Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 337
Star 4
Fork 1
P
Paddle-Lite
提交 e80f61c5 编写于 作者: R Ray Liu 提交者： GitHub
浏览文件
操作

Merge branch 'develop' into develop

上级 f8c8f8a6 26771012
隐藏空白更改
内联 并排
Showing with 769 addition and 171 deletion
.gitignore
浏览文件 @ e80f61c5
......@@ -92,3 +92,4 @@ metal/images/
metal/paddle-mobile/paddle-mobile/CPU/libpaddle-mobile.a
*.xcuserdatad/
*/xcuserdata/
/venv/
CMakeLists.txt
浏览文件 @ e80f61c5
......@@ -20,6 +20,7 @@ set(CMAKE_CXX_FLAGS "-O3 -s -DNDEBUG ${CMAKE_CXX_FLAGS}")
if(IS_IOS)
set(CMAKE_CXX_FLAGS "-mfpu=neon -marm -fobjc-abi-version=2 -fobjc-arc \
-std=gnu++11 -stdlib=libc++ -isysroot ${CMAKE_OSX_SYSROOT} ${CMAKE_CXX_FLAGS}")
add_compile_options(-fembed-bitcode)
else()
set(CMAKE_CXX_FLAGS "-std=c++11 ${CMAKE_CXX_FLAGS}")
endif()
......
README.md
浏览文件 @ e80f61c5
......@@ -110,7 +110,8 @@ Paddle-Mobile是PaddlePaddle组织下的项目，是一个致力于嵌入式平
### 开发文档
开发文档主要是关于编译、运行等问题。做为开发者，它可以和贡献文档共同结合使用。
[开发文档链接](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/doc/development_doc.md)
* [iOS](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/doc/development_ios.md)
* [Android](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/doc/development_android.md)
### 贡献文档
- [贡献文档链接](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/CONTRIBUTING.md)
......
benchmark/arm_benchmark.md 0 → 100644
浏览文件 @ e80f61c5
|mobilenet arm v7|1线程|2线程|4线程|
|------------|----|-----|-----|
|麒麟970(ms)|108.180|63.935|37.545|
|麒麟960(ms)|108.588|63.073|36.822|
|高通845(ms)|85.952|48.890|28.641|
|高通835(ms)|105.434|62.752|37.131|
|||||
|mobilenetssd arm v7|1线程|2线程|4线程|
|麒麟970(ms)|212.686|127.205|77.485|
|麒麟960(ms)|212.641|125.338|75.250|
|高通845(ms)|182.863|95.671|56.857|
|高通835(ms)|213.849|127.717|77.006|
|||||
|googlenet(v1) arm v7|1线程|2线程|4线程|
|麒麟970(ms)|335.288|234.559|161.295|
|麒麟960(ms)|354.443|232.642|157.815|
|高通845(ms)|282.007|173.146|122.148|
|高通835(ms)|341.250|233.354|158.554|
|||||
|squeezenet arm v7|1线程|2线程|4线程|
|麒麟970(ms)|83.726|57.944|36.923|
|麒麟960(ms)|85.835|55.762|36.496|
|高通845(ms)|71.301|41.618|28.785|
|高通835(ms)|82.407|56.176|36.455|
|||||
|yolo arm v7|1线程|2线程|4线程|
|麒麟970(ms)|129.658|79.993|49.969|
|麒麟960(ms)|130.208|78.791|48.390|
|高通845(ms)|109.244|61.736|40.600|
|高通835(ms)|130.402|80.863|50.359|
测试机型信息：
麒麟970:荣耀v10 (2.36GHz * 4 + 1.8GHz * 4)
麒麟960:华为mate9 (2.36GHz * 4 + 1.8GHz * 4)
骁龙835:小米6 (2.45GHz * 4 + 1.9GHz * 4)
骁龙845:OPPO FindX (2.80GHz * 4 + 1.8GHz * 4)
\ No newline at end of file
benchmark/metal_benchmark.md 0 → 100644
浏览文件 @ e80f61c5
|mobilenetfssd|速度|
|------------|-----|
|A9(ms)|33.78|
|A10(ms)|24.05|
|A11(ms)|17.15|
|||
|genet|速度|
|A9(ms) |3.49|
|A10(ms)|2.54|
|A11(ms)|1.43|
\ No newline at end of file
