未验证 提交 f0940bb9 编写于 作者: H huzhiqiang 提交者: GitHub

[doc]modify release doc, add explanation for `build_extra` (#3005)

上级 b92e9f9a
# 支持OP列表
## Ops
## Ops (共计158个算子)
### Basic Operators (默认编译的算子)
- affine_channel
- anchor_generator
- arg_max
- assign
- assign_value
- attention_padding_mask
- axpy
- batch_norm
- beam_search
- beam_search_decode
- bilinear_interp
- box_clip
- box_coder
- calib
- calib_once
- cast
- collect_fpn_proposals
- concat
- conditional_block
- conv2d
- conv2d_transpose
- crop
- decode_bboxes
- density_prior_box
- depthwise_conv2d
- distribute_fpn_proposals
- dropout
- elementwise_add
- elementwise_div
- elementwise_max
- elementwise_mul
- elementwise_sub
- equal
- exp
- expand
- fake_channel_wise_dequantize_max_abs
......@@ -56,29 +42,86 @@
- fusion_elementwise_max_activation
- fusion_elementwise_mul_activation
- fusion_elementwise_sub_activation
- gather
- gelu
- grid_sampler
- hard_sigmoid
- instance_norm
- io_copy
- io_copy_once
- layout
- leaky_relu
- log
- matmul
- mean
- mul
- multiclass_nms
- nearest_interp
- pad2d
- pool2d
- prelu
- prior_box
- range
- reduce_mean
- relu
- relu6
- relu_clipped
- reshape
- reshape2
- rsqrt
- scale
- search_fc
- sequence_topk_avg_pooling
- shuffle_channel
- sigmoid
- slice
- softmax
- softsign
- split
- sqrt
- square
- squeeze
- squeeze2
- stack
- subgraph
- swish
- tanh
- transpose
- transpose2
- unsqueeze
- unsqueeze2
- yolo_box
### Extra Operators (打开 `--build_extra=ON`开关才会编译)
- anchor_generator
- assign
- assign_value
- attention_padding_mask
- axpy
- beam_search
- beam_search_decode
- box_clip
- calib_once
- collect_fpn_proposals
- conditional_block
- crop
- decode_bboxes
- distribute_fpn_proposals
- equal
- gather
- generate_proposals
- greater_equal
- greater_than
- grid_sampler
- gru
- gru_unit
- hard_sigmoid
- im2sequence
- increment
- instance_norm
- io_copy
- io_copy_once
- is_empty
- layer_norm
- layout
- layout_once
- leaky_relu
- less_equal
- less_than
- lod_reset
- log
- logical_and
- logical_not
- logical_or
......@@ -87,37 +130,18 @@
- lookup_table_v2
- lrn
- match_matrix_tensor
- matmul
- mean
- merge_lod_tensor
- mul
- multiclass_nms
- nearest_interp
- negative
- norm
- not_equal
- pad2d
- pool2d
- power
- prelu
- prior_box
- range
- read_from_array
- reduce_max
- reduce_mean
- reduce_prod
- reduce_sum
- relu
- relu6
- relu_clipped
- reshape
- reshape2
- roi_align
- rsqrt
- scale
- search_aligned_mat_mul
- search_attention_padding_mask
- search_fc
- search_grnn
- search_group_padding
- search_seq_arithmetic
......@@ -133,33 +157,15 @@
- sequence_reshape
- sequence_reverse
- sequence_softmax
- sequence_topk_avg_pooling
- shape
- shuffle_channel
- sigmoid
- slice
- softmax
- softsign
- split
- split_lod_tensor
- sqrt
- square
- squeeze
- squeeze2
- stack
- subgraph
- swish
- tanh
- top_k
- transpose
- transpose2
- uniform_random
- unsqueeze
- unsqueeze2
- var_conv_2d
- while
- write_to_array
- yolo_box
## Kernels
......
