TensorFlow的安装与CNN测试

0.说明

在Google开源该框架之后便使用真实K40m卡测试，由于生产环境是CentOS6.6的操作系统，但是该框架需要在Python2.7环境下执行，CentOS6.6下折腾了一天没搞定，后来换成CentOS7，顺利跑通

1.系统环境

• python >=2.7
• numpy >=1.9
• gcc >=4.8.2
• cuda 7.0
• java >=1.8
• cudnn 6.5 v2

2.安装部署

 

1. #安装依赖，kernel-devel是为了安装cuda
2. yum -y install gcc python-devel kernel-devel
3. #安装pip
4. wget --no-check-certificate https://github.com/pypa/pip/archive/1.5.5.tar.gz
5. tar zvxf 1.5.5.tar.gz
6. cd pip-1.5.5/
7. python setup.py install
8. #安装tensorflow,此时确保服务器可以联网，会自动下载安装numpy和six
9. pip install http://dlp.iflytek.com/soft/tensorflow-0.5.0-cp27-none-linux_x86_64.whl

3.测试用例CNN

• 下载训练数据集

 

1. wget http://dlp.iflytek.com/soft/cifar-10-binary.tar.gz
2. tar -zxvf -C /tmp/cifar10_data

• 执行脚本(默认cpu) 
  cd /root/tensorflow-master/tensorflow/models/image/cifar10 
  python cifar10_train.py 
  

• 采用gpu执行 
  python cifar10_multi_gpu_train.py --num_gpus=4 
  

4.脚本相关说明

 

1. #查看帮助
2. python cifar10_train.py --help
3. --batch_size BATCH_SIZE #一批数据的图片数量，默认是包含128个examples
4. Number of images to process in a batch.
5. --data_dir DATA_DIR #训练数据集目录，默认是/tmp/cifar10_data
6. Path to the CIFAR-10 data directory.
7. --train_dir TRAIN_DIR #训练目录
8. Directory where to write event logs and checkpoint.
9. --max_steps MAX_STEPS #最大步数，默认是1000000
10. Number of batches to run.
11. --log_device_placement LOG_DEVICE_PLACEMENT
12. Whether to log device placement.
13. --nolog_device_placement

 

---

Tutorials and Machine Learning Examples — TensorFlow 
http://tensorflow.org/tutorials/deep_cnn/index.md