doc/design_doc.md
浏览文件 @ e80f61c5
......@@ -3,7 +3,7 @@
#### 以下是 paddle-mobile 代码的执行流程图:
![执行流程图](http://otkwwi4x8.bkt.clouddn.com/2018-07-02-15305189473720.png)
![执行流程图](http://mms-graph.bj.bcebos.com/paddle-mobile/git_images/flow_chart.png)
#### 主要分为: Loader 模块、 Program 模块、 Executor 模块、 op 模块、 kernel 模块、scope variable Tensor 模块
......@@ -14,12 +14,12 @@
先来看一下模型, 模型分为两种结构:
一种为参数文件是散开的, 如下图, 红框为模型结构的 protobuf 文件, 其余为参数文件
![模型描述](http://otkwwi4x8.bkt.clouddn.com/2018-07-02-15305190629577.png)
![模型描述](http://mms-graph.bj.bcebos.com/paddle-mobile/git_images/model_desc.png)
另一种为参数文件结合在一起的, 如下图, 红框内为模型结构描述的 protobuf 文件, 另一个文件为结合在一起的参数文件
![模型描述combined](http://otkwwi4x8.bkt.clouddn.com/2018-07-02-15305191057130.png)
![模型描述combined](http://mms-graph.bj.bcebos.com/paddle-mobile/git_images/model_desc_combined.png)
loader 模块的作用是将模型结构信息 load 进内存, 将红框内的 protobuf 文件 load 进内存, 并对模型结构进行优化(如将几个细粒度的 op 融合成 粗粒度的 op, 如将 conv、 add、 batchnorm、 relu 融合为 conv\_add\_batchnorm\_relu).
......@@ -161,7 +161,7 @@ sh build.sh android yolo
### 五. kernel
kernel 为 op 的底层运算实现, 主要有两个函数, Init 和 Compute, 分别用来初始化、预处理 和 运算操作, 值得提出的是, kernel 会根据泛型特化到不同的平台, 如图所示:
![设备特化]![](http://otkwwi4x8.bkt.clouddn.com/2018-07-02-15305191401976.png)
![设备特化](http://mms-graph.bj.bcebos.com/paddle-mobile/git_images/devices.png)
不同平台的 kernel 实现, 为同一个 kernel 类不同泛型的特化实现, 目前有三个平台, arm、mali、fpga, 图中的 central-arm-func\ 目录为 op kernel 的 arm 实现, 它承担了 arm\ 目录下 kernel 的底层实现, 同时 arm 处理器作为中央处理器, central-arm-func\ 也可以作为其他协处理器的底层实现, 如: fpga 的某一个 op kernel 还没有 fpga 协处理器的实现, 就可以直接调用使用这里的 arm 实现.
......
doc/development_doc.md → doc/development_android.md
浏览文件 @ e80f61c5
### iOS&Android开发文档
# iOS开发文档
## 编译
```sh
# 在 paddle-mobile 目录下:
cd tools
sh build.sh ios
# 如果只想编译某个特定模型的 op, 则需执行以下命令
sh build.sh ios googlenet
# 在这个文件夹下, 你可以拿到生成的 .a 库
cd ../build/release/ios/build
```
#### 常见问题:
1. No iOS SDK's found in default search path ...
这个问题是因为 tools/ios-cmake/ios.toolchain.cmake 找不到你最近使用的 iOS SDK 路径, 所以需要自己进行指定,
以我当前的环境为例: 在 tools/ios-cmake/ios.toolchain.cmake 143行前添加我本地的 iOS SDK 路径: set(CMAKE_IOS_SDK_ROOT "/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk")
## 集成
```
将上一步生成的:
libpaddle-mobile.a
/src/ios_io/ 下的
PaddleMobile.h
```
拖入工程
#### oc 接口
接口如下:
```
/*
创建对象
*/
- (instancetype)init;
/*
load 模型, 开辟内存
*/
- (BOOL)load:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath;
/*
进行预测, means 和 scale 为训练模型时的预处理参数, 如训练时没有做这些预处理则直接使用 predict
*/
- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim means:(NSArray<NSNumber *> *)means scale:(float)scale;
/*
进行预测
*/
- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim;
/*
清理内存
*/
- (void)clear;
```
# Android开发文档
用户可通过如下两种方式，交叉编译Android平台上适用的paddle-mobile库：
......
