cudnn

I created the exact same network with native and backend tensorflow but after many hours of testing using number of different parameters, still couldn't figure out why keras outperforms the native tensorflow and produces better(slightly but better) results. Does Keras implement a different weight initializer method? or performs different weight decay approach other than tf.train.inverse_time_decay? P.s. the score difference is always like Keras with Tensorflow: ~0.9850 - 0.9885 - ~45 sec. avg. training time for 1 epoch Tensorflow Native ~0.9780 - 0.9830 - ~23 sec. My environment is: Python 3.5

1.降级内核 1.1 安装新内核 由于最新ubuntu系统的内核等级过高，安装低版本驱动会失败，所以要将系统内核版本降低。 安装4.4.0-98版本内核。 sudo apt-get install linux -headers-4.4.0-98- generic linux-image-4.4.0-98-generic linux-image-extra-4.4.0-98-generic 1.2更新grub 编辑 /etc/default/grub sudo vi /etc/default/grub 使用方向键移动到 GRUB_HIDDEN_TIMEOUT=0 按 i 键 进入编辑模式 在前面加入# 注释该行，开机时会出现高级选项 #GRUB_HIDDEN_TIMEOUT=0 Esc键退出编辑 :w 保存 :q 退出 更新引导 sudo update-grub 如果当前驱动版本过高，重启后可能会出现重复登录的情况。可以提前把要安装的显卡驱动下载好，以备安装。 重启计算机，进入高级选项，选择新安装的内核。 sudo reboot 如果重复登录，CTRL+ALT+F1进入命令行模式登录。 1.3卸载旧内核 查看当前内核 uname -r 查看所有内核 $ dpkg --get-selections| grep linux 卸载多余内核 sudo apt remove linux

I was trying to use TensorFlow with GPU and got the following error: I tensorflow/core/common_runtime/gpu/gpu_device.cc:838] Creating TensorFlow device (/gpu:0) -> (device: 0, name: Tesla K20m, pci bus id: 0000:02:00.0) E tensorflow/stream_executor/cuda/cuda_dnn.cc:347] Loaded runtime CuDNN library: 5005 (compatibility version 5000) but source was compiled with 5103 (compatibility version 5100). If using a binary install, upgrade your CuDNN library to match. If building from sources, make sure the library loaded at runtime matches a compatible version specified during compile configuration. F

在配置Tensorflow环境过程中，由于版本和驱动不对应，导致调试时一直报错，请后续者在配置前，务必参看CUDA和CUDNN与Tensorflow的对应关系后再动手，避免不必要的时间浪费。本文系统环境是Ubuntu18.04,显卡是GeForce GTX960M。链接2在较新的硬件安装时用，可大辐节省时间。 本文参考链接1： https://blog.csdn.net/BigData_Mining/article/details/99670642 本文参考链接2： https://www.cnblogs.com/ManWingloeng/p/11489122.html 一、安装N卡驱动： 1，查看N卡信息。 1 nvidia-smi 2，删除旧版N卡驱动。 1 sudo apt-get remove nvidia-* 2 sudo apt-get autoremove 3，获取系统推荐驱动（括号带recommended的），并安装该驱动，我的是440。 1 ubuntu-drivers devices 1 sudo apt-get install nvidia-settings nvidia-driver-440 nvidia-prime 4，重启。 1 sudo reboot 二、安装CUDA10.1及CUDNN7.6.1 1，查看适合本机的CUDA及CUDNN版本。 1

This question already has answers here : Which TensorFlow and CUDA version combinations are compatible? (5 answers) Closed 12 months ago . Currently, I need to install gpu version tensorflow on a machine that existing an older version CUDA and cuDNN, and my question is how can I get it done without re-installing CUDA and cuDNN since other package (see pytorch) need it. Check CUDA version with below command according to this post . $ nvcc --version Check cuDNN version with below command according to this post . $ cat /usr/local/cuda/include/cudnn.h | grep CUDNN_MAJOR -A 2 Choose tensorflow-gpu

一、环境准备 1.安装CUDA、cudnn 首先确定显卡是否能GPU加速，查看显卡版本下载对应的驱动 CUDA和cudnn必须在宿主机中安装而不是在镜像中 之后下载CUDA、与cudnn 具体下载方式网上很多 安装时一定要先确定你要安装的tensorflow-gpu版本 例：我安装的是tensorflow-gpu 1.8.0 所以对应-CUDA 9.0 与cudnn 7.0 之前因为安装了9.1又安回9.0的时候出了很多问题 所以最好一次到位 cuda 版本 cat /usr/local/cuda/version.txt cudnn 版本 cat /usr/local/cuda/include/cudnn.h | grep CUDNN_MAJOR -A 2 2.docker安装 1.docker版本用的是docker-ce_18.06.0_ce_3-0 有时候docker对版本还是挺敏感的 2.就是我踩的大坑 nvidia-docker 当时就觉得docker和nvidia-docker是一个东西所以就只按了docker 只有安装了nvidia-docker 才能把宿主机的cuda和cudnn映射到docker镜像中 当时报的错是在镜像中x86_64-linux-gnu中少了很多cuda相关共享库 3.Tensorflow-gpu-1.8.0 1.docker pull