Skip to content

TensorFlow_API_Install_Guide.md

Latest commit

 

History

History
73 lines (62 loc) · 2.09 KB

TensorFlow_API_Install_Guide.md

File metadata and controls

73 lines (62 loc) · 2.09 KB

TensorFlow Detector API

create by afterloe
version is 1.0.3
MIT License

TensorFlow的Detector API的安装经过折腾半个月，现在基本跑通，现在写下这篇博文。首先TensorFLow Detector API的版本还在迭代，现在TensorFLow 2.x出现有一段时间， 然后该API迁移工作并没有完成，所有不能使用TensorFlow 2.x的代码，具体安装步骤如下：

前提准备

硬件

x86 Linux Ubuntu 18.04.1
GeForce GTX 1050 Ti (4G Memory)

软件

NVIDIA Driver:  430
CUDA Version:   10.0
cuDNN Version:  7_7.6.5.32 for CUDA 10.0
tensorflow-gpu: 1.15.2
protobuf:       3.0.0
python:         3.7.5

TensorFlow api使用的版本为1.15.2，准备完毕后进行安装阶段，另外CUDA与cuDNN的安装和配置请参考CUDA_Install_Guide.md

cuda-repo-ubuntu1804-10-0-local-10.0.130-410.48_1.0-1_amd64.deb | libcudnn7_7.6.5.32-1+cuda10.0_amd64.deb

框架安装与配置

下载源码

cd ~
git clone https://github.com/tensorflow/models.git
mv models tensorflow_api

安装框架必要依赖

sudo apt install protobuf-compiler python-pil python-lxml python-tk -y
pip3 install Cython contextlib2 jupyter matplotlib pillow lxml tensorboard

下载coco API 并安装

cd ~
git clone https://github.com/cocodataset/cocoapi.git
cd cocoapi/PythonAPI
vim Makefile  # 将 python修改为 python3
make
ln -s pycocotools ~/tensorflow_api/research

编译Protobuf文件

cd ~/tensorflow_api/research
protoc object_detection/protos/*.proto --python_out=.

linux 下安装protoc可以从官网下载二进制可执行文件，放到/usr/local/bin， 执行protoc --version显示libprotoc 3.0.0表示成功，在执行编译

添加库到环境变量

vim ~/.profile
export PYTHONPATH=$PYTHONPATH:/home/afterloe/tensorflow_api/research:tensorflow_api/research/slim

测试安装

cd ~/tensorflow_api/research
python3 object_detection/builders/model_builder_test.py