doc/development_ios.md 0 → 100644
浏览文件 @ e80f61c5
# iOS开发文档
## CPU
需要: xcode
### 编译
```sh
# 在 paddle-mobile 目录下:
cd tools
sh build.sh ios
# 如果只想编译某个特定模型的 op, 则需执行以下命令
sh build.sh ios googlenet
# 在这个文件夹下, 你可以拿到生成的 .a 库
cd ../build/release/ios/build
```
#### 常见问题:
1. No iOS SDK's found in default search path ...
这个问题是因为 tools/ios-cmake/ios.toolchain.cmake 找不到你最近使用的 iOS SDK 路径, 所以需要自己进行指定,
以我当前的环境为例: 在 tools/ios-cmake/ios.toolchain.cmake 143行前添加我本地的 iOS SDK 路径: set(CMAKE_IOS_SDK_ROOT "/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk")
### 集成
```
将上一步生成的:
libpaddle-mobile.a
/src/ios_io/ 下的
PaddleMobile.h
```
拖入工程
#### oc 接口
接口如下:
```
/*
创建对象
*/
- (instancetype)init;
/*
load 模型, 开辟内存
*/
- (BOOL)load:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath;
/*
进行预测, means 和 scale 为训练模型时的预处理参数, 如训练时没有做这些预处理则直接使用 predict
*/
- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim means:(NSArray<NSNumber *> *)means scale:(float)scale;
/*
进行预测
*/
- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim;
/*
清理内存
*/
- (void)clear;
```
## GPU
需要: xcode、cocoapods
```
# 在 paddle-mobile 目录下:
cd metal
pod install
open paddle-mobile.xcworkspace
```
metal/README.md
浏览文件 @ e80f61c5
## Paddle-Mobile
This folder is used to develop metal version for ios gpu
需要: xcode、 cocoapods
```
pod install
open paddle-mobile.xcworkspace
```
Demo 所需依赖的模型可在[这里](http://mms-graph.bj.bcebos.com/paddle-mobile%2FmodelsAndImages.zip)下载
tools/python/modeltools/.gitignore 0 → 100644
浏览文件 @ e80f61c5
# Created by .ignore support plugin (hsz.mobi)
### Python template
# Byte-compiled / optimized / DLL files
__pycache__/
*.py[cod]
*$py.class
# C extensions
*.so
# Distribution / packaging
.Python
build/
develop-eggs/
dist/
downloads/
eggs/
.eggs/
lib/
lib64/
parts/
sdist/
var/
wheels/
*.egg-info/
.installed.cfg
*.egg
MANIFEST
# PyInstaller
# Usually these files are written by a python script from a template
# before PyInstaller builds the exe, so as to inject date/other infos into it.
*.manifest
*.spec
# Installer logs
pip-log.txt
pip-delete-this-directory.txt
# Unit test / coverage reports
htmlcov/
.tox/
.coverage
.coverage.*
.cache
nosetests.xml
coverage.xml
*.cover
.hypothesis/
.pytest_cache/
# Translations
*.mo
*.pot
# Django stuff:
*.log
local_settings.py
db.sqlite3
# Flask stuff:
instance/
.webassets-cache
# Scrapy stuff:
.scrapy
# Sphinx documentation
docs/_build/
# PyBuilder
target/
# Jupyter Notebook
.ipynb_checkpoints
# pyenv
.python-version
# celery beat schedule file
celerybeat-schedule
# SageMath parsed files
*.sage.py
# Environments
.env
.venv
env/
venv/
ENV/
env.bak/
venv.bak/
# Spyder project settings
.spyderproject
.spyproject
# Rope project settings
.ropeproject
# mkdocs documentation
/site
# mypy
.mypy_cache/
/yolo/datas/
/mobilenet/datas/
python/tools/mdl2fluid/op_types.py → tools/python/modeltools/core/op_types.py
浏览文件 @ e80f61c5
......@@ -5,22 +5,28 @@ layer_mdl_conv = 'ConvolutionLayer'
layer_mdl_deepwise_conv = 'DepthwiseConvolutionLayer'
layer_mdl_relu = 'ReluLayer'
layer_mdl_pointwise_add = 'PointwiseConvolutionLayer'
layer_mdl_pooling = 'PoolingLayer'
