前言论文: https://arxiv.org/abs/2106.04263
代码: https://github.com/Atten4Vis/DemystifyLocalViT
本文记录复现 DemystifyLocalViT 在图像语义分割中的应用过程。
作者将该模型语义分割的代码公开在 https://github.com/Atten4Vis/DemystifyLocalViT/tree/master/downstreams/segmentation
Ubuntu 18.4
CUDA 11.2
Conda+Python 3.7.11
mmcv 1.4.8
mmsegmentation 0.23.0
环境说明
作者的 DemystifyLocalViT 中的 Segmentation 部分是使用 mmSegmentation 框架实现的。
并且使用的mmSeg框架版本为0.11.0, 这要求mmcv版本在1.1.4~1.3.9之间。
可能由于CUDA版本问题(其他项目依赖, 不能改), 我mmcv-full==1.3.9版本死活装不上去, 因此使用当前(20220416)最新版本的mmcv1.4.8 + mmSeg0.23.0, 然后将作者的代码迁移过来。
参考 mmSeg官方文档 (中文)
先使用Conda创建环境, 安装pytorch和mmcv, mmsegmentation
conda create -n mmlab python=3.7.11 -y
conda activate mmlab
conda install pytorch=1.11.0 torchvision cudatoolkit=11.3 -c pytorch
pip install mmcv-full==1.4.8 -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.11.0/index.html
# 安装mmsegmentation
git clone https://github.com/open-mmlab/mmsegmentation.git
cd mmsegmentation
pip install -e .
# 链接数据集地址
ln -s <your_dataset_path> data
cd ..
安装完mmSeg之后建议先根据 官网教程验证安装 。
配置项目## 克隆DWNet的储存库, 并复制些有用的文件到我们最新版的mmseg中
#git clone https://github.com/Atten4Vis/DemystifyLocalViT.git
## 1 将配置文件复制过来
#cp -r DemystifyLocalViT/downstreams/segmentation/configs/dwnet mmsegmentation/configs
cd mmsegmentation
# 1.下载配置文件 upernet_dynamic_dwnet_tiny_patch4_window7_512x512_160k_ade20k.py
# 注: 这里配置文件名中的window7是指的配置中的窗口大小, 不是运行环境
cd configs
mkdir dwnet
cd dwnet
wget https://raw.githubusercontent.com/Atten4Vis/DemystifyLocalViT/master/downstreams/segmentation/configs/dwnet/upernet_dynamic_dwnet_tiny_patch4_window7_512x512_160k_ade20k.py
cd ../..
# 2.1.下载模型 dwnet.py
cd mmseg/models/backbones
wget https://raw.githubusercontent.com/Atten4Vis/DemystifyLocalViT/master/downstreams/segmentation/mmseg/models/backbones/dwnet.py
# 2.2.在 mmsegmentation/mmseg/models/backbones/__init__.py中完成DWNet的注册。
...
cd ../../..
# 3.下载模型权重文件 upernet_dynamic_dwnet_tiny.pth
mkdir models
cd models
mkdir dwnet
cd dwnet
wget https://github.com/Atten4Vis/DemystifyLocalViT/releases/download/prerelease/upernet_dynamic_dwnet_tiny.pth
cd ../..
# 4.下载数据集
cd data
mkdir ade
cd ade
wget http://data.csail.mit.edu/places/ADEchallenge/ADEChallengeData2016.zip
http://data.csail.mit.edu/places/ADEchallenge/ADEChallengeData2016.zip
unzip ADEChallengeData2016.zip
cd ../..
# 5.执行训练
python tools/train.py configs/dwnet/upernet_dynamic_dwnet_tiny_patch4_window7_512x512_160k_ade20k.py
# 6.执行测试
python tools/test.py configs/dwnet/upernet_dynamic_dwnet_tiny_patch4_window7_512x512_160k_ade20k.py models/dwnet/upernet_dynamic_dwnet_tiny.pth --eval mIoU
> python tools/test.py configs/dwnet/upernet_dynamic_dwnet_tiny_patch4_window7_512x512_160k_ade20k.py models/dwnet/upernet_dynamic_dwnet_tiny.pth --eval mIoU
2022-04-18 14:38:07,999 - mmseg - INFO - Multi-processing start method is `None`
2022-04-18 14:38:07,999 - mmseg - INFO - OpenCV num_threads is `
2022-04-18 14:38:08,083 - mmseg - INFO - Loaded 2000 images
/home/dl/mmsegmentation/mmseg/models/losses/cross_entropy_loss.py:226: UserWarning: Default ``avg_non_ignore`` is False, if you would like to ignore the certain label and average loss over non-ignore labels, which is the same with PyTorch official cross_entropy, set ``avg_non_ignore=True``.
