论文复现｜Panoptic Deeplab(全景分割PyTorch)

摘要：这是发表于CVPR 2020的一篇论文的复现模型。

本文分享自华为云社区《Panoptic Deeplab(全景分割PyTorch)》，作者：HWCloudAI 。

这是发表于CVPR 2020的一篇论文的复现模型，B. Cheng et al, “Panoptic-DeepLab: A Simple, Strong, and Fast Baseline for Bottom-Up Panoptic Segmentation”, CVPR 2020，此模型在原论文的基础上，使用HRNet作为backbone，得到了高于原论文的精度，PQ达到了63.7%，mIoU达到了80.3%，AP达到了37.3%。该算法会载入Cityscapes上的预训练模型（HRNet），我们提供了训练代码和可用于训练的模型，用于实际场景的微调训练。训练后生成的模型可直接在ModelArts平台部署成在线服务。

具体算法介绍：https://marketplace.huaweicloud.com/markets/aihub/modelhub/detail/?id=33d3239f-8f0b-4432-a842-f787662ed6a0

注意事项：

1.本案例使用框架:PyTorch1.4.0

2.本案例使用硬件:GPU: 1*NVIDIA-V100NV32(32GB) | CPU: 8 核 64GB

3.运行代码方法: 点击本页面顶部菜单栏的三角形运行按钮或按Ctrl+Enter键运行每个方块中的代码

4.JupyterLab的详细用法: 请参考《ModelAtrs JupyterLab使用指导》

5.碰到问题的解决办法: 请参考《ModelAtrs JupyterLab常见问题解决办法》

1.下载数据和代码

运行下面代码，进行数据和代码的下载

本案例使用cityscapes数据集。

import os

import moxing as mox

# 数据代码下载

mox.file.copy_parallel('s3://obs-aigallery-zc/algorithm/panoptic-deeplab','./panoptic-deeplab')

2.模型训练

2.1依赖库加载

#!/usr/bin/env python3

# -*- coding: utf-8 -*-

from __future__ import print_function

import os

root_path = './panoptic-deeplab/'

os.chdir(root_path)

# 获取当前目录结构信息，以便进行代码调试

print('os.getcwd():', os.getcwd())

import time

import argparse

import time

import datetime

import math

import sys

import shutil

import moxing as mox # ModelArts上专用的moxing模块，可用于与OBS的数据交互，API文档请查看：https://github.com/huaweicloud/ModelArts-Lab/tree/master/docs/moxing_api_doc

from PIL import ImageFile

ImageFile.LOAD_TRUNCATED_IMAGES = True

2.2训练参数设置

parser = argparse.ArgumentParser(description='Panoptic Deeplab')

parser.add_argument('--training_dataset', default='/home/ma-user/work/panoptic-deeplab/', help='Training dataset directory') # 在ModelArts中创建算法时，必须进行输入路径映射配置，输入映射路径的前缀必须是/home/work/modelarts/inputs/，作用是在启动训练时，将OBS的数据拷贝到这个本地路径中供本地代码使用。

parser.add_argument('--train_url', default='./output', help='the path to save training outputs') # 在ModelArts中创建训练作业时，必须指定OBS上的一个训练输出位置，训练结束时，会将输出映射路径拷贝到该位置

parser.add_argument('--num_gpus',  default=1, type=int, help='num of GPUs to train')

parser.add_argument('--eval', default='False', help='whether to eval')

parser.add_argument('--load_weight', default='trained_model/model/model_final.pth',type=str) # obs路径 断点模型 pth文件 如果是评估 则是相对于src的路径

parser.add_argument('--iteration', default=100, type=int)

parser.add_argument('--learning_rate', default=0.001, type=float)

parser.add_argument('--ims_per_batch', default=8, type=int)

args, unknown = parser.parse_known_args() # 必须将parse_args改成parse_known_args，因为在ModelArts训练作业中运行时平台会传入一个额外的init_method的参数

# dir

fname = os.getcwd()

project_dir = os.path.join(fname, "panoptic-deeplab")

detectron2_dir = os.path.join(fname, "detectron2-0.3+cu102-cp36-cp36m-linux_x86_64.whl")

panopticapi_dir = os.path.join(fname, "panopticapi-0.1-py3-none-any.whl")

cityscapesscripts_dir = os.path.join(fname, "cityscapesScripts-2.1.7-py3-none-any.whl")

requirements_dir = os.path.join(project_dir, "requirements.txt")

output_dir = "/home/work/modelarts/outputs/train_output"

# config strings

evalpath = ''

MAX_ITER = 'SOLVER.MAX_ITER ' + str(args.iteration+90000)

BASE_LR = 'SOLVER.BASE_LR ' + str(args.learning_rate)

IMS_PER_BATCH = 'SOLVER.IMS_PER_BATCH ' + str(args.ims_per_batch)

SCRIPT_PATH = os.path.join(project_dir, "tools_d2/train_panoptic_deeplab.py")

CONFIG_PATH = os.path.join(fname, "configs/config.yaml")