layer_mdl_softmax = 'SoftmaxLayer'
# fluid ops
op_fluid_fusion_conv_add = 'fusion_conv_add'
op_fluid_relu = 'relu'
op_fluid_pooling = 'pool2d'
op_fluid_softmax = 'softmax'
# dict mdk layer --- fluid op
mdl2fluid_op_layer_dict = {
layer_mdl_conv: op_fluid_fusion_conv_add,
layer_mdl_deepwise_conv: op_fluid_fusion_conv_add,
layer_mdl_relu: op_fluid_relu,
layer_mdl_pointwise_add: op_fluid_fusion_conv_add
layer_mdl_pointwise_add: op_fluid_fusion_conv_add,
layer_mdl_pooling: op_fluid_pooling,
layer_mdl_softmax: op_fluid_softmax
}
mdl_outputs_key = "outputs"
mdl_inputs_key = "inputs"
mdl_weight_key = "weights"
mdl_weight_key = "weight"
mdl_attrs_key = "params"
# dict of mdl-input _out param to fluid input out attrs
......@@ -39,13 +45,30 @@ fusion_conv_add_dict = {
relu_dict = {
mdl_inputs_key: 'X',
mdl_outputs_key: 'Out',
mdl_weight_key: ()
# mdl_weight_key: ()
}
pool2d_dict = {
mdl_inputs_key: 'X',
mdl_outputs_key: 'Out',
# mdl_weight_key: (),
mdl_attrs_key: ('pooling_type', 'global_pooling')
}
softmax_dict = {
mdl_inputs_key: 'X',
mdl_outputs_key: 'Out',
mdl_weight_key: (),
mdl_attrs_key: ()
}
# mdl layers --- fluid ops
op_io_dict = {
'fusion_conv_add': fusion_conv_add_dict,
'relu': relu_dict
'relu': relu_dict,
'pool2d': pool2d_dict,
'softmax': softmax_dict
}
# fluid attr key --- mdl params key
......@@ -60,64 +83,3 @@ fluid_attrs_type_dict = {
'strides': 6,
'groups': 6
}
# '': "bias_term", 是不是要add 目前 yolo的模型都是 bias_term = 1
# attrs {
# name: "axis"
# type: INT
# i: 1
# }
# attrs_name = {
# 'name': "workspace_size_MB",
# 'type': 'INT',
# 'i': '4096'
# }
# attrs
# {
# name: "data_format"
# type: STRING
# s: "AnyLayout"
# }
# attrs
# {
# name: "use_mkldnn"
# type: BOOLEAN
# b: false
# }
# attrs
# {
# name: "use_cudnn"
# type: BOOLEAN
# b: true
# }
# attrs
# {
# name: "dilations"
# type: INTS
# ints: 1
# ints: 1
# }
# attrs
# {
# name: "groups"
# type: INT
# i: 1
# }
# attrs
# {
# name: "paddings"
# type: INTS
# ints: 0
# ints: 0
# }
# attrs
# {
# name: "strides"
# type: INTS
# ints: 1
# ints: 1
# }
tools/python/modeltools/mobilenet/converter_mobilenet.py 0 → 100644
浏览文件 @ e80f61c5
import json
import os
from core import framework_pb2 as framework_pb2, op_types as types
from mobilenet.swicher import Swichter
import shutil
def load_mdl(mdl_json_path):
# print('mdl json path : ' + mdl_json_path)
with open(mdl_json_path, 'r') as f:
return json.load(f)
class Converter:
'convert mdlmodel to fluidmodel'
def __init__(self, base_dir, mdl_json_path):
self.mdl_json_path = base_dir + mdl_json_path
self.base_dir = base_dir
print mdl_json_path
self.mdl_json = load_mdl(self.mdl_json_path)
self.program_desc = framework_pb2.ProgramDesc()
self.weight_list_ = []
self.deepwise_weight_list_ = []
# print(json_dick)
# layers = (json_dick['layer'])
# for layer in layers:
# print(layer)
def convert(self):
print 'convert begin.....'
# add block_desc
block_desc = self.program_desc.blocks.add()
block_desc.idx = 0
block_desc.parent_idx = -1
self.package_ops(block_desc)
self.package_vars(block_desc)
print 'blocks: '
print self.program_desc.blocks
print 'convert end.....'