'Default ``avg_non_ignore`` is False, if you would like to '
load checkpoint from local path: models/dwnet/upernet_dynamic_dwnet_tiny.pth
/home/dl/mmsegmentation/tools/test.py:249: UserWarning: SyncBN is only supported with DDP. To be compatible with DP, we convert SyncBN to BN. Please use dist_train.sh which can avoid this error.
'SyncBN is only supported with DDP. To be compatible with DP, '
[>>>>>>>>>>>>>>>>>>>>>>>>>>>] 2000/2000, 3.0 task/s, elapsed: 665s, ETA: 0sper class results:
+---------------------+-------+-------+
| Class | IoU | Acc |
+---------------------+-------+-------+
| wall | 74.51 | 86.95 |
| building | 81.71 | 91.64 |
| sky | 93.84 | 97.21 |
| floor | 78.77 | 90.32 |
| tree | 71.87 | 87.01 |
| ceiling | 81.25 | 90.56 |
| road | 81.97 | 89.64 |
| bed | 87.05 | 95.01 |
| windowpane | 59.56 | 76.34 |
| grass | 65.73 | 83.84 |
| cabinet | 57.41 | 69.92 |
| sidewalk | 62.65 | 78.48 |
| person | 77.81 | 90.52 |
| earth | 32.16 | 45.79 |
| door | 42.19 | 55.54 |
| table | 54.19 | 71.68 |
| mountain | 53.22 | 72.77 |
| plant | 51.14 | 60.59 |
| curtain | 67.9 | 81.51 |
| chair | 53.13 | 68.15 |
| car | 82.34 | 91.19 |
| water | 46.18 | 59.28 |
| painting | 67.1 | 86.53 |
| sofa | 60.57 | 80.31 |
| shelf | 41.55 | 56.79 |
| house | 50.66 | 67.21 |
| sea | 42.3 | 63.55 |
| mirror | 59.5 | 66.68 |
| rug | 50.02 | 58.25 |
| field | 29.92 | 45.88 |
| armchair | 37.08 | 51.43 |
| seat | 57.46 | 80.3 |
| fence | 36.06 | 50.07 |
| desk | 45.57 | 67.58 |
| rock | 34.54 | 51.15 |
| wardrobe | 43.42 | 62.41 |
| lamp | 58.71 | 71.32 |
| bathtub | 70.73 | 78.86 |
| railing | 32.47 | 45.06 |
| cushion | 51.9 | 66.33 |
| base | 25.58 | 33.95 |
| box | 20.9 | 27.54 |
| column | 38.1 | 49.18 |
| signboard | 35.5 | 48.11 |
| chest of drawers | 39.24 | 54.82 |
| counter | 30.97 | 35.94 |
| sand | 28.66 | 55.3 |
| sink | 68.58 | 76.58 |
| skyscraper | 57.75 | 74.37 |
| fireplace | 72.86 | 86.27 |
| refrigerator | 70.09 | 84.08 |
| grandstand | 39.06 | 61.05 |
| path | 19.23 | 26.7 |
| stairs | 28.09 | 35.65 |
| runway | 65.81 | 87.47 |
| case | 44.03 | 58.9 |
| pool table | 90.51 | 94.46 |
| pillow | 50.98 | 60.66 |
| screen door | 59.52 | 73.15 |
| stairway | 30.88 | 35.85 |
| river | 10.58 | 23.88 |
| bridge | 62.17 | 74.15 |
| bookcase | 36.94 | 53.49 |
| blind | 40.64 | 44.84 |
| coffee table | 53.48 | 77.98 |
| toilet | 82.43 | 89.29 |
| flower | 41.28 | 57.67 |
| book | 42.8 | 65.98 |
| hill | 10.69 | 14.6 |
