GitHub - LPeng625/mmsegmentation: OpenMMLab Semantic Segmentation Toolbox and Benchmark.

运行步骤

一、环境配置

1、使用配置好的conda环境

下载环境：https://pan.baidu.com/s/1YWSuoMlPPF8O6JXnHiNSLQ?pwd=mdvs -> 解压环境 -> 激活 conda: activate openmmlab

打包conda环境（该指令打包时候用，配置时不需要！！！）

conda pack -n openmmlab -o openmmlab.tar.gz --ignore-editable-packages

2、按照官网配置

二、基础配置

1、基础配置更改

配置1：configs/birenet/birenet_fcn_1xb4-306k_cityscapes-256x512.py

调整输入图片尺寸

crop_size = (512, 512)

原配置是4卡，改成单卡eta_min需要设置为原来的1/4

eta_min=1e-4 / 4,

原配置是4卡，改成单卡lr需要设置为原来的1/4

optimizer = dict(type='SGD', lr=0.05 / 4, momentum=0.9, weight_decay=0.0005)

添加mDice和mFscore指标

val_evaluator = dict(type='IoUMetric', iou_metrics=['mIoU', 'mDice', 'mFscore'], output_dir='work_dir/birenet/format_results')
test_evaluator = val_evaluator

配置2：configs/base/models/birenet.py

norm_cfg = dict(type='BN', requires_grad=True) #TODO SyncBN->BN
num_classes=2, #TODO 19->2

配置3：configs/base/datasets/cityscapes_256x512.py

调整输入图片尺寸

crop_size = (256, 512)

配置4：configs/base/datasets/cityscapes.py

调整数据集路径

data_root = ''

配置5：configs/base/default_runtime.py

添加tensorboard和wandb

vis_backends=[dict(type='LocalVisBackend'),
              # tensorboard 路径eg:work_dir/birenet/20230510_200431/vis_data
              dict(type='TensorboardVisBackend'),
              # wandb 特征图路径eg：work_dir/birenet/20230511_150800/vis_data/wandb/run-20230511_150804-9z6vfqe3/files/media/images
              # 备注：需要visualization=dict(type='SegVisualizationHook', draw=True, interval=815))，draw=True才能看到特征图
              dict(type='WandbVisBackend')]

配置6：configs/base/schedules/schedule_306k.py

更改 max_iters 和 val_interval、interval

train_cfg = dict(
    type='IterBasedTrainLoop', max_iters=306000, val_interval=204) 
logger=dict(type='LoggerHook', interval=204,

测试/验证期间结果可视化特征图存储路径eg

work_dir/birenet/20230511_150800/vis_data/vis_image

wandb特征图存储路径eg

work_dir/birenet/20230511_150800/vis_data/wandb/run-20230511_150804-9z6vfqe3/files/media/images
visualization=dict(type='SegVisualizationHook', draw=True, interval=815))

2、更改类别及后缀

mmseg/datasets/cityscapes.py 更改类别

CLASSES = ('background','road')
PALETTE = [[0, 0, 0], [255, 255, 255]]

更改后缀

img_suffix='.png',
seg_map_suffix='_road.png',

3、常用指令

train 单gpu

python tools/train.py ${配置文件} --work-dir ${工作路径} --resume  load_from=${检查点}
load_from=${检查点} 指定检查点
--resume 默认从最新的检查点恢复
eg:
CUDA_VISIBLE_DEVICES=0  python tools/train.py  configs/birenet/birenet_1xb8-300epoch_deepglobe-1024x1024.py  --work-dir work_dir/birenet

多gpu

sh tools/dist_train.sh ${配置文件} ${GPU数量} --work-dir  ${工作路径} --resume  load_from=${检查点}
eg:
./tools/dist_train.sh  configs/birenet/birenet_1xb8-300epoch_deepglobe-1024x1024.py 2
--work-dir --work-dir work_dir/birenet

test

python tools/test.py ${配置文件} ${模型权重文件} [可选参数]
eg:
CUDA_VISIBLE_DEVICES=5 python tools/test.py configs/birenet/birenet_1xb8-300epoch_deepglobe-1024x1024.py  work_dir/birenet/best_mIoU_iter_XXX.pth

测试FPS

CUDA_VISIBLE_DEVICES=0 python tools/analysis_tools/benchmark.py configs/birenet/birenet_1xb8-300epoch_deepglobe-1024x1024.py  work_dir/birenet/best_mIoU_iter_XXX.pth

测试FLOPs/Params

CUDA_VISIBLE_DEVICES=0 python tools/analysis_tools/get_flops.py  configs/birenet/birenet_1xb8-300epoch_deepglobe-1024x1024.py   --shape 1024 1024

