torchseg

tag v0.4 use pytorch0.4

use pytorch to do image semantic segmentation

• train+val pspnet python test/pspnet_test.py --batch_size=8 --net_name=pspnet --augmentation=False --learning_rate=0.01 --optimizer=sgd --backbone_name=resnet101 --midnet_scale=10 --upsample_type=bilinear --backbone_pretrained=True

current version use pytorch1.1+ (for DDP and sync batchnorm support)

• DDP trainning

  python test/pspnet_test.py --test dist --mp_dist True --note dist --batch_size 4

• support sync batchnorm

  python test/pspnet_test.py --test dist --mp_dist True --note dist --batch_size 4 --use_sync_bn True

experiments

1. 100 epoch
2. 200 epoch
3. require GPU number >=2, input_shape=(713,713)

offical result

requriments

• for detail version see requirements.txt
• test on python3

  conda install pytorch=1.1 torchvision cudatoolkit=9.2 -c pytorch
  pip install opencv-python
  pip install tensorboardX
  pip install easydict
  pip install imgaug
  pip install pandas
  pip install torchsummary
  ...

code reference

• https://github.com/meetshah1995/pytorch-semseg

  test pspnet model on cityscapes dataset
  Overall Acc:     0.865857910768811
  Mean Acc :   0.4482797176755918
  FreqW Acc :      0.7728509434255326
  Mean IoU :   0.36876733235206416
  note for official website https://hszhao.github.io/projects/pspnet/
  the miou is 0.8+

• https://github.com/CSAILVision/semantic-segmentation-pytorch
• cityscapes dataset evaluation
• dual attention

blog reference

• paper and dataset for semantic segmentation introduction

benchmark

cityscapes

• use code from https://github.com/mcordts/cityscapesScripts (only support for python2)
• change to python2 environment source activate env2
• clone them and install them by pip install .
• export CITYSCAPES_RESULTS=/media/sdc/yzbx/benchmark_output/cityscapes
• export CITYSCAPES_DATASET=/media/sdb/CVDataset/ObjectSegmentation/archives/Cityscapes_archives/gtFine_trainvaltest
• open terminal and input: csEvalPixelLevelSemanticLabeling
• the image size and id transform can be view in test/benchmark_test.py
• note: the benchmark is very slow, even for only 500 val images, about 5-10s/per image
• a failed result can be see in #3

voc2012

rm -rf output/results/VOC2012/Segmentation/comp6_test_cls
mv xxx output/results/VOC2012/Segmentation/comp6_test_cls
tar -czvf results.tgz results

useful script

• python utils/summary_to_csv.py output best val/iou to orange tab file
• python test/dp_log.py view pid and detail
• pipreqs --ignore models/examples,models/mxnet,utils/model_hyperopt.py --force . pip requirements
• python tools/fbms_train.py motion segmentation
• python tools/pspnet_test.py semantic segmetnation
• python notebooks/log_statistic.py --app dump_tasks --note xxx --dump_group dump motion segmentation experiment in table format.