| bench | 38.36 | 49.08 |
| countertop | 54.45 | 71.18 |
| stove | 75.66 | 83.81 |
| palm | 49.91 | 70.4 |
| kitchen island | 33.69 | 59.09 |
| computer | 66.19 | 78.67 |
| swivel chair | 44.92 | 60.9 |
| boat | 68.29 | 78.67 |
| bar | 27.72 | 33.96 |
| arcade machine | 44.97 | 47.79 |
| hovel | 31.79 | 35.37 |
| bus | 81.86 | 89.83 |
| towel | 56.75 | 65.18 |
| light | 52.41 | 59.77 |
| truck | 28.21 | 35.42 |
| tower | 27.77 | 35.0 |
| chandelier | 66.25 | 78.44 |
| awning | 22.87 | 27.1 |
| streetlight | 23.24 | 28.77 |
| booth | 49.58 | 52.85 |
| television receiver | 65.01 | 75.92 |
| airplane | 57.94 | 65.04 |
| dirt track | 7.45 | 16.1 |
| apparel | 31.15 | 47.0 |
| pole | 20.06 | 25.56 |
| land | 4.21 | 7.02 |
| bannister | 10.05 | 14.81 |
| escalator | 23.01 | 26.02 |
| ottoman | 42.57 | 56.1 |
| bottle | 32.97 | 48.27 |
| buffet | 46.11 | 54.5 |
| poster | 28.57 | 38.07 |
| stage | 15.52 | 24.34 |
| van | 41.29 | 60.72 |
| ship | 45.55 | 47.46 |
| fountain | 19.45 | 20.15 |
| conveyer belt | 69.46 | 90.24 |
| canopy | 11.18 | 14.57 |
| washer | 62.97 | 67.66 |
| plaything | 15.63 | 25.52 |
| swimming pool | 38.25 | 68.51 |
| stool | 37.8 | 49.99 |
| barrel | 40.56 | 66.37 |
| basket | 27.09 | 36.52 |
| waterfall | 54.61 | 63.26 |
| tent | 84.77 | 97.99 |
| bag | 10.94 | 13.71 |
| minibike | 51.07 | 58.24 |
| cradle | 79.27 | 93.83 |
| oven | 43.34 | 62.98 |
| ball | 39.68 | 45.41 |
| food | 52.07 | 62.52 |
| step | 9.94 | 11.85 |
| tank | 55.94 | 59.83 |
| trade name | 29.85 | 37.78 |
| microwave | 74.1 | 81.89 |
| pot | 38.97 | 48.09 |
| animal | 56.87 | 60.89 |
| bicycle | 54.09 | 72.34 |
| lake | 48.25 | 55.53 |
| dishwasher | 55.37 | 71.9 |
| screen | 67.28 | 84.68 |
| blanket | 12.37 | 14.05 |
| sculpture | 49.82 | 60.43 |
| hood | 55.38 | 61.45 |
| sconce | 40.37 | 50.29 |
| vase | 32.64 | 46.22 |
| traffic light | 29.17 | 47.19 |
| tray | 3.72 | 7.58 |
| ashcan | 35.5 | 44.42 |
| fan | 55.26 | 68.46 |
| pier | 49.24 | 66.17 |
| crt screen | 2.66 | 7.45 |
| plate | 51.61 | 64.61 |
| monitor | 5.19 | 6.29 |
| bulletin board | 41.7 | 50.1 |
| shower | 1.09 | 4.05 |
| radiator | 51.56 | 58.36 |
| glass | 10.9 | 12.11 |
| clock | 30.79 | 37.01 |
| flag | 31.0 | 36.27 |
+---------------------+-------+-------+
Summary:
+-------+-------+-------+
| aAcc | mIoU | mAcc |
+-------+-------+-------+
| 81.15 | 45.72 | 57.06 |
+-------+-------+-------+
Process finished with exit code 0
欢迎分享，转载请注明来源：内存溢出
微信扫一扫
支付宝扫一扫
评论列表（0条）