【PaddlePaddle Hackathon】31. 为 Paddle Inference 添加新的前端语言

[chenyanlann]

PR types

New features

PR changes

为 Paddle Inference 添加 Java Apis

Describe

Task: #35977

Paddle Inference java API

Paddle Inference java API 基于 capi 和 jni 实现，需要您提前准备好C预测库。

安装

1.下载C预测库

您可以选择直接下载paddle_inference_c预测库，或通过源码编译的方式安装，源码编译方式参考官网文档，注意这里cmake编译时打开-DON_INFER=ON,在编译目录下得到paddle_inference_c_install_dir。

2.准备预测部署模型

下载 resnet50 模型后解压，得到 Paddle Combined 形式的模型。

wget https://paddle-inference-dist.bj.bcebos.com/Paddle-Inference-Demo/resnet50.tgz
tar zxf resnet50.tgz

#### 获得 resnet50 目录结构如下
resnet50/
├── inference.pdmodel
├── inference.pdiparams
└── inference.pdiparams.info

3.准备预测执行目录

git clone github.com/paddlepaddle/paddle/paddle/fluid/inference/javaapi

3. 编译动态链接库和jar包

在javaapi目录下执行

./build_gpu.sh {c预测库目录} {jni头文件目录} {jni系统头文件目录}

以笔者的目录结构为例
./build_gpu.sh /root/paddle_c/paddle_inference_c /usr/lib/jvm/java-8-openjdk-amd64/include /usr/lib/jvm/java-8-openjdk-amd64/include/linux

执行完成后，会在当前目录下生成JavaInference.jar和libpaddle_inference.so

5.运行单测，验证

在javaapi目录下执行

./test.sh {c预测库目录} {.pdmodel文件目录} {.pdiparams文件目录}

以笔者的目录结构为例
./test.sh "/root/paddle_c/paddle_inference_c_2.2/paddle_inference_c"  "/root/paddle_c/resnet50/inference.pdmodel" "/root/paddle_c/resnet50/inference.pdiparams"

在Java中使用Paddle预测

首先创建预测配置

Config config = new Config();
config.setCppModel(model_file, params_file);

创建predictor

Predictor predictor = Predictor.createPaddlePredictor(config);

获取输入Tensor

String inNames = predictor.getInputNameById(0);
Tensor inHandle = predictor.getInputHandle(inNames);

设置输入数据(假设只有一个输入)

inHandle.Reshape(4, new int[]{1, 3, 224, 224});
float[] inData = new float[1*3*224*224];
inHandle.CopyFromCpu(inData);

运行预测

predictor.Run();

获取输出Tensor

String outNames = predictor.getOutputNameById(0);
Tensor outHandle = predictor.getOutputHandle(outNames);
float[] outData = new float[outHandle.GetSize()];
outHandle.CopyToCpu(outData);

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ chenyanlann
❌ chenyanlan

---

chenyanlan seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}