Any
,
Set
import
unittest
class
TrtConvertMulticlassNMSTest
(
TrtLayerAutoScanTest
):
def
is_program_valid
(
self
,
program_config
:
ProgramConfig
)
->
bool
:
return
True
def
create_inference_config
(
self
,
use_trt
=
True
)
->
paddle_infer
.
Config
:
if
use_trt
:
config
=
paddle_infer
.
Config
()
config
.
disable_glog_info
()
config
.
enable_use_gpu
(
100
,
0
)
config
.
set_optim_cache_dir
(
self
.
cache_dir
)
config
.
switch_ir_debug
()
config
.
enable_tensorrt_engine
(
max_batch_size
=
self
.
trt_param
.
max_batch_size
,
workspace_size
=
self
.
trt_param
.
workspace_size
,
min_subgraph_size
=
self
.
trt_param
.
min_subgraph_size
,
precision_mode
=
self
.
trt_param
.
precision
,
use_static
=
self
.
trt_param
.
use_static
,
use_calib_mode
=
self
.
trt_param
.
use_calib_mode
)
if
len
(
self
.
dynamic_shape
.
min_input_shape
)
!=
0
and
self
.
dynamic_shape
.
min_input_shape
.
keys
(
)
==
self
.
dynamic_shape
.
max_input_shape
.
keys
(
)
and
self
.
dynamic_shape
.
min_input_shape
.
keys
(
)
==
self
.
dynamic_shape
.
opt_input_shape
.
keys
():
config
.
set_trt_dynamic_shape_info
(
self
.
dynamic_shape
.
min_input_shape
,
self
.
dynamic_shape
.
max_input_shape
,
self
.
dynamic_shape
.
opt_input_shape
,
self
.
dynamic_shape
.
disable_trt_plugin_fp16
)
return
config
else
:
config
=
paddle_infer
.
Config
()
config
.
switch_ir_debug
(
True
)
config
.
set_optim_cache_dir
(
self
.
cache_dir
)
config
.
disable_glog_info
()
return
config
def
sample_program_configs
(
self
):
def
generate_boxes
(
batch
,
num_boxes
):
return
np
.
arange
(
batch
*
num_boxes
*
4
,
dtype
=
np
.
float32
).
reshape
([
batch
,
num_boxes
,
4
])
def
generate_scores
(
batch
,
num_boxes
,
num_classes
):
return
np
.
arange
(
batch
*
num_classes
*
num_boxes
,
dtype
=
np
.
float32
).
reshape
(
[
batch
,
num_classes
,
num_boxes
])
# return np.random.rand(batch, num_classes, num_boxes).astype(np.float32)
for
batch
in
[
1
,
2
]:
self
.
batch
=
batch
for
nms_eta
in
[
0.8
,
1.1
]:
for
num_boxes
,
num_classes
in
[[
80
,
100
],
[
40
,
200
],
[
20
,
400
]]:
self
.
num_boxes
,
self
.
num_classes
=
num_boxes
,
num_classes
for
score_threshold
in
[
0.01
,
]:
ops_config
=
[{
"op_type"
:
"multiclass_nms"
,
"op_inputs"
:
{
"BBoxes"
:
[
"input_bboxes"
],
"Scores"
:
[
"input_scores"
],
},
"op_outputs"
:
{
"Out"
:
[
"nms_output_boxes"
],
},
"op_attrs"
:
{
"background_label"
:
-
1
,
"score_threshold"
:
score_threshold
,
"nms_top_k"
:
num_boxes
,
"keep_top_k"
:
num_boxes
,
"nms_threshold"
:
0.3
,
"normalized"
:
False
,
"nms_eta"
:
nms_eta
}
}]
ops
=
self
.
generate_op_config
(
ops_config
)
program_config
=
ProgramConfig
(
ops
=
ops
,
weights
=
{},
inputs
=
{
"input_bboxes"
:
TensorConfig
(
data_gen
=
partial
(
generate_boxes
,
batch
,
num_boxes
)),
"input_scores"
:
TensorConfig
(
data_gen
=
partial
(
generate_scores
,
batch
,
num_boxes
,
num_classes
))
},
outputs
=
[
"nms_output_boxes"
])
yield
program_config
def
sample_predictor_configs
(
self
,
program_config
)
->
(
paddle_infer
.
