未验证 提交 256dad09 编写于 作者: F flame 提交者: GitHub

R language support (#22417) (#22468)

* R-language inference support
上级 20fa4f0d
FROM ubuntu:18.04
MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com>
WORKDIR /workspace
ENV PATH /opt/python3/bin:/root/.local/bin:$PATH
ENV LD_LIBRARY_PATH $LD_LIBRARY_PATH:/opt/python3/lib
# Install Python
ADD https://mirrors.tuna.tsinghua.edu.cn/anaconda/miniconda/Miniconda3-4.7.12.1-Linux-x86_64.sh miniconda3.sh
RUN /bin/bash miniconda3.sh -b -p /opt/python3/ && \
rm -f miniconda3.sh
RUN mkdir -p ~/.pip && \
echo "[global]" >> ~/.pip/pip.conf && \
echo "trusted-host = mirrors.aliyun.com" >> ~/.pip/pip.conf && \
echo "index-url = https://mirrors.aliyun.com/pypi/simple" >> ~/.pip/pip.conf
RUN echo "channels:" >> ~/.condarc && \
echo " - conda-forge" >> ~/.condarc && \
echo " - https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/" >> ~/.condarc && \
echo " - https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/" >> ~/.condarc && \
echo " - defaults" >> ~/.condarc && \
echo "custom_channels:" >> ~/.condarc && \
echo " conda-forge: https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud" >> ~/.condarc && \
echo " msys2: https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud" >> ~/.condarc && \
echo " bioconda: https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud" >> ~/.condarc && \
echo " menpo: https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud" >> ~/.condarc && \
echo " pytorch: https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud" >> ~/.condarc && \
echo " simpleitk: https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud" >> ~/.condarc && \
echo "show_channel_urls: true" >> ~/.condarc && \
echo "channel_priority: strict" >> ~/.condarc
# Install R
RUN conda install -y r -c conda-forge
# Install PaddlePaddle
RUN /opt/python3/bin/python -m pip install \
https://paddle-wheel.bj.bcebos.com/0.0.0-cpu-mkl/paddlepaddle-0.0.0-cp37-cp37m-linux_x86_64.whl
# Install reticulate, R interface to Python
RUN Rscript -e 'install.packages("reticulate", repos="https://cran.rstudio.com")'
COPY example example
RUN cd example && \
curl -O https://paddle-inference-dist.cdn.bcebos.com/mobilenet-test-model-data.tar.gz && \
tar -zxvf mobilenet-test-model-data.tar.gz && rm mobilenet-test-model-data.tar.gz
# R support
English | [简体中文](./README_cn.md)
Use paddle in R.
## Install
### Use docker
Download [`Dockerfile`](./Dockerfile), run
``` bash
docker build -t paddle-rapi:latest .
```
### Local installation
First, make sure `Python` is installed, assuming that the path is `/opt/python3.7`.
``` bash
python -m pip install paddlepaddle # CPU version
python -m pip install paddlepaddle-gpu # GPU version
```
Install the R libraries needed to use paddle.
``` r
install.packages("reticulate") # call Python in R
install.packages("RcppCNPy") # use numpy.ndarray in R
```
## Use Paddle inference in R
First, load PaddlePaddle in R.
``` r
library(reticulate)
library(RcppCNPy)
use_python("/opt/python3.7/bin/python3.7")
paddle <- import("paddle.fluid.core")
```
Create an `AnalysisConfig`, which is the configuration of the paddle inference engine.
``` r
config <- paddle$AnalysisConfig("")
```
Set model path.
``` r
config$set_model("model/__model__", "model/__params__")
```
Use zero copy inference.
``` r
config$switch_use_feed_fetch_ops(FALSE)
config$switch_specify_input_names(TRUE)
```
Other configuration options and descriptions are as fallows.