desc_serialize_to_string = self.program_desc.SerializeToString()
outputmodel_ = self.base_dir + 'datas/target/outputmodel/'
if os.path.exists(outputmodel_):
shutil.rmtree(outputmodel_)
os.makedirs(outputmodel_, 0777)
# todo copy weight files
# if os.path.exists(outputmodel_):
# shutil.rmtree(outputmodel_)
# shutil.copytree('yolo/datas/multiobjects/float32s_nchw_with_head/', 'mobilenet/datas/target/outputmodel/')
f = open(outputmodel_ + "__model__", "wb")
f.write(desc_serialize_to_string)
f.close()
def package_ops(self, block_desc):
self.add_op_feed(block_desc)
# add ops with layer
if 'layer' in self.mdl_json:
layers_ = self.mdl_json['layer']
for layer in layers_:
desc_ops_add = block_desc.ops.add()
# print layer
# for i in layer:
# print i
if 'name' in layer:
l_name = layer['name']
if 'type' in layer:
self.package_ops_type(desc_ops_add, layer)
if 'weight' in layer:
self.package_ops_weight2inputs(desc_ops_add, layer)
if 'output' in layer:
self.package_ops_outputs(desc_ops_add, layer)
if 'input' in layer:
self.package_ops_inputs(desc_ops_add, layer)
self.package_ops_attrs(desc_ops_add, layer)
self.add_op_fetch(block_desc)
def add_op_feed(self, block_desc):
desc_ops_add = block_desc.ops.add()
inputs_add = desc_ops_add.inputs.add()
inputs_add.parameter = 'X'
inputs_add.arguments.append('feed')
desc_ops_add.type = 'feed'
outputs_add = desc_ops_add.outputs.add()
outputs_add.parameter = 'Out'
outputs_add.arguments.append('data')
attrs_add = desc_ops_add.attrs.add()
attrs_add.name = 'col'
# boolean
attrs_add.type = 0
attrs_add.i = 0
def add_op_fetch(self, block_desc):
desc_ops_add = block_desc.ops.add()
inputs_add = desc_ops_add.inputs.add()
inputs_add.parameter = 'X'
inputs_add.arguments.append('conv_pred_87')
desc_ops_add.type = 'fetch'
outputs_add = desc_ops_add.outputs.add()
outputs_add.parameter = 'Out'
outputs_add.arguments.append('fetch')
attrs_add = desc_ops_add.attrs.add()
attrs_add.name = 'col'
# boolean
attrs_add.type = 0
attrs_add.i = 0
@staticmethod
def package_ops_attrs(desc_ops_add, layer):
# print l_params
# print desc_ops_add.type
if desc_ops_add.type == types.op_fluid_fusion_conv_add:
Converter.pack_fusion_conv_add_attr(desc_ops_add, layer)
elif desc_ops_add.type == types.op_fluid_relu:
# fusion_conv_add : attrs
attrs_add = desc_ops_add.attrs.add()
attrs_add.name = 'use_mkldnn'
# boolean
attrs_add.type = 6
attrs_add.b = 0
@staticmethod
def pack_fusion_conv_add_attr(desc_ops_add, layer):
# fusion_conv_add : attrs
attrs_add = desc_ops_add.attrs.add()
attrs_add.name = 'workspace_size_MB'
# 0-->INT
attrs_add.type = 0
attrs_add.i = 4096
attrs_add = desc_ops_add.attrs.add()
attrs_add.name = 'data_format'
# 2-->STRING
attrs_add.type = 2
attrs_add.s = 'AnyLayout'
attrs_add = desc_ops_add.attrs.add()
attrs_add.name = 'use_mkldnn'
# boolean
attrs_add.type = 6
attrs_add.b = 0
attrs_add = desc_ops_add.attrs.add()
attrs_add.name = 'use_cudnn'
# boolean
attrs_add.type = 6
attrs_add.b = 1
attrs_add = desc_ops_add.attrs.add()
attrs_add.name = 'dilations'
# ints
attrs_add.type = 3
attrs_add.ints.append(1)
attrs_add.ints.append(1)
attrs_add = desc_ops_add.attrs.add()
attrs_add.name = 'axis'
# int
attrs_add.type = 0
attrs_add.i = 1
if 'param' in layer:
l_params = layer['param']
attrs_add = desc_ops_add.attrs.add()
attrs_add.name = 'paddings'
# ints
attrs_add.type = 3
attrs_add.ints.append(l_params[types.fusion_conv_add_attrs_dict.get('paddings')])
attrs_add.ints.append(l_params[types.fusion_conv_add_attrs_dict.get('paddings')])
attrs_add = desc_ops_add.attrs.add()
attrs_add.name = 'strides'
# ints
attrs_add.type = 3
attrs_add.ints.append(l_params[types.fusion_conv_add_attrs_dict.get('strides')])
attrs_add.ints.append(l_params[types.fusion_conv_add_attrs_dict.get('strides')])
attrs_add = desc_ops_add.attrs.add()
attrs_add.name = 'groups'
# int
attrs_add.type = 0
attrs_add.i = l_params[types.fusion_conv_add_attrs_dict.get('groups')]
