配置训练HRNet的环境

配置HRNet的环境

网络介绍

simple-HRNet是一个简化版的HRNet没有官方那么复杂的，也更好配置。
是基于官方代码，论文地址

网络配置

软硬件环境：
电脑一：
python 3.7.6
Ubuntu 18.04 LTS
CUDA:10.1
Pytorch:1.4.0
torchvision:0.5.0
Driver Version: 440.100
NVIDIA TITAN X (Pascal) 12G
电脑二：
python 3.6.2
Ubuntu 18.04 LTS
CUDA:9.0
Pytorch:1.1.0
torchvision:0.3.0
Driver Version: 435.21
NVIDIA Quadro P6000 24G

参考Linux下基于Pytorch框架用HRNet提取骨架特征点+解决CUDA和Torch版本不匹配问题
下载网络

git clone https://github.com/stefanopini/simple-HRNet.git
#安装环境文件
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple/ -r requirements.txt

requirements里面的torch版本是≥1.4.0，CUDA版本要和torch对应的，不要装错了
cuda版本查看 cat /usr/local/cuda/version.txt
pytorch版本，cuda版本，系统cuda版本查询和对应关系
用下面代码测试torch是否使用cuda。输出是True就安装正确了

import torch
torch.cuda.is_available()

文件结构

weights下要下载对应的预训练的权重
在./models/detectors下

git clone https://github.com/eriklindernoren/PyTorch-YOLOv3

把文件名改为yolo后，在yolo文件下

#安装环境文件
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple/ -r requirements.txt
cd weights/
#下载weight
bash download_weights.sh

回到simple-HRNet文件夹下

git submodule update --init --recursive

运行测试Class usage

import cv2
import matplotlib.pyplot as plt
from SimpleHRNet import SimpleHRNet

model = SimpleHRNet(48, 17, "./weights/pose_hrnet_w48_384x288.pth")
image = cv2.imread("./test.JPG", cv2.IMREAD_COLOR) #这里的文件路径和文件名根据自己要识别的图片来
joints = model.predict(image)
print(joints)

运行特征点提取测试

python scripts/extract-keypoints.py --filename video.mp4

会出现ffmpeg报错，或是cv2没有安装的报错
继续配置环境

apt-get -y --no-install-recommends update 
apt-get -y --no-install-recommends upgrade 
apt-get install -y --no-install-recommends build-essential cmake git libatlas-base-dev libprotobuf-dev apt-file
apt-get install -y --no-install-recommends libleveldb-dev libsnappy-dev libhdf5-serial-dev libboost-all-dev
apt-get install -y --no-install-recommends libgflags-dev libgoogle-glog-dev libviennacl-dev libcanberra-gtk-module
apt-get install -y --no-install-recommends libopencv-dev opencl-headers ocl-icd-opencl-dev protobuf-compiler liblmdb-dev
apt-get install -y --no-install-recommends pciutils python3-setuptools python3-dev python3-pip
apt-get update 
apt-file update
#我安装上opencv后不能使用，然后安装下面包就可以了
apt-file search libSM.so.6 
apt-get install libsm6 
apt-get -y --no-install-recommends update 
apt-get -y --no-install-recommends upgrade
#重新安装ffmpeg
pip uninstall ffmpeg-python
sudo add-apt-repository ppa:djcj/hybrid  
sudo apt-get update  
sudo apt-get install ffmpeg  
conda install ffmpeg-python # 我测试过用pip安装ffmepg,并不行。
pip install json_tricks

训练HRNet

安装cocoapi

git clone https://github.com/cocodataset/cocoapi.git 
cd cocoapi-master/PythonAPI
# 使用 make
make install
# 或是下面代码
python3 setup.py install --user

安装nms

在simple-HRNet文件夹下

cd misc
make
#或是
cd misc/nms
python setup_linux.py build_ext --inplace

但是这样还是会包找不到cpu_nms和gpu_nms包的问题
所以需要打开misc/nms/nms.py文件
改变下面代码，把报错的注释掉，并没用上

# ------------------------------------------------------------------------------
# Copyright (c) Microsoft
# Licensed under the MIT License.
# Modified from py-faster-rcnn (https://github.com/rbgirshick/py-faster-rcnn)
# ------------------------------------------------------------------------------