Config
,
List
[
int
],
float
):
def
generate_dynamic_shape
(
attrs
):
# The last dim of input_bboxes should be static.
self
.
dynamic_shape
.
min_input_shape
=
{
"input_bboxes"
:
[
1
,
self
.
num_boxes
,
4
],
"input_scores"
:
[
1
,
self
.
num_classes
,
self
.
num_boxes
],
}
self
.
dynamic_shape
.
max_input_shape
=
{
"input_bboxes"
:
[
8
,
self
.
num_boxes
,
4
],
"input_scores"
:
[
8
,
self
.
num_classes
,
self
.
num_boxes
],
}
self
.
dynamic_shape
.
opt_input_shape
=
{
"input_bboxes"
:
[
self
.
batch
,
self
.
num_boxes
,
4
],
"input_scores"
:
[
self
.
batch
,
self
.
num_classes
,
self
.
num_boxes
],
}
def
clear_dynamic_shape
():
self
.
dynamic_shape
.
min_input_shape
=
{}
self
.
dynamic_shape
.
max_input_shape
=
{}
self
.
dynamic_shape
.
opt_input_shape
=
{}
def
generate_trt_nodes_num
(
attrs
,
dynamic_shape
):
return
1
,
2
attrs
=
[
program_config
.
ops
[
i
].
attrs
for
i
in
range
(
len
(
program_config
.
ops
))
]
# for static_shape
clear_dynamic_shape
()
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Float32
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
attrs
,
False
),
1e-5
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Half
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
attrs
,
False
),
1e-2
# for dynamic_shape
generate_dynamic_shape
(
attrs
)
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Float32
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
attrs
,
True
),
1e-5
# self.trt_param.precision = paddle_infer.PrecisionType.Half
# yield self.create_inference_config(), generate_trt_nodes_num(
# attrs, True), (1e-2, 1e-2)
def
assert_tensors_near
(
self
,
atol
:
float
,
rtol
:
float
,
tensor
:
Dict
[
str
,
np
.
array
],
baseline
:
Dict
[
str
,
np
.
array
]):
# the order of tensorrt outputs are not consistent with paddle
for
key
,
arr
in
tensor
.
items
():
if
key
==
"nms_output_boxes"
:
basline_arr
=
np
.
array
(
sorted
(
baseline
[
key
].
reshape
((
-
1
,
6
)),
key
=
lambda
i
:
[
i
[
0
],
i
[
1
]]))
arr
=
np
.
array
(
sorted
(
arr
.
reshape
((
-
1
,
6
)),
key
=
lambda
i
:
[
i
[
0
],
i
[
1
]]))
else
:
basline_arr
=
np
.
array
(
baseline
[
key
].
reshape
((
-
1
,
1
)))
arr
=
np
.
array
(
arr
.
reshape
((
-
1
,
1
)))
self
.
assertTrue
(
basline_arr
.
shape
==
arr
.
shape
,
"The output shapes are not equal, the baseline shape is "
+
str
(
basline_arr
.
shape
)
+
', but got '
+
str
(
arr
.
shape
))
diff
=
abs
(
basline_arr
-
arr
)
np
.
testing
.
assert_allclose
(
basline_arr
,
arr
,
rtol
=
rtol
,
atol
=
atol
,
err_msg
=
'Output has diff, Maximum absolute error: {}'
.
format
(
np
.
amax
(
diff
)))
def
assert_op_size
(
self
,
trt_engine_num
,
paddle_op_num
):
# tensorrt op num is not consistent with paddle
return
True
def
test
(
self
):
self
.
trt_param
.
workspace_size
=
1
<<
25
self
.
run_test
()
if
__name__
==
"__main__"
:
unittest
.
main
()
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_multiclass_nms3.py
浏览文件 @
f85f2e83
...
@@ -71,8 +71,10 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest):
...
