- 全体構成
- nvidia 提供の cuda10-cudnn7 コンテナをベースとする
- tensorRT は nvidia のサイトからダウンロードする必要がある
- その他は Dockerfile 内で完結
- ビルド方法
$ git clone https://github.com/kagyuu/Build_Tensorflow_GPU.git
$ cd Build_Tensorflow_GPU
$ docker-compose build
$ docker-compose run nvidia /bin/bash
途中でビルドが止まってしまうので、コンテナにログインして再度ビルドを実行するととりあえずビルドは終わる
$ docker-compose run nivid /bin/bash
# cd ~/tensorflow
# bazel build //tensorflow/tools/pip_package:build_pip_package --config=opt
...
INFO: Elapsed time: 1843.411s, Critical Path: 168.01s
INFO: 3364 processes: 3364 local.
INFO: Build completed successfully, 4486 total actions
# ./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/
ホストからパッケージを取得する
$ docker ps
... ... ... NAMES
... ... ... tensorflow_nvidia_run_67e67b06722a
$ docker cp tensorflow_nvidia_run_67e67b06722a:/tmp/tensorflow_pkg/tensorflow-1.13.0rc0-cp36-cp36m-linux_x86_64.whl ./
version: '2.3'
services:
nvidia:
build: .
image: nvidia/my-tensorflow-gpu
runtime: nvidia
command: bash
FROM nvidia/cuda:10.0-cudnn7-devel-ubuntu18.04
WORKDIR /root/
RUN apt-get update && apt-get -y dist-upgrade
RUN apt-get -y install curl wget vim htop git swig build-essential
# Don't ask timezone when you install expect that depends on tzdata.
ENV DEBIAN_FRONTEND=noninteractive
RUN apt-get -y install expect
# Install Bazel
# https://docs.bazel.build/versions/master/install-ubuntu.html
RUN apt-get -y install openjdk-8-jdk
RUN echo "deb [arch=amd64] http://storage.googleapis.com/bazel-apt stable jdk1.8" | tee /etc/apt/sources.list.d/bazel.list
RUN curl https://bazel.build/bazel-release.pub.gpg | apt-key add -
RUN apt-get update && apt-get -y install bazel
# Download Tensorflow
RUN git clone https://github.com/tensorflow/tensorflow
RUN cd tensorflow && git checkout r1.13
# cf. Build from Source
# https://www.tensorflow.org/install/source#install_tensorflow_python_dependencies
RUN apt-get -y install python3 python3-dev python3-pip
RUN /usr/bin/pip3 install -U six numpy wheel mock h5py
RUN /usr/bin/pip3 install -U keras_applications==1.0.6 --no-deps
RUN /usr/bin/pip3 install -U keras_preprocessing==1.0.5 --no-deps
# TensorRT
# https://developer.nvidia.com/tensorrt
COPY nv-tensorrt-repo-ubuntu1804-cuda10.0-trt5.0.2.6-ga-20181009_1-1_amd64.deb /tmp
RUN dpkg -i /tmp/nv-tensorrt-repo-ubuntu1804-cuda10.0-trt5.0.2.6-ga-20181009_1-1_amd64.deb
RUN apt-key add /var/nv-tensorrt-repo-cuda10.0-trt5.0.2.6-ga-20181009/7fa2af80.pub
RUN apt-get update && apt-get -y install tensorrt python3-libnvinfer-dev uff-converter-tf
# Build Tensorflow
WORKDIR /root/tensorflow/
ENV TMP=/tmp
RUN expect -c " \
set timeout 10; \
spawn ./configure; \
expect \"Please specify the location of python.\"; send \"/usr/bin/python3\r\"; \
expect \"Please input the desired Python library path to use\"; send \"\r\"; \
expect \"XLA JIT support?\"; send \"n\r\"; \
expect \"OpenCL SYCL support?\"; send \"n\r\"; \
expect \"ROCm support?\"; send \"n\r\"; \
expect \"CUDA support?\"; send \"y\r\"; \
expect \"CUDA SDK version\"; send \"\r\"; \
expect \"where CUDA 10.0 toolkit\"; send \"\r\"; \
expect \"the cuDNN version\"; send \"\r\"; \
expect \"where cuDNN 7 library\"; send \"\r\"; \
expect \"TensorRT support?\"; send \"y\r\"; \
expect \"where TensorRT is installed\"; send \"\r\"; \
expect \"installed NCCL version\"; send \"\r\"; \
expect \"Cuda compute capabilities\"; send \"7.5\r\"; \
expect \"use clang as CUDA compiler\"; send \"\r\"; \
expect \"which gcc should be used\"; send \"\r\"; \
expect \"MPI support?\"; send \"\r\"; \
expect \"optimization flags\"; send \"\r\"; \
expect \"Android builds?\"; send \"\r\"; \
expect eof; \
"
# The following command will be failed by missing cuda librry.
# You shoud retry in login shell, "docker-compose run nvidia /bin/bash".
RUN bazel build //tensorflow/tools/pip_package:build_pip_package --config=opt; exit 0
Deep Learning
![[PukiWiki]](image/pukiwiki.png "[PukiWiki]")
