Code Monkey home page Code Monkey logo

paddlepaddle_yolact's Introduction

YOLACT

其实并不是完美复刻,只能说大部分情况跟随了原仓库。

更新日记

2020/04/22:初次见面

需要补充

从ImageNet预训练的resnet50dcn开始训练

概述

mask AP为30.5。飞桨论文复现挑战赛参赛作品。

官方权重转换

1_pytorch2paddle.py将yolact作者的pytorch模型转换为paddlepaddle模型。(地址https://github.com/dbolya/yolact) 再对转换好的模型finetune,减少训练时间。 唯一import torch的地方。 这个脚本在本地windows上跑,再把pretrained_resnet50dcn模型上传到AIStudio。如果不允许,请从头训练。

模型下载

30.5 mask AP:链接:https://pan.baidu.com/s/12jRS1satM_OoJ-dSMatC5Q 提取码:z58d

训练

train.py用于训练 有两种训练模型。0-从头训练,1-读取模型继续训练。通过指定--pattern参数指定。 如果你要从头训练,键入下面命令:

nohup python train.py --initial_step=0 --conf_loss=ce_loss --pattern=0>> train.txt 2>&1 &

如果你从转换后的pretrained_resnet50dcn模型训练,键入下面命令:

nohup python train.py --initial_step=800000 --steps=830000 --conf_loss=ce_loss --pattern=1 --model_path=./pretrained_resnet50dcn>> train.txt 2>&1 &

我试过,训练30000步之后就能达到30% mAP以上。

如果想接着从上次训练后的模型(比如是step1066300-ep073-loss5.715.pd)继续训练,键入下面命令:

nohup python train.py --initial_step=1066300 --steps=1230000 --conf_loss=ce_loss --pattern=1 --model_path=./weights/step1066300-ep073-loss5.715.pd>> train.txt 2>&1 &

评估

test_dev.py用于跑COCO test-dev的图片,最后会生成一个json文件 results/detections_test-dev2017_yolactplus_results.json 提交到 https://competitions.codalab.org/competitions/20796#participate 进行评分。

eval.py用于计算val2017的mAP

预测

demo.py用于预测images/test/下的图片,结果保存在images/res/目录下。 一些参数(model_path、backbone_name、obj_threshold、nms_threshold等)在代码中改。

传送门

cv算法交流q群:645796480 但是关于仓库的疑问尽量在Issues上提,避免重复解答。

广告位招租

可联系微信wer186259

Citation

@inproceedings{yolact-iccv2019,
  author    = {Daniel Bolya and Chong Zhou and Fanyi Xiao and Yong Jae Lee},
  title     = {YOLACT: {Real-time} Instance Segmentation},
  booktitle = {ICCV},
  year      = {2019},
}
@misc{yolact-plus-arxiv2019,
  title         = {YOLACT++: Better Real-time Instance Segmentation},
  author        = {Daniel Bolya and Chong Zhou and Fanyi Xiao and Yong Jae Lee},
  year          = {2019},
  eprint        = {1912.06218},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}

paddlepaddle_yolact's People

Contributors

miemie2013 avatar

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.