Pensieve Books, Programming

여러가지 DL 라이브러리를 도커(Docker)에서 세팅해보기

03 Jun 2019

Build MXNet 1.4.1 with C++ API on Ubuntu 16.04 with CUDA 10.1 & cuDNN7.5.1

1. Run a New Docker Container
docker run --runtime=nvidia -dit --name mxnet nvidia/cuda:10.1-cudnn7-devel-ubuntu16.04 bash
docker exec -it mxnet bash
2. Install Dependencies via apt
apt update && apt install -y vim git cmake cpio libopenblas-dev liblapack-dev graphviz
3. Install OpenCV 3.4.6
cd
git clone https://github.com/opencv/opencv
cd opencv
git checkout 3.4.6
mkdir build && cd build
cmake -DBUILD_TESTS=Off -DBUILD_PERF_TESTS=Off ../
make -j$(nproc) install
4. Install Python
# We'll just install Anaconda, downloadable from
# https://www.anaconda.com/distribution/#download-section
# with following options "Linux, Python3.7, Anaconda3-2019.03-Linux-x86_64.sh"
docker cp Anaconda3-2019.03-Linux-x86_64.sh mxnet:/root
bash Anaconda3-2019.03-Linux-x86_64.sh
5. Build MXNet
cd
git clone https://github.com/apache/incubator-mxnet mxnet
cd mxnet
git checkout 1.4.1
git submodule update --recursive --init
mkdir build && cd build
cmake -DUSE_CUDA=1 -DUSE_CUDNN=1 -DUSE_MKLDNN=1 -DUSE_CPP_PACKAGE=1 ../
make -j$(nproc)
Reference
https://mxnet.apache.org/versions/master/install/build_from_source.html
https://mxnet.apache.org/versions/master/install/c_plus_plus.html

Convert CNTK Model to ONNX Model

Going to build CNTK 2.7 CPU only version. CNTK 2.7 supports ONNX 1.4.1.

Download Anaconda
https://www.anaconda.com/distribution/#download-section
Linux, Python3.7, Anaconda3-2019.03-Linux-x86_64.sh
Run a New Docker Container and Install Anaconda
docker run -dit --name cntk2onnx ubuntu:16.04 bash
docker cp Anaconda3-2019.03-Linux-x86_64.sh cntk2onnx:/root
docker exec -it cntk2onnx bash

apt update && apt install -y openmpi-bin bzip2
cd && bash Anaconda3-2019.03-Linux-x86_64.sh

exit
docker exec -it cntk2onnx bash

Create Anaconda Environment and Install CNTK via pip

conda create --name cntk-py36 python=3.6
conda activate cntk-py36
pip install https://cntk.ai/PythonWheel/CPU-Only/cntk-2.7-cp36-cp36m-linux_x86_64.whl

To confirm that CNTK 2.7 is properly installed, run:

LD_LIBRARY_PATH=/root/anaconda3/envs/cntk-py36/lib python -c "import cntk; print(cntk.__version__)"

Convert a CNTK Model to ONNX Model

LD_LIBRARY_PATH=/root/anaconda3/envs/cntk-py36/lib python -c "import cntk; cntk.Function.load('model0.model').save('model0.onnx',format=cntk.ModelFormat.ONNX)"

Install “tensorflow-onnx”

Download Anaconda
https://www.anaconda.com/distribution/#download-section
Linux, Python3.7, Anaconda3-2019.03-Linux-x86_64.sh
Run a New Docker Container and Install Anaconda
docker run -dit --name onnx2tf ubuntu:16.04 bash
docker cp Anaconda3-2019.03-Linux-x86_64.sh onnx2tf:/root
docker exec -it onnx2tf bash

apt update && apt install -y bzip2
cd && bash Anaconda3-2019.03-Linux-x86_64.sh

exit
docker exec -it onnx2tf bash
Install TensorFlow via pip
pip install tensorflow==1.13.1
Install ONNX
apt install -y git cmake g++ protobuf-compiler libprotoc-dev
git clone https://github.com/onnx/onnx.git
cd onnx
git checkout v1.4.1
git submodule update --init --recursive
export ONNX_ML=1
python setup.py install
Install tensorflow-onnx
cd
git clone https://github.com/onnx/tensorflow-onnx
cd tensorflow-onnx
git checkout v1.4.1
python setup.py install

Show ChangeLog