# attrs_add.i = 1
#
# op_attrs_tupl = types.op_io_dict.get(desc_ops_add.type) \
# .get(types.mdl_attrs_key)
#
#
#
#
# # group stride padding
# print '----------------------'
# for i, val in enumerate(op_attrs_tupl):
# attrs_add = desc_ops_add.attrs.add()
# attr_name = op_attrs_tupl[i]
# print attr_name
# attrs_add.name = attr_name
# attrs_add.type = types.fluid_attrs_type_dict.get(attr_name)
# attrs_add.
# print l_params[types.fusion_conv_add_attrs_dict.get(attr_name)]
# for p in l_params:
# attrs_add = desc_ops_add.attrs.add()
@staticmethod
def package_ops_inputs(desc_ops_add, layer):
l_inputs = layer['input']
for i in l_inputs:
inputs_add = desc_ops_add.inputs.add()
# print i
inputs_add.parameter = types.op_io_dict.get(desc_ops_add.type).get(types.mdl_inputs_key)
inputs_add.arguments.append(i)
@staticmethod
def package_ops_outputs(desc_ops_add, layer):
l_outputs = layer['output']
for o in l_outputs:
# print o
outputs_add = desc_ops_add.outputs.add()
dict = types.op_io_dict.get(desc_ops_add.type)
print 'desc_ops_add.type: ' + desc_ops_add.type
print dict
outputs_add.parameter = dict.get(types.mdl_outputs_key)
outputs_add.arguments.append(o)
def package_ops_weight2inputs(self, desc_ops_add, layer):
l_weights = layer['weight']
for w in l_weights:
self.weight_list_.append(w)
if layer['type'] == types.layer_mdl_deepwise_conv:
# print l_weights[0]
self.deepwise_weight_list_.append(l_weights[0])
op_weight_tup = types.op_io_dict.get(desc_ops_add.type).get(types.mdl_weight_key)
if op_weight_tup is not None:
# print len(op_weight_tup)
for i, val in enumerate(op_weight_tup):
# print i
# print val
inputs_add = desc_ops_add.inputs.add()
inputs_add.parameter = op_weight_tup[i]
inputs_add.arguments.append(l_weights[i])
# for w in l_weights:
# inputs_add = desc_ops_add.inputs.add()
# # print w
# inputs_add.parameter = op_weight_tup[0]
# inputs_add.arguments.append(w)
@staticmethod
def package_ops_type(desc_ops_add, layer):
l_type = layer['type']
# print l_type
# print mdl2fluid_op_layer_dict.get(l_type)
desc_ops_add.type = types.mdl2fluid_op_layer_dict.get(l_type)
def package_vars(self, block_desc):
vars_add = block_desc.vars.add()
vars_add.name = 'feed'
vars_add.type.type = 9 # 9 is FEED_MINIBATCH
vars_add.persistable = 1
# fetch
vars_add = block_desc.vars.add()
vars_add.name = 'fetch'
vars_add.type.type = 10 # 10 is fetch list
vars_add.persistable = 1
json_matrix_ = self.mdl_json['matrix']
# print json_matrix_
for j in json_matrix_:
vars_add = block_desc.vars.add()
vars_add.name = j
vars_add.type.type = 7 # 7 is lodtensor
# print j
tensor = vars_add.type.lod_tensor.tensor
tensor.data_type = 5 # 5 is FP32
# print json_matrix_
dims_of_matrix = json_matrix_.get(j)
# dims_size = len(dims_of_matrix)
# print dims_size
# if dims_size == 4:
# tensor.dims.append(dims_of_matrix[0]) # N
# tensor.dims.append(dims_of_matrix[3]) # C
# tensor.dims.append(dims_of_matrix[1]) # H
# tensor.dims.append(dims_of_matrix[2]) # W
# else:
# issues in mdl model filter swich n and c
if j in self.deepwise_weight_list_ and len(dims_of_matrix) == 4:
print j
tensor.dims.append(dims_of_matrix[1])
tensor.dims.append(dims_of_matrix[0])
tensor.dims.append(dims_of_matrix[2])
tensor.dims.append(dims_of_matrix[3])
print tensor.dims
else:
for dims in dims_of_matrix:
# print dims
tensor.dims.append(dims)
if j in self.weight_list_:
vars_add.persistable = 1
dims_size = len(dims_of_matrix)
# print dims_size
# if dims_size == 4:
# # convert weight from nhwc to nchw
# Swichter().nhwc2nchw_one_slice_add_head(
# 'yolo/datas/multiobjects/float32s_nhwc/' + j + '.bin',
# 'yolo/datas/multiobjects/float32s_nchw_with_head/' + j,
# 'yolo/datas/multiobjects/float32s_nchw/' + j + '.tmp',
# dims_of_matrix[0],
# dims_of_matrix[1],
# dims_of_matrix[2],
# dims_of_matrix[3]
# )
# else:
# Swichter().copy_add_head(
# 'yolo/datas/multiobjects/float32s_nhwc/' + j + '.bin',
# 'yolo/datas/multiobjects/float32s_nchw_with_head/' + j,
# 'yolo/datas/multiobjects/float32s_nchw/' + j + '.tmp'
# )
else:
