• 设为首页
  • 点击收藏
  • 手机版
    手机扫一扫访问
    迪恩网络手机版
  • 关注官方公众号
    微信扫一扫关注
    公众号

zym1119/DeepLabv3_MobileNetv2_PyTorch: A PyTorch Implementation of MobileNetv2+D ...

原作者: [db:作者] 来自: 网络 收藏 邀请

开源软件名称(OpenSource Name):

zym1119/DeepLabv3_MobileNetv2_PyTorch

开源软件地址(OpenSource Url):

https://github.com/zym1119/DeepLabv3_MobileNetv2_PyTorch

开源编程语言(OpenSource Language):

Python 100.0%

开源软件介绍(OpenSource Introduction):

DeepLabv3_MobileNetv2

This is a PyTorch implementation of MobileNet v2 network with DeepLab v3 structure used for semantic segmentation.

The backbone of MobileNetv2 comes from paper:

Inverted Residuals and Linear Bottlenecks: Mobile Networks for Classification, Detection and Segmentation

And the segment head of DeepLabv3 comes from paper:

Rethinking Atrous Convolution for Semantic Image Segmentation

Please refer to these papers about details like Atrous Convolution, Inverted Residuals, Depthwise Convolution or ASPP if you have some confusion about these blocks.

Results

After training for 150 epochs, without any further tuning, the first training result on test set is like: img

Feel free to change any config or code in this repo :-)

How to use?

First you need to install dependencies of this implementation. This implementation is written under Python 3.5 with following libs:

torch 0.4.0
torchvision 0.2.1
numpy 1.14.5
opencv-python 3.4.1.15
tensorflow 1.8.0 (necessary for tensorboardX)
tensorboardX 1.2

use sudo pip install lib to install them

Then, prepare cityscapes dataset or your own dataset. Currently, cityscapes is the only supported dataset without any modification.

Cityscapes dataset should have the following hierachy:

dataset_root
|   trainImages.txt
|   trainLabels.txt
|   valImages.txt
|   valLabels.txt 
|
└───gtFine(Label Folder)
|   └───train(train set)
|   |   └───aachen(city)
|   |   └───bochum
|   |   └───...
|   |
|   └───test(test set)
|   └───val(val set)
|
└───leftImg8bit(Image Folder)
    └───train
    └───test
    └───val

Don't worry about txt files if you don't have them, this program can generate unexist txt files automatically.

Third, modify config.py to fit your own training policy or configuration

At last, run python main.py --root /your/path/to/dataset/ or just run python main.py

After training, tensorboard is also available to observe training procedure using tensorboard --logdir=./exp_dir/summaries

Tips

I have changed a little from origin MobileNetv2 and DeepLabv3 network, here are the changes:

1. The multi-grid blocks have the same structure with the 7-th layer in MobileNetv2 while 
the rest layers of MobileNetv2 are discarded.
2. The lr decay is determined by epoch not iterations as in DeepLab and the input image 
is randomly cropped by 512 instead of 513 in DeepLab.
3. During training, a input image is first resized so that the shorter side is 600 pixel, 
then cropped into 512 pixels square and sent into network.

If you have some question, please leave an issue.

ImageNet pre-trained weights are loaded from Randl's github, really helpful!

TO-DO

  1. add cityscapes visualization tools(Done)
  2. fine-tune training policy
  3. use ImageNet pre-trained model(Done)

Logs

Date Changes
7.11 fix bugs in network.Test(), add cityscapes output visualization function
7.12 fix bugs in network.plot_curve(), add checkpoint split to avoid out of memory, add save loss in network.save_checkpoint()
7.13 fix bugs in figure save, add checkpoint@150epoch
7.27 upload ImageNet pre-trained weight



鲜花

握手

雷人

路过

鸡蛋
该文章已有0人参与评论

请发表评论

全部评论

专题导读
热门推荐
阅读排行榜

扫描微信二维码

查看手机版网站

随时了解更新最新资讯

139-2527-9053

在线客服(服务时间 9:00~18:00)

在线QQ客服
地址:深圳市南山区西丽大学城创智工业园
电邮:jeky_zhao#qq.com
移动电话:139-2527-9053

Powered by 互联科技 X3.4© 2001-2213 极客世界.|Sitemap