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import numpy as np

#from cpu_nms import cpu_nms
#from gpu_nms import gpu_nms


def py_nms_wrapper(thresh):
    def _nms(dets):
        return nms(dets, thresh)
    return _nms


#def cpu_nms_wrapper(thresh):
#    def _nms(dets):
#        return cpu_nms(dets, thresh)
#    return _nms


#def gpu_nms_wrapper(thresh, device_id):
#    def _nms(dets):
#        return gpu_nms(dets, thresh, device_id)
#    return _nms

数据集的放置

以下面结构放置

训练

python scripts/train_coco.py --help

python scripts/train_coco.py --help
usage: train_coco.py [-h] [--exp_name EXP_NAME] [--epochs EPOCHS]
                     [--batch_size BATCH_SIZE] [--num_workers NUM_WORKERS]
                     [--lr LR] [--disable_lr_decay]
                     [--lr_decay_steps LR_DECAY_STEPS]
                     [--lr_decay_gamma LR_DECAY_GAMMA] [--optimizer OPTIMIZER]
                     [--weight_decay WEIGHT_DECAY] [--momentum MOMENTUM]
                     [--nesterov]
                     [--pretrained_weight_path PRETRAINED_WEIGHT_PATH]
                     [--checkpoint_path CHECKPOINT_PATH] [--log_path LOG_PATH]
                     [--disable_tensorboard_log] [--model_c MODEL_C]
                     [--model_nof_joints MODEL_NOF_JOINTS]
                     [--model_bn_momentum MODEL_BN_MOMENTUM]
                     [--disable_flip_test_images]
                     [--image_resolution IMAGE_RESOLUTION]
                     [--coco_root_path COCO_ROOT_PATH]
                     [--coco_bbox_path COCO_BBOX_PATH] [--seed SEED]
                     [--device DEVICE]

optional arguments:
  -h, --help            show this help message and exit
  --exp_name EXP_NAME, -n EXP_NAME
                        experiment name. A folder with this name will be
                        created in the log_path.
  --epochs EPOCHS, -e EPOCHS
                        number of epochs
  --batch_size BATCH_SIZE, -b BATCH_SIZE
                        batch size
  --num_workers NUM_WORKERS, -w NUM_WORKERS
                        number of DataLoader workers
  --lr LR, -l LR        initial learning rate
  --disable_lr_decay    disable learning rate decay
  --lr_decay_steps LR_DECAY_STEPS
                        learning rate decay steps
  --lr_decay_gamma LR_DECAY_GAMMA
                        learning rate decay gamma
  --optimizer OPTIMIZER, -o OPTIMIZER
                        optimizer name. Currently, only `SGD` and `Adam` are
                        supported.
  --weight_decay WEIGHT_DECAY
                        weight decay
  --momentum MOMENTUM, -m MOMENTUM
                        momentum
  --nesterov            enable nesterov
  --pretrained_weight_path PRETRAINED_WEIGHT_PATH, -p PRETRAINED_WEIGHT_PATH
                        pre-trained weight path. Weights will be loaded before
                        training starts.
  --checkpoint_path CHECKPOINT_PATH, -c CHECKPOINT_PATH
                        previous checkpoint path. Checkpoint will be loaded
                        before training starts. It includes the model, the
                        optimizer, the epoch, and other parameters.
  --log_path LOG_PATH   log path. tensorboard logs and checkpoints will be
                        saved here.
  --disable_tensorboard_log, -u
                        disable tensorboard logging
  --model_c MODEL_C     HRNet c parameter
  --model_nof_joints MODEL_NOF_JOINTS
                        HRNet nof_joints parameter
  --model_bn_momentum MODEL_BN_MOMENTUM
                        HRNet bn_momentum parameter
  --disable_flip_test_images
                        disable image flip during evaluation
  --image_resolution IMAGE_RESOLUTION, -r IMAGE_RESOLUTION
                        image resolution
  --coco_root_path COCO_ROOT_PATH
                        COCO dataset root path
  --coco_bbox_path COCO_BBOX_PATH
                        path of detected bboxes to use during evaluation
  --seed SEED, -s SEED  seed
  --device DEVICE, -d DEVICE
                        device

训练时，我使用的两台电脑在w48_384x288结构下，一个epochs需要2个多小时，batch_size 16时显存10G左右。