``` r
config$enable_profile() # turn on inference profile
config$enable_use_gpu(gpu_memory_mb, gpu_id) # use GPU
config$disable_gpu() # disable GPU
config$gpu_device_id() # get GPU id
config$switch_ir_optim(TRUE) # turn on IR optimize(default is TRUE)
config$enable_tensorrt_engine(workspace_size,
max_batch_size,
min_subgraph_size,
paddle$AnalysisConfig$Precision$FLOAT32,
use_static,
use_calib_mode
) # use TensorRT
config$enable_mkldnn() # use MKLDNN
config$delete_pass(pass_name) # delete IR pass
```
Create inference engine.
``` r
predictor <- paddle$create_paddle_predictor(config)
```
Get input tensor(assume single input), and set input data
``` r
input_names <- predictor$get_input_names()
input_tensor <- predictor$get_input_tensor(input_names[1])
input_shape <- as.integer(c(1, 3, 300, 300)) # shape has integer type
input_data <- np_array(data, dtype="float32")$reshape(input_shape)
input_tensor$copy_from_cpu(input_data)
```
Run inference.
``` r
predictor$zero_copy_run()
```
Get output tensor(assume single output).
``` r
output_names <- predictor$get_output_names()
output_tensor <- predictor$get_output_tensor(output_names[1])
```
Parse output data, and convert to `numpy.ndarray`
``` r
output_data <- output_tensor$copy_to_cpu()
output_data <- np_array(output_data)
```
Click to see the full [R mobilenet example](./example/mobilenet.r) and the corresponding [Python mobilenet example](./example/mobilenet.py) the above. For more examples, see [R inference example](./example).
## Quick start
Download [Dockerfile](./Dockerfile) and [example](./example) to local directory, and build docker image
``` bash
docker build -t paddle-rapi:latest .
```
Create and enter container
``` bash
docker run --rm -it paddle-rapi:latest bash
```
Run the following command in th container
```
cd example
chmod +x mobilenet.r
./mobilenet.r
```
# R 语言支持
在 R 中使用 PaddlePaddle
## 环境安装
首先确保已安装Python,假设路径为`/opt/python3.7`
使用Python安装Paddle
``` bash
/opt/python3.7/bin/python3.7 -m pip install paddlepaddle # CPU
/opt/python3.7/bin/python3.7 -m pip install paddlepaddle-gpu # GPU
```
安装r运行paddle预测所需要的库
``` r
install.packages("reticulate") # 调用Paddle
```
## 在 R 中使用Paddle预测
首先在 R 中引入paddle预测环境
``` r
library(reticulate)
use_python("/opt/python3.7/bin/python")
paddle <- import("paddle.fluid.core")
```
创建一个AnalysisConfig,用于设置预测引擎的各选项
``` r
config <- paddle$AnalysisConfig("")
```
禁用feed和fetch OP,以使用 zero copy 预测
``` r
config$switch_use_feed_fetch_ops(FALSE)
config$switch_specify_input_names(TRUE)
```
设置模型路径有两种形式:
- `model` 目录中存在一个模型文件和多个参数文件
- `model` 目录中存在一个模型文件`__model__`和一个参数文件`__params__`
分别对应如下设置
``` r
config$set_model("model")
config$set_model("model/__model__", "model/__params__")
```
其他一些配置选项及说明如下
``` r
config$enable_profile() # 打开预测profile
config$enable_use_gpu(gpu_memory_mb, gpu_id) # 开启GPU预测
config$disable_gpu() # 禁用GPU
config$gpu_device_id() # 返回使用的GPU ID
config$switch_ir_optim(TRUE) # 开启IR优化(默认开启)
config$enable_tensorrt_engine(workspace_size,
max_batch_size,
min_subgraph_size,
paddle$AnalysisConfig$Precision$Float32,
use_static,
use_calib_mode
) # 开启TensorRT
config$enable_mkldnn() # 开启MKLDNN
config$disable_glog_info() # 禁用预测中的glog日志
config$delete_pass(pass_name) # 预测的时候删除指定的pass
```
创建预测引擎
``` r
predictor <- paddle$create_paddle_predictor(config)
```
获取输入tensor(为简单起见,此处假设只有一个输入),并设置输入tensor中的数据(注意需要使用np_array以传入numpy.ndarray类型的数据)
``` r
input_names <- predictor$get_input_names()
input_tensor <- predictor$get_input_tensor(input_names[1])
input_shape <- as.integer(c(1, 3, 300, 300)) # shape 为int类型
input_data <- np_array(data, dtype="float32")$reshape(input_shape)
input_tensor$copy_from_cpu(input_data)
```
运行预测引擎
``` r
predictor$zero_copy_run()
```
获取输出tensor(为简单起见,此处假设只有一个输出)
``` r
output_names <- predictor$get_output_names()
output_tensor <- predictor$get_output_tensor(output_names[1])
```
获取输出tensor中的数据,注意需要转为numpy.ndarray
``` r
output_data <- output_tensor$copy_to_cpu()
output_data <- np_array(output_data)
```
点击查看完整的[R预测示例](./example/mobilenet.r)及对应的[python预测示例](./example/mobilenet.py)
### 快速运行
[Dockerfile](./Dockerfile)[example](./example)下载到本地,使用以下命令构建docker镜像
``` bash
docker build -t paddle-rapi:latest .
