pytorch-deeplab (semantic segmentation, distributed training)

Introduction

This is a PyTorch implementation of DeepLab-V3-Plus. It can use Modified Aligned Xception and ResNet as backbone. Currently, we train DeepLab V3 Plus using Pascal VOC 2012, SBD, Cityscapes, and basicDataset(customer-made) datasets.

Prepare basicDataset(customer-made) datasets

/datasets
    /train
        /image
            a.jpg
            ...
        /label
            a.png
            ...
    /val
        /image
            b.jpg
            ...
        /label
            b.png
            ...
    /test
        /image
            c.jpg
            ...
        /label
            c.png
            ...
            

Train, Val, and Test selection

args.testValTrain: '-1: infer, 0: test, 1: testval, 2: train, 3: trainval, 4: trainvaltest'

infer: only do inference
test: do inference and calculate metrics such as miou and fwiou
...

How to train

1 Use slurm

1.1 foreground running

srun -N1 --cpus-per-task 16 --gres gpu:4 python train_main.py --testValTrain 4

1.2 background running

set "args.testValTrain = 2, 3, or 4" and then run the command below

sbatch background_running.slurm

2 directly use gpu machines

CUDA_VISIBLE_DEVICES=0,1,2,3 python train_main.py --testValTrain 4 --resume run/basicDataset/deeplab-resnet/experiment_7/checkpoint.pth.tar --loss_type ce

How to test or val

1 Use slurm

1.1 foreground running

srun -N1 --cpus-per-task 8 --gres gpu:4 python train_main.py --testValTrain 1 --resume run/basicDataset/deeplab-resnet/experiment_*/model_best.pth.tar 

1.2 background running

set "args.testValTrain = 1" and then run the command below

sbatch background_running.slurm

2 directly use gpu machines

CUDA_VISIBLE_DEVICES=0,1,2,3 python train_main.py --testValTrain 1 --resume run/basicDataset/deeplab-resnet/experiment_*/model_best.pth.tar --dump_image

Inference only

This is for development.

python inference.py --resume run/basicDataset/deeplab-resnet/experiment_*/model_best.pth.tar --gpu_id .......

Acknowledgement

pytorch-deeplab-xception

PyTorch-Encoding

Synchronized-BatchNorm-PyTorch

drn

How to push and pull code on GitHub

1. 初始化本地仓库

git init

2. 关联本地仓库到远程仓库

git remote add origin https://github.com/User/Repo.git

3. 添加要提交的文件到暂存区

git add *

4. 提交代码到文件控制仓库

git commit -m "init commint"

5. 将远程主机的更新，全部取回本地

git fetch origin

6. 拉取远程分支代码到本地

git pull origin main 

如果报错用这个

git pull origin main --allow-unrelated-histories

7. 提交本地分支(main)代码到远程分支(main)

git push -u origin main

Q: 本地仓库的东西push到远程仓库时报错

A: 因为本地仓库的文件和远程仓库不一样，所以要先用命令git pull -f origin main将远程仓库的文件拉到本地. 可是这样做之后再git push还是没有用，还是报同样的错误。 于是就用git push -f origin main强制push就成功了。（注意：大家千万不要随便用-f的操作，因为f意味着强制push，会覆盖掉远程的所有代码！）

当我们多人协作写一个项目的时候，我们会发现上传代码到远程github（码云等）时，拉取上传会很麻烦，很有可能会将我们本来改好的代码直接覆盖掉，这很不利于我们的更新操作。因此，下面我给大家介绍一下如何操作可以避免覆盖问题的发生：

1、先将本地代码放到暂存区 git stash

2、将远程github（码云等）上面的代码拉取下来 git pull

3、将第一步暂存区的代码放回本地 git stash pop

4、下面继续我们平时的正常上传代码的操作即可： git add . 或者 git add -A git commit -m '操作内容' git push