对 mIoU, mAcc, aAcc ,mFscore 指标画图

python tools/analysis_tools/analyze_logs.py work_dir/birenet/20240501_221552/vis_data/20240501_221552.json --keys mIoU mAcc aAcc mFscore --legend mIoU mAcc aAcc mFscore

对 loss 指标画图

python tools/analysis_tools/analyze_logs.py work_dir/birenet/20240508_195106/vis_data/20240508_195106.json --keys loss --legend loss

4、测试/验证期间通过wandb实现特征图可视化

配置1：configs/base/default_runtime.py

添加tensorboard和wandb
vis_backends=[dict(type='LocalVisBackend'),
              # tensorboard 路径eg:work_dir/birenet/20230510_200431/vis_data
              dict(type='TensorboardVisBackend'),
              # wandb 特征图路径eg：work_dir/birenet/20230511_150800/vis_data/wandb/run-20230511_150804-9z6vfqe3/files/media/images
              dict(type='WandbVisBackend')]

配置2：configs/base/schedules/schedule_160k.py

测试/验证期间结果可视化
特征图存储路径eg：
work_dir/birenet/20230511_150800/vis_data/vis_image

wandb特征图存储路径eg：
work_dir/birenet/20230511_150800/vis_data/wandb/run-20230511_150804-9z6vfqe3/files/media/images
visualization=dict(type='SegVisualizationHook', draw=True, interval=815))

5、通过脚本实现特征图可视化

脚本一：

tools/visualization/visualization.py

特征图存储路径eg：
tools/visualization/work_dirs/vis_data/vis_image/out_file_cityscapes_0.png

脚本二：

python tools/visualization/feature_map_visual.py \
tools/visualization/img/1.png \
configs/ann/ann_r50-d8_4xb2-40k_cityscapes-512x1024.py \
tools/visualization/test.pth \
--gt_mask tools/visualization/img/1_label.png

特征图存储路径（wandb）eg：
mmsegmentation/mmsegmentation/vis_data/wandb/run-20230511_151617-mb9txy83/files/media/images

Acknowledgments

We thank the authors of https://github.com/open-mmlab/mmsegmentation for open-source code.

Name		Name	Last commit message	Last commit date
Latest commit History 1,011 Commits
.circleci		.circleci
.dev_scripts		.dev_scripts
.github		.github
configs		configs
demo		demo
docker		docker
docs		docs
mmseg		mmseg
projects		projects
requirements		requirements
resources		resources
tests		tests
tools		tools
.gitignore		.gitignore
.owners.yml		.owners.yml
.pre-commit-config.yaml		.pre-commit-config.yaml
CITATION.cff		CITATION.cff
LICENSE		LICENSE
MANIFEST.in		MANIFEST.in
README.md		README.md
README_zh-CN.md		README_zh-CN.md
dataset-index.yml		dataset-index.yml
model-index.yml		model-index.yml
requirements.txt		requirements.txt
setup.cfg		setup.cfg
setup.py		setup.py
workdoc		workdoc

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Repository files navigation

运行步骤

一、环境配置

1、使用配置好的conda环境

2、按照官网配置

二、基础配置

1、基础配置更改

配置1：configs/birenet/birenet_fcn_1xb4-306k_cityscapes-256x512.py

配置2：configs/base/models/birenet.py

配置3：configs/base/datasets/cityscapes_256x512.py

配置4：configs/base/datasets/cityscapes.py

配置5：configs/base/default_runtime.py

配置6：configs/base/schedules/schedule_306k.py

2、更改类别及后缀

3、常用指令

4、测试/验证期间通过wandb实现特征图可视化

配置1：configs/base/default_runtime.py

配置2：configs/base/schedules/schedule_160k.py

5、通过脚本实现特征图可视化

Acknowledgments

About

Uh oh!

Releases

Packages

Languages

License

LPeng625/mmsegmentation

Folders and files

Latest commit

History

Repository files navigation

运行步骤

一、环境配置

1、使用配置好的conda环境

2、按照官网配置

二、基础配置

1、基础配置更改

配置1：configs/birenet/birenet_fcn_1xb4-306k_cityscapes-256x512.py

配置2：configs/base/models/birenet.py

配置3：configs/base/datasets/cityscapes_256x512.py

配置4：configs/base/datasets/cityscapes.py

配置5：configs/base/default_runtime.py

配置6：configs/base/schedules/schedule_306k.py

2、更改类别及后缀

3、常用指令

4、测试/验证期间通过wandb实现特征图可视化

配置1：configs/base/default_runtime.py

配置2：configs/base/schedules/schedule_160k.py

5、通过脚本实现特征图可视化

Acknowledgments

About

Resources

License

Code of conduct

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Languages

Packages