C++ Train Demo¶

Introduction¶

我们都知道，Paddle Lite 可以做移动端预测，事实上 Paddle Lite 支持在移动端做模型训练。本文给出使用Paddle Lite 做训练的例子，这一例子对应的任务是“波士顿房价预测”，又称作 “fit-a-line”。

你可以通过 book 库中的文档和源码进一步了解“波士顿房价预测”这一任务的定义及其建模过程，其使用线性回归（ Linear Regression ）模型做建模。本文主要介绍如何将其迁移至 Paddle Lite 进行训练。

Requirements¶

一部安卓手机，用于运行训练程序
装了 Paddle (version >= 1.7.0) 的 python

Quick start¶

Step1 build paddle lite¶

请按照 paddle lite 官方文档的教程编译 full_publish 的 paddle lite lib。以 Linux 上编译为例，其具体的命令为：

## 配置环境
wget -c https://mms-res.cdn.bcebos.com/cmake-3.10.3-Linux-x86_64.tar.gz --no-check-certificate
tar xzf cmake-3.10.3-Linux-x86_64.tar.gz
export PATH=${PWD}'/cmake-3.10.3-Linux-x86_64/bin':$PATH

wget https://dl.google.com/android/repository/android-ndk-r17c-linux-x86_64.zip
unzip android-ndk-r17c-linux-x86_64.zip
export NDK_ROOT=/opt/android-ndk-r17c

## 编译
git clone https://github.com/PaddlePaddle/Paddle-Lite.git
cd Paddle-Lite
./lite/tools/build.sh \
  --arm_os=android \
  --arm_abi=armv7 \
  --build_extra=ON \
  --arm_lang=gcc \
  --android_stl=c++_static \
  --build_train=ON full_publish

产物:

Paddle-Lite/build.lite.android.armv7.gcc/inference_lite_lib.android.armv7/cxx/lib/libpaddle_full_api_shared.so

Step2 编译 lr_trainer¶

cd Paddle-Lite/lite/demo/cxx/train_demo/cplus_train/
sh run_build.sh /path/to/your/Paddle-Lite/build.lite.android.armv7.gcc/ /path/to/your/android-ndk-r17c

产物:

bin/
`-- demo_trainer

Step3 download model and run it!¶

在你的笔记本电脑上，用 usb 连接到手机，开启开发者模式，在任意目录下执行：

local_path=/data/local/tmp/linear_regression
adb shell "mkdir "${local_path}

# download model and push to mobile
wget http://paddle-tar.bj.bcebos.com/paddle-lite/lite_lr_model.tar.gz
tar -zxvf lite_lr_model.tar.gz
adb push lite_lr_model/housing.data ${local_path}
adb push lite_lr_model/model_dir ${local_path}

# push lib and executable file to moblie
adb push libpaddle_full_api_shared.so ${local_path}
adb push demo_trainer ${local_path}
adb shell chmod +x ${local_path}/demo_trainer

# run it!
adb shell "export LD_LIBRARY_PATH="${local_path}" && export LIBRARY_PATH="${local_path}" && cd "${local_path}" && ./demo_trainer true"

期望结果：

sample 0: Loss: 564.317
sample 1: Loss: 463.9
sample 2: Loss: 1197.54
sample 3: Loss: 1093.83
sample 4: Loss: 1282.76
sample 5: Loss: 792.097
sample 6: Loss: 491.776
sample 7: Loss: 698.496
sample 8: Loss: 248.445
sample 9: Loss: 325.135

与 Paddle 训练结果做校对¶

前 10 个 Loss 值¶

为了验证 Paddle 与 Paddle Lite 的一致性，我们控制模型参数一致、数据一致、batch size = 1 的情况下，训练10 个 batch，记录了二者的 loss 值。

python + paddle 命令:

  fluid train.py --num_steps=10 --batch_size=1

python + paddle 结果:

Train cost, Step 0, Cost 564.317017
Train cost, Step 1, Cost 463.900238
Train cost, Step 2, Cost 1197.537354
Train cost, Step 3, Cost 1093.833008
Train cost, Step 4, Cost 1282.760254
Train cost, Step 5, Cost 792.097351
Train cost, Step 6, Cost 491.775848
Train cost, Step 7, Cost 698.496033
Train cost, Step 8, Cost 248.444885
Train cost, Step 9, Cost 325.135132

c++ 与 paddle lite命令：

./demo_trainer true

c++ 与 Paddle Lite 结果：

sample 0: Loss: 564.317
sample 1: Loss: 463.9
sample 2: Loss: 1197.54
sample 3: Loss: 1093.83
sample 4: Loss: 1282.76
sample 5: Loss: 792.097
sample 6: Loss: 491.776
sample 7: Loss: 698.496
sample 8: Loss: 248.445
sample 9: Loss: 325.135

Loss 曲线¶

控制训练时的 batch size 为 20，每个 epoch 对训练数据做全局 shuffle，训练 100 个 epoch 后，Paddle 和 Paddle Lite 的 loss 曲线对比如下。

lr_loss

如果想复现上述效果，paddle+python 的运行命令为：

git clone https://github.com/PaddlePaddle/book.git
cd book/01.fit_a_line
python train.py

Paddle Lite + c++ 的运行命令为：

./demo_trainer false