```
启动一个容器
``` bash
docker run --rm -it paddle-rapi:latest bash
```
运行示例
``` bash
cd example && chmod +x mobilenet.r
./mobilenet.r
```
#!/usr/bin/env python3.7
# pylint: skip-file
import functools
import numpy as np
from paddle.fluid.core import AnalysisConfig
from paddle.fluid.core import AnalysisPredictor
from paddle.fluid.core import create_paddle_predictor
def main():
config = set_config()
predictor = create_paddle_predictor(config)
data, result = parse_data()
input_names = predictor.get_input_names()
input_tensor = predictor.get_input_tensor(input_names[0])
shape = (1, 3, 300, 300)
input_data = data[:-4].astype(np.float32).reshape(shape)
input_tensor.copy_from_cpu(input_data)
predictor.zero_copy_run()
output_names = predictor.get_output_names()
output_tensor = predictor.get_output_tensor(output_names[0])
output_data = output_tensor.copy_to_cpu()
def set_config():
config = AnalysisConfig("")
config.set_model("model/__model__", "model/__params__")
config.switch_use_feed_fetch_ops(False)
config.switch_specify_input_names(True)
config.enable_profile()
return config
def parse_data():
""" parse input and output data """
with open('data/data.txt', 'r') as fr:
data = np.array([float(_) for _ in fr.read().split()])
with open('data/result.txt', 'r') as fr:
result = np.array([float(_) for _ in fr.read().split()])
return (data, result)
if __name__ == "__main__":
main()
#!/usr/bin/env Rscript
library(reticulate) # call Python library
use_python("/opt/python3.7/bin/python")
np <- import("numpy")
paddle <- import("paddle.fluid.core")
set_config <- function() {
config <- paddle$AnalysisConfig("")
config$set_model("data/model/__model__", "data/model/__params__")
config$switch_use_feed_fetch_ops(FALSE)
config$switch_specify_input_names(TRUE)
config$enable_profile()
return(config)
}
zero_copy_run_mobilenet <- function() {
data <- np$loadtxt("data/data.txt")
data <- data[0:(length(data) - 4)]
result <- np$loadtxt("data/result.txt")
result <- result[0:(length(result) - 4)]
config <- set_config()
predictor <- paddle$create_paddle_predictor(config)
input_names <- predictor$get_input_names()
input_tensor <- predictor$get_input_tensor(input_names[1])
input_data <- np_array(data, dtype="float32")$reshape(as.integer(c(1, 3, 300, 300)))
input_tensor$copy_from_cpu(input_data)
predictor$zero_copy_run()
output_names <- predictor$get_output_names()
output_tensor <- predictor$get_output_tensor(output_names[1])
output_data <- output_tensor$copy_to_cpu()
output_data <- np_array(output_data)$reshape(as.integer(-1))
#all.equal(output_data, result)
}
if (!interactive()) {
zero_copy_run_mobilenet()
}
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