Building OpenPose (Pytorch_Realtime_Multi-Person_Pose_Estimation) environment on Docker : training part

過去の記事

PC上に学習環境を直接構築する方法は以下を参照
修正中

前提条件

ローカルマシンにGPUの環境が構築されている前提とします．
（nvidia-smiでGPUが確認できる）

Dockerのインストールは，以下のサイトなどを参考に．．．
https://qiita.com/ttsubo/items/c97173e1f04db3cbaeda

Dockerの基本操作

以下参照：
https://qiita.com/19503/private/a6f0d43c8cfc44748f62

Dockerコンテナを用いた環境構築

実行マシンのGPUやCUDAの世代の影響を避けるため，実行環境のGPUを直接操作するのではなく，以下の様にDockerコンテナを利用した方法を推奨します．

# 1.ホストPC上の操作
# setup COCO data(coco2017)
COCO_DIR=/home/your_work_dir/your_coco_data_dir
mkdir $COCO_DIR/images
cd $COCO_DIR/images
wget http://images.cocodataset.org/zips/train2017.zip
wget http://images.cocodataset.org/zips/val2017.zip
cd $COCO_DIR
wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip

cd /home/your_work_dir
git clone https://github.com/pytorch/pytorch.git -b v0.2.0
cd pytorch
# Dockerfileを，後述するlast-one-Docerfileの様に編集（名前はDockerfileから変更しない事）
docker build -t last-one:v0.01 -f ./Dockerfile .
# --ipc="host"はホストPCとコンテナ間でメモリを共有するオプション，これが無いとメモリ不足により実行中に落ちる（以下参考）
# https://discuss.pytorch.org/t/unable-to-write-to-file-torch-18692-1954506624/9990
docker run --gpus all --ipc="host" -v /home/your_work_dir/your_coco_data_dir:/home/dat -itd --name last-one_cont last-one:v0.01
docker exec -it last-one_cont /bin/bash

# 2.Dockerコンテナでの操作

pip intall pycocotools
cd /home/dev/Pytorch_Realtime_Multi-Person_Pose_Estimation/preprocessing
COCO_DIR=/home/dat

apt-get update
pip install numpy==1.16.6 opencv-python==3.1.0.0 easydict==1.9

python generate_json_mask.py --ann_path $COCO_DIR/annotations/person_keypoints_train2017.json --json_path out/train2017_json.json --mask_dir out/train2017_maskdir --filelist_path out/train2017_filelst.txt --masklist_path out/train2017_masklst.txt
python generate_json_mask.py --ann_path $COCO_DIR/annotations/person_keypoints_val2017.json --json_path out/val2017_json.json --mask_dir out/val2017_maskdir --filelist_path out/val2017_filelst.txt --masklist_path out/val2017_masklst.txt

# if you encountered "import _tkinter # If this fails your Python may not be configured for Tk" error, you must install tk-dev as below
# apt-get update
# apt-get install tk-dev

# 関連ファイルのパスをviエディタで編集（以下例）
vi train2017_filelst.txt
%s/^/\/home\/dat\/images\/train2017\//g
vi val2017_filelst.txt
%s/^/\/home\/dat\/images\/val2017\//g
vi train2017_masklst.txt
%s/^/\/home\/dev\/Pytorch_Realtime_Multi-Person_Pose_Estimation\/preprocessing\//g
vi val2017_masklst.txt
%s/^/\/home\/dev\/Pytorch_Realtime_Multi-Person_Pose_Estimation\/preprocessing\//g

# --gpuオプションは自身のGPU環境によって変更が必要，かつ過去記事にしたがってtrain/train_pose.py l255のCUDA_VISIBLE_DEVICESを修正
TRAIN_DIR=/home/dev/Pytorch_Realtime_Multi-Person_Pose_Estimation/preprocessing/out
python train_pose.py --gpu 0 1 2 --train_dir $TRAIN_DIR/train2017_filelst.txt $TRAIN_DIR/train2017_masklst.txt $TRAIN_DIR/train2017_json.json --val_dir $TRAIN_DIR/val2017_filelst.txt $TRAIN_DIR/val2017_masklst.txt $TRAIN_DIR/val2017_json.json --config config.yml > $LOG

last-one-Docerfile

FROM nvidia/cuda:8.0-cudnn6-devel-ubuntu16.04

RUN echo "deb http://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1604/x86_64 /" > /etc/apt/sources.list.d/nvidia-ml.list

RUN apt-get update && apt-get install -y --no-install-recommends \
         build-essential \
         cmake \
         git \
         wget \
         vim \
         ca-certificates \
         libjpeg-dev \
         libpng-dev &&\
     rm -rf /var/lib/apt/lists/*

RUN wget https://repo.anaconda.com/archive/Anaconda2-2.4.0-Linux-x86_64.sh -O ./anaconda.sh && \
     chmod +x ./anaconda.sh && \
     ./anaconda.sh -b -p /opt/conda && \
     rm ./anaconda.sh && \
     /opt/conda/bin/conda install conda-build && \
     /opt/conda/bin/conda create -y --name pytorch-py27 python=2.7.13 numpy pyyaml scipy ipython mkl&& \
     /opt/conda/bin/conda clean -ya

ENV PATH /opt/conda/envs/pytorch-py27/bin:$PATH
RUN conda install --name pytorch-py27 -c soumith magma-cuda80
# This must be done before pip so that requirements.txt is available
WORKDIR /opt/pytorch
COPY . .

# install torch
RUN TORCH_CUDA_ARCH_LIST="3.5 5.2 6.0 6.1+PTX" TORCH_NVCC_FLAGS="-Xfatbin -compress-all" \
    CMAKE_PREFIX_PATH="$(dirname $(which conda))/../" \
    pip install -v .

# install torchvision
RUN git clone https://github.com/pytorch/vision.git -b v0.2.0
WORKDIR /opt/pytorch/vision
RUN pip install -v .

# clone OpenPose code
WORKDIR /home/dev
RUN git clone https://github.com/last-one/Pytorch_Realtime_Multi-Person_Pose_Estimation.git

# setup coco tools
RUN git clone https://github.com/cocodataset/cocoapi.git

RUN pip install pandas

WORKDIR /workspace