# 预测库说明
# `build_extra`参数说明:
Paddle-Lite的编译结果为预测库文件(包括静态库和动态库),具体编译过程参考[源码编译](./source_compile)
Lite预测库分为**基础预测库****全量预测库(with_extra)**:基础预测库只包含基础CV算子(OP),体积较小;全量预测库包含所有Lite算子,体积较大,支持模型较多
Lite预测库分为**基础预测库****全量预测库**:基础预测库只打包了基础模型需要的基础算子,预测库体积较小;全量预测库打包了所有的Lite算子,可以支持更多的模型,但是预测库的体积也更大。 编译时由编译选项 `build_extra`(默认为OFF)控制,`--build_extra=OFF`时编译基础预测库,`--build_extra=ON`时编译全量的预测库
编译时由编译选项 `build_extra`(默认为OFF)控制,`--build_extra=OFF`时编译**基础预测库**`--build_extra=ON`时编译**全量预测库**
## 基础预测库
## 基础预测库( [基础OP列表](../advanced_user_guides/support_operation_list.html#basic-operators) )
### 编译方法
编译时设置`--build_extra=OFF` (默认值) 或不指定即可编译出基础预测库。例如:
```
./lite/tools/build.sh --arm_os=android --arm_abi=armv8 --arm_lang=gcc --android_stl=c++_static tiny_publish
```
### 基础预测库支持的功能
(1)支持基础CV模型
### 支持功能
2)支持基础的in8量化模型
1)87个[基础OP](../advanced_user_guides/support_operation_list.html#basic-operators) (2)9个基础模型 (3)3个in8量化模型
(3)支持[benchmark测试](../benchmark/benchmark)
### 支持的模型
### 基础预测库支持的基础模型:
1. fluid基础模型(paddle model 提供的基础模型9个)
1. fluid基础模型(来源:[paddle-models](https://github.com/PaddlePaddle/models)
```
mobileNetV1 mnasnet yolov3 ssd_mobilenetv1 shufflenet_v2
mobileNetV2 resnet50 unet squeezenet_v11
mobilenetV1 mnasnet yolov3 ssd_mobilenetv1 shufflenet_v2
mobilenetV2 resnet50 unet squeezenet_v11
```
2. int8量化模型模型
2. int8量化模型
```
mobilenet_v1 mobilenet_v2 resnet50
```
### 特点
轻量级预测库,体积更小,支持常用的基础模型。
## 全量预测库
轻量级预测库,体积更小,支持常用模型。
### 编译方法
编译时设置`--build_extra=ON` 即可编译出全量预测库。例如:
编译时设置`--build_extra=OFF` (默认值) 编译出基础预测库。例如:
```
./lite/tools/build.sh --arm_os=android --arm_abi=armv8 --arm_lang=gcc --android_stl=c++_static --build_extra=ON tiny_publish
./lite/tools/build.sh --arm_os=android --arm_abi=armv8 --arm_lang=gcc --android_stl=c++_static tiny_publish
```
### 全量预测库功能
(1) 基础预测库所有功能
(2)支持所有Paddle-Lite中注册的所有算子
## 全量预测库( [OP列表](../advanced_user_guides/support_operation_list.html#op) )
### 支持功能
Paddle-Lite中的全量算子( [基础OP](../advanced_user_guides/support_operation_list.html#basic-operators) + [Extra OP](../advanced_user_guides/support_operation_list.html#extra-operators-build-extra-on)
### 特点
支持更多的硬件平台和算子,可以支持更多模型但体量更大。
包含更多算子、支持更多模型,但体量更大。
### 编译方法
设置`--build_extra=ON` 可编译出全量预测库。例如:
```
./lite/tools/build.sh --arm_os=android --arm_abi=armv8 --arm_lang=gcc --android_stl=c++_static --build_extra=ON tiny_publish
```
......@@ -3,10 +3,17 @@
## 编译版本介绍
- ARM_Version=`armv7/armv8` 编译时选择的arm版本,v7或者v8
- build_extra=`ON/OFF` 是否编译全量算子,OFF时只编译CV相关基础算子
- arm_stl=`c++_static/c++_shared` 选择动态或者静态链接NDK STL库
- target: `tiny_publish/full_publish` 是否编译full_api预测库,`tiny_publish`时只编译light_api预测库
- ARM_Version=`armv7/armv8` arm版本,可选择armv7或者armv8
- arm_os=`android\ios\ios64\armlinux` 安装平台,支持的arm端移动平台包括 `ios\ios64``armlinux``android`
- arm_lang=`gcc/clang` 源码编译时的编译器,默认为`gcc`编译器
- arm_stl=`c++_static/c++_shared` Lite预测库链接STL库的方式,支持静态或动态链接
- build_extra=`ON/OFF` 是否编译全量OP,OFF时只编译CV相关基础OP,[参数详情](./library.html)
- `tiny_publish/full_publish` 编译模式,`tiny_publish`编译移动端部署库、`full_publish`编译部署库的同时编译第三方依赖库
## Android
......@@ -35,16 +42,16 @@
|ARM Version|arm_os|with_extra|下载|
|:-------:|:-----:|:-----:|:-----:|
|armv7|Ios|OFF|[release/v2.3](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.3.0/inference_lite_lib.ios.armv7.tar.gz)|
|armv7|ios|OFF|[release/v2.3](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.3.0/inference_lite_lib.ios.armv7.tar.gz)|
|armv7|ios|ON|[release/v2.3](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.3.0/inference_lite_lib.ios.armv7.with_extra.tar.gz)|
|armv8|Ios64|OFF|[release/v2.3](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.3.0/inference_lite_lib.ios64.armv8.tar.gz)|
|armv8|Ios64|ON|[release/v2.3](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.3.0/inference_lite_lib.ios64.armv8.with_extra.tar.gz)|
|armv8|ios64|OFF|[release/v2.3](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.3.0/inference_lite_lib.ios64.armv8.tar.gz)|
|armv8|ios64|ON|[release/v2.3](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.3.0/inference_lite_lib.ios64.armv8.with_extra.tar.gz)|
## opt 工具
| 运行系统 | 下载 |
| :---------: | :--------------: |
| :---------: | :--------------: |
| Linux | [release/v2.3](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.3.0/opt) |
| MacOs | [release/v2.3](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.3.0/opt_mac) |
......
Markdown is supported
0% .
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