CONFIG_CMD = '--config-file ' + CONFIG_PATH

EVAL_CMD = ''

GPU_CMD = ''

OPTS_CMD = MAX_ITER + ' ' + BASE_LR + ' ' + IMS_PER_BATCH

RESUME_CMD = ''

#functions

def merge_cmd(scirpt_path, config_cmd, gpu_cmd, eval_cmd, resume_cmd, opts_cmd):

 return "python " + scirpt_path + " "+ config_cmd + " " + gpu_cmd + " " + eval_cmd + " " + resume_cmd + " " + OPTS_CMD

if args.eval == 'True':

 assert args.load_weight, 'load_weight empty when trying to evaluate' # 如果评估时为空，则报错

 if args.load_weight != 'trained_model/model/model_final.pth':

 #将model拷贝到本地，并获取模型路径

 modelpath, modelname = os.path.split(args.load_weight)

 mox.file.copy_parallel(args.load_weight, os.path.join(fname, modelname))

 evalpath = os.path.join(fname,modelname)

 else:

 evalpath = os.path.join(fname,'trained_model/model/model_final.pth')

    EVAL_CMD = '--eval-only MODEL.WEIGHTS ' + evalpath

else:

    GPU_CMD = '--num-gpus ' + str(args.num_gpus)

 if args.load_weight:

        RESUME_CMD = '--resume'

 if args.load_weight != 'trained_model/model/model_final.pth':

 modelpath, modelname = os.path.split(args.load_weight)

 mox.file.copy_parallel(args.load_weight, os.path.join('/cache',modelname))

 with open('/cache/last_checkpoint','w') as f: #创建last_checkpoint文件

 f.write(modelname)

 f.close()

 else:

 os.system('cp ' + os.path.join(fname, 'trained_model/model/model_final.pth') + ' /cache/model_final.pth')

 with open('/cache/last_checkpoint','w') as f: #创建last_checkpoint文件

 f.write('model_final.pth')

 f.close()

os.environ['DETECTRON2_DATASETS'] = args.training_dataset #添加数据库路径环境变量

cmd = merge_cmd(SCRIPT_PATH, CONFIG_CMD, GPU_CMD, EVAL_CMD, RESUME_CMD, OPTS_CMD)

# os.system('mkdir -p ' + args.train_url)

print('*********Train Information*********')

print('Run Command: ' + cmd)

print('Num of GPUs: ' + str(args.num_gpus))

print('Evaluation: ' + args.eval)

if args.load_weight:

 print('Load Weight: ' + args.load_weight)

else:

 print('Load Weight: None (train from scratch)')

print('Iteration: ' + str(args.iteration))

print('Learning Rate: ' + str(args.learning_rate))

print('Images Per Batch: ' + str(args.ims_per_batch))

2.3安装依赖库

安装依赖库需要几分钟，请耐心等待

def install_dependecies(r,d, p, c):

 os.system('pip uninstall pytorch> out1.txt')

 os.system('pip install  torch==1.7.0> out2.txt')

 os.system('pip install --upgrade pip')

 os.system('pip install --upgrade numpy')

 os.system('pip install torchvision==1.7.0> out3.txt')

 os.system('pip install pydot')

 os.system('pip install --upgrade pycocotools')

 os.system('pip install tensorboard')

 os.system('pip install -r ' + r + ' --ignore-installed PyYAML')

 os.system('pip install ' + d)

 os.system('pip install ' + p)

 os.system('pip install ' + c)

 os.system('pip install pyyaml ==5.1.0')

# 安装依赖

print('*********Installing Dependencies*********')

install_dependecies(requirements_dir,detectron2_dir, panopticapi_dir, cityscapesscripts_dir)

*********Installing Dependencies*********

2.4开始训练

print('*********Training Begin*********')

print(cmd)

start = time.time()

ret = os.system(cmd+ " >out.txt")

if ret == 0:

 print("success")

else:

 print('fail')

end_time=time.time()

print('done')

print(end_time-start)

if args.eval == 'False':

 os.system('mv /cache/model_final.pth ' + os.path.join(fname, 'output/model_final.pth')) #/cache模型移动到输出文件夹

if os.path.exists(os.path.join(fname, 'pred_results')):

 os.system('mv ' + os.path.join(fname, 'pred_results') + ' ' + args.train_url)

训练完成之后，可以在out.txt中看运行日志
在./panoptic-deeplab/output/pred_results/文件目录下，有该模型全景分割，实例分割，语义分割的评估结果

3.模型测试

3.1加载测试函数

from test import *

3.2开始预测

if __name__ == '__main__':

 img_path = r'/home/ma-user/work/panoptic-deeplab/cityscapes/leftImg8bit/val/frankfurt/frankfurt_000000_003920_leftImg8bit.png' # TODO 修改测试图片路径

 model_path = r'/home/ma-user/work/panoptic-deeplab/output/model_final.pth' # TODO 修改模型路径

 my_model = ModelClass(model_path)

    result = my_model.predict(img_path)

 print(result)

点击关注，第一时间了解华为云新鲜技术~