Ubuntu16.04从源码安装DensePose

Installing DensePose

DensePose官网 http://densepose.org/

DensePose代码 https://github.com/facebookresearch/Densepose

DensePose安装系统要求：有NVIDIA GPU的Linux系统，满足这一条件便可按照下面的步骤安装。

安装Anaconda

官网下载Linux版的Anaconda3，运行bash Anaconda3-5.2.0-Linux-x86_64.sh。

安装caffe2(pytorch)

创建caffe2的conda环境

conda create -n densepose python=2.7

进入新创建的环境

source activate densepose 

安装一些需要的库

(densepose)$ conda install -y future gflags leveldb mkl mkl-include numpy opencv protobuf
(densepose)$ conda install flask graphviz hypothesis jupyter matplotlib pydot pyyaml requests scikit-image scipy setuptools tornado

下载pytorch及其依赖的第三方库

# git clone --recursive https://github.com/pytorch/pytorch.git 
(densepose)$ git clone -b v0.4.1 --recursive https://github.com/pytorch/pytorch.git # 下载v0.4.1分支代码
(densepose)$ cd pytorch
(densepose)$ git submodule update --init

编译并安装

(densepose)$ rm -rf build && mkdir build && cd build

(densepose)$ cmake -DCMAKE_PREFIX_PATH=~/anaconda3/envs/densepose -DCMAKE_INSTALL_PREFIX=~/anaconda3/envs/densepose -DUSE_NATIVE_ARCH=ON ..

(densepose)$ make -j32 install

测试是否安装成功

# To check if Caffe2 build was successful
(densepose)$ python2 -c 'from caffe2.python import core' 2>/dev/null && echo "Success" || echo "Failure"
# 输出Success即安装成功,否则失败

# To check if Caffe2 GPU build was successful
(densepose)$ python2 -c 'from caffe2.python import workspace; print(workspace.NumCudaDevices())'
# 输出8或者其他,表示显卡个数

安装cocoapi

# COCOAPI=/path/to/clone/cocoapi
(densepose)$ git clone https://github.com/cocodataset/cocoapi.git $COCOAPI
(densepose)$ cd $COCOAPI/PythonAPI
# Install into global site-packages
(densepose)$ make install
# Alternatively, if you do not have permissions or prefer
# not to install the COCO API into global site-packages
(densepose)$ python2 setup.py install --user

安装densepose

安装densepose

(densepose)$ mkdir $DENSEPOSE_DIR && cd $DENSEPOSE_DIR
(densepose)$ git clone https://github.com/facebookresearch/densepose
(densepose)$ cd densepose
(densepose)$ pip install -r requirements.txt  # install Python dependencies

(densepose)$ make  # set up python modules
(densepose)$ python2 detectron/tests/test_spatial_narrow_as_op.py  # check that Detectron tests pass

(densepose)$ make ops  # Build the custom operators library
(densepose)$ python2 detectron/tests/test_zero_even_op.py  # check that the custom operator tests pass

下载densepose数据:

(densepose)$ cd DensePoseData
(densepose)$ bash get_densepose_uv.sh

如果需要训练,下载DensePose-COCO数据集:

(densepose)$ bash get_DensePose_COCO.sh

如果评价(evaluation),还需下载其他需要的文件:

(densepose)$ bash get_eval_data.sh

Inference with Pretrained Models

首先从这里下载与DensePose_ResNet101_FPN_s1x-e2e.yaml相匹配的训练好的模型,并放在weights文件夹下,然后执行下面命令:

(densepose)$ python2 tools/infer_simple.py \
    --cfg configs/DensePose_ResNet101_FPN_s1x-e2e.yaml \
    --output-dir DensePoseData/infer_out/ \
    --image-ext jpg \
    --wts  weights/DensePose_ResNet101_FPN_s1x-e2e.pkl \
    DensePoseData/demo_data/demo_im.jpg

Ref

• Caffe2与Detectron的安装笔记与Mask-RCNN的测试 - CSDN博客
• ubuntu16.04 caffe2(pytorch)安装
• how-to-let-cmake-find-cuda
• DensePose-INSTALL
• DensePose-GETTING_STARTED