@@ -71,8 +71,10 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest):
# return np.random.rand(batch, num_classes, num_boxes).astype(np.float32)
# return np.random.rand(batch, num_classes, num_boxes).astype(np.float32)
for
batch
in
[
1
,
2
]:
for
batch
in
[
1
,
2
]:
for
num_boxes
in
[
4
,
12
]:
self
.
batch
=
batch
for
num_classes
in
[
2
,
6
]:
for
nms_eta
in
[
0.8
,
1.1
]:
for
num_boxes
,
num_classes
in
[[
80
,
100
],
[
40
,
200
],
[
20
,
400
]]:
self
.
num_boxes
,
self
.
num_classes
=
num_boxes
,
num_classes
for
score_threshold
in
[
for
score_threshold
in
[
0.01
,
0.01
,
]:
]:
...
@@ -94,7 +96,7 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest):
...
@@ -94,7 +96,7 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest):
"keep_top_k"
:
num_boxes
,
"keep_top_k"
:
num_boxes
,
"nms_threshold"
:
0.3
,
"nms_threshold"
:
0.3
,
"normalized"
:
False
,
"normalized"
:
False
,
"nms_eta"
:
1.1
"nms_eta"
:
nms_eta
}
}
}]
}]
ops
=
self
.
generate_op_config
(
ops_config
)
ops
=
self
.
generate_op_config
(
ops_config
)
...
@@ -114,12 +116,26 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest):
...
@@ -114,12 +116,26 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest):
"nms_output_boxes"
,
"nms_output_num"
,
"nms_output_boxes"
,
"nms_output_num"
,
"nms_output_index"
"nms_output_index"
])
])
yield
program_config
yield
program_config
def
sample_predictor_configs
(
def
sample_predictor_configs
(
self
,
program_config
)
->
(
paddle_infer
.
Config
,
List
[
int
],
float
):
self
,
program_config
)
->
(
paddle_infer
.
Config
,
List
[
int
],
float
):
def
generate_dynamic_shape
(
attrs
):
# The last dim of input_bboxes should be static.
self
.
dynamic_shape
.
min_input_shape
=
{
"input_bboxes"
:
[
1
,
self
.
num_boxes
,
4
],
"input_scores"
:
[
1
,
self
.
num_classes
,
self
.
num_boxes
],
}
self
.
dynamic_shape
.
max_input_shape
=
{
"input_bboxes"
:
[
8
,
self
.
num_boxes
,
4
],
"input_scores"
:
[
8
,
self
.
num_classes
,
self
.
num_boxes
],
}
self
.
dynamic_shape
.
opt_input_shape
=
{
"input_bboxes"
:
[
self
.
batch
,
self
.
num_boxes
,
4
],
"input_scores"
:
[
self
.
batch
,
self
.
num_classes
,
self
.
num_boxes
],
}
def
clear_dynamic_shape
():
def
clear_dynamic_shape
():
self
.
dynamic_shape
.
min_input_shape
=
{}
self
.
dynamic_shape
.
min_input_shape
=
{}
self
.
dynamic_shape
.
max_input_shape
=
{}
self
.
dynamic_shape
.
max_input_shape
=
{}
...
@@ -141,6 +157,15 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest):
...
@@ -141,6 +157,15 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest):
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
attrs
,
False
),
1e-2
attrs
,
False
),
1e-2
# for dynamic_shape
generate_dynamic_shape
(
attrs
)
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Float32
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
attrs
,
True
),
1e-5
# self.trt_param.precision = paddle_infer.PrecisionType.Half
# yield self.create_inference_config(), generate_trt_nodes_num(
# attrs, True), (1e-2, 1e-2)
def
assert_tensors_near
(
self
,
atol
:
float
,
rtol
:
float
,
def
assert_tensors_near
(
self
,
atol
:
float
,
rtol
:
float
,
tensor
:
Dict
[
str
,
np
.
array
],
tensor
:
Dict
[
str
,
np
.
array
],
baseline
:
Dict
[
str
,
np
.
array
]):
baseline
:
Dict
[
str
,
np
.
array
]):
...
@@ -176,7 +201,7 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest):
...
@@ -176,7 +201,7 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest):
return
True
return
True
def
test
(
self
):
def
test
(
self
):
self
.
trt_param
.
workspace_size
=
1
<<
2
0
self
.
trt_param
.
workspace_size
=
1
<<
2
5
self
.
run_test
()
self
.
run_test
()
...
...
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