vars_add.persistable = 0
mdl_path = "datas/sourcemodels/cls231_0802/mobileNetModel.json"
base_dir = "/Users/xiebaiyuan/PaddleProject/paddle-mobile/tools/python/modeltools/mobilenet/"
converter = Converter(base_dir, mdl_path)
converter.convert()
python/tools/mdl2fluid/swicher.py → tools/python/modeltools/mobilenet/swicher.py
浏览文件 @ e80f61c5
......@@ -58,7 +58,7 @@ class Swichter:
to_file = open(to_file_name, "wb")
tmp = tmp_file.read()
head = self.read_head('/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/yolo/conv1_biases')
head = self.read_head('yolo/datas/yolo/conv1_biases')
to_file.write(head)
to_file.write(tmp)
tmp_file.close()
......@@ -77,7 +77,7 @@ class Swichter:
to_file = open(to_file_name, "wb")
# tmp_file = open(tmp_file_name, "wb")
head = self.read_head('/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/yolo/conv1_biases')
head = self.read_head('yolo/datas/yolo/conv1_biases')
to_file.write(head)
to_file.write(from_file.read())
from_file.close()
......@@ -96,7 +96,7 @@ class Swichter:
to_file = open(to_file_name, "wb")
# tmp_file = open(tmp_file_name, "wb")
head = self.read_head('/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/yolo/conv1_biases')
head = self.read_head('yolo/datas/yolo/conv1_biases')
to_file.write(head)
to_file.write(read)
from_file.close()
......@@ -104,12 +104,12 @@ class Swichter:
pass
# Swichter().nhwc2nchw_one_slice_add_head(
# '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nhwc/conv1_0.bin',
# '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw_with_head/conv1_0',
# '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw/.tmp',
# '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/modeltools/multiobjects/float32s_nhwc/conv1_0.bin',
# '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/modeltools/multiobjects/float32s_nchw_with_head/conv1_0',
# '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/modeltools/multiobjects/float32s_nchw/.tmp',
# 32,
# 3, 3, 3)
# Swichter().read_head('/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/yolo/conv1_biases')
# Swichter().read_head('/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/modeltools/yolo/conv1_biases')
# Swichter().copy_add_head('datas/model.0.0.weight', 'datas/conv1_0', '')
python/tools/mdl2fluid/loader.py → tools/python/modeltools/tools/loader.py
浏览文件 @ e80f61c5
import datetime
import json
import os
import google.protobuf as pbg
import framework_pb2 as framework_pb2
def loadmdl(json_path):
......
python/tools/mdl2fluid/model_reader.py → tools/python/modeltools/tools/model_reader.py
浏览文件 @ e80f61c5
import os
import framework_pb2 as framework_pb2
from core import framework_pb2 as framework_pb2
def read_model(model_path):
......@@ -16,7 +16,7 @@ def read_model(model_path):
# print desc.blocks
except IOError:
print ": File not found. Creating a new file."
print ": File not found."
def get_file_size(file_path):
......@@ -26,5 +26,5 @@ def get_file_size(file_path):
return round(fsize, 2)
path = "newyolo/__model__"
path = '/Users/xiebaiyuan/PaddleProject/paddle-mobile/tools/python/modeltools/mobilenet/datas/sourcemodels/mobilenet_example/mobilenet/__model__'
read_model(path)
python/tools/mdl2fluid/mdl2fluid.py → tools/python/modeltools/yolo/mdl2fluid.py
浏览文件 @ e80f61c5
import json
import os
import framework_pb2 as framework_pb2
import op_types as types
from swicher import Swichter
from core import framework_pb2 as framework_pb2, op_types as types
from yolo.swicher import Swichter
import shutil
......@@ -40,10 +38,10 @@ class Converter:
print self.program_desc.blocks
print 'convert end.....'
desc_serialize_to_string = self.program_desc.SerializeToString()
shutil.rmtree('newyolo/')
shutil.copytree('multiobjects/float32s_nchw_with_head', 'newyolo/')
shutil.rmtree('yolo/datas/newyolo/')
shutil.copytree('yolo/datas/multiobjects/float32s_nchw_with_head/', 'yolo/datas/newyolo/')
f = open("newyolo/__model__", "wb")
f = open("yolo/datas/newyolo/__model__", "wb")
f.write(desc_serialize_to_string)
f.close()
......@@ -312,9 +310,9 @@ class Converter:
if dims_size == 4:
# convert weight from nhwc to nchw
Swichter().nhwc2nchw_one_slice_add_head(
'/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nhwc/' + j + '.bin',
'/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw_with_head/' + j,
'/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw/' + j + '.tmp',
'yolo/datas/multiobjects/float32s_nhwc/' + j + '.bin',
'yolo/datas/multiobjects/float32s_nchw_with_head/' + j,
'yolo/datas/multiobjects/float32s_nchw/' + j + '.tmp',
dims_of_matrix[0],
dims_of_matrix[1],
dims_of_matrix[2],
......@@ -322,14 +320,14 @@ class Converter:
)
else:
Swichter().copy_add_head(
'/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nhwc/' + j + '.bin',
'/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw_with_head/' + j,
'/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw/' + j + '.tmp'
'yolo/datas/multiobjects/float32s_nhwc/' + j + '.bin',
'yolo/datas/multiobjects/float32s_nchw_with_head/' + j,
'yolo/datas/multiobjects/float32s_nchw/' + j + '.tmp'
)
else:
vars_add.persistable = 0
mdl_path = "/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/YOLO_Universal.json"
mdl_path = "yolo/datas/multiobjects/YOLO_Universal.json"
converter = Converter(mdl_path)
converter.convert()
tools/python/modeltools/yolo/swicher.py 0 → 100644
浏览文件 @ e80f61c5
from array import array
class Swichter:
def __init__(self):
pass
def nhwc2nchw_one_slice(self, from_file_name, to_file_name, batch, channel, height, width):
from_file = open(from_file_name, "rb")
to_file = open(to_file_name, "wb")
float_array = array("f")
float_array.fromfile(from_file, width * height * batch * channel)
float_write_array = array("f")
for b in range(batch):
for c in range(channel):
for h in range(height):
for w in range(width):
float_value = float_array[b * channel * width * height
+ channel * (h * width + w) + c]
float_write_array.append(float_value)
float_write_array.tofile(to_file)
from_file.close()
to_file.close()
def copy(self, from_file_name, to_file_name):
from_file = open(from_file_name, "rb")
to_file = open(to_file_name, "wb")
to_file.write(from_file.read())
from_file.close()
to_file.close()
def nhwc2nchw_one_slice_add_head(self, from_file_name, to_file_name, tmp_file_name, batch, channel, height, width):
from_file = open(from_file_name, "rb")
tmp_file = open(tmp_file_name, "wb+")
float_array = array("f")
float_array.fromfile(from_file, width * height * batch * channel)
float_write_array = array("f")
for b in range(batch):
for c in range(channel):
for h in range(height):
for w in range(width):
float_value = float_array[b * channel * width * height
+ channel * (h * width + w) + c]
float_write_array.append(float_value)
float_write_array.tofile(tmp_file)
tmp_file.close()
from_file.close()
tmp_file = open(tmp_file_name, "rb")
to_file = open(to_file_name, "wb")
tmp = tmp_file.read()
head = self.read_head('yolo/datas/yolo/conv1_biases')
to_file.write(head)
to_file.write(tmp)
tmp_file.close()
to_file.close()
def read_head(self, head_file):
from_file = open(head_file, "rb")
read = from_file.read(24)
# print read
from_file.close()
# print read
return read
def copy_add_head(self, from_file_name, to_file_name, tmp_file_name):
from_file = open(from_file_name, "rb")
to_file = open(to_file_name, "wb")
# tmp_file = open(tmp_file_name, "wb")
head = self.read_head('yolo/datas/yolo/conv1_biases')
to_file.write(head)
to_file.write(from_file.read())
from_file.close()
to_file.close()
pass
def copy_padding_add_head(self, from_file_name, to_file_name, tmp_file_name, padding):
print'padding = %d' % padding
from_file = open(from_file_name, "rb")
# print len(from_file.read())
from_file.seek(padding, 0)
read = from_file.read()
print len(read)
to_file = open(to_file_name, "wb")
# tmp_file = open(tmp_file_name, "wb")
head = self.read_head('yolo/datas/yolo/conv1_biases')
to_file.write(head)
to_file.write(read)
from_file.close()
to_file.close()
pass
# Swichter().nhwc2nchw_one_slice_add_head(
# '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/modeltools/multiobjects/float32s_nhwc/conv1_0.bin',
# '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/modeltools/multiobjects/float32s_nchw_with_head/conv1_0',
# '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/modeltools/multiobjects/float32s_nchw/.tmp',
# 32,
# 3, 3, 3)
# Swichter().read_head('/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/modeltools/yolo/conv1_biases')
# Swichter().copy_add_head('datas/model.0.0.weight', 'datas/conv1_0', '')
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