Skip to content

Latest commit

 

History

History
103 lines (71 loc) · 3.81 KB

README.md

File metadata and controls

103 lines (71 loc) · 3.81 KB
Raw

DSDL: Standard Description Language for DataSet

Abstract

Data is the cornerstone of artificial intelligence. The efficiency of data acquisition, exchange, and application directly impacts the advances in technologies and applications. Over the long history of AI, a vast quantity of data sets have been developed and distributed. However, these datasets are defined in very different forms, which incurs significant overhead when it comes to exchange, integration, and utilization -- it is often the case that one needs to develop a new customized tool or script in order to incorporate a new dataset into a workflow.

To overcome such difficulties, we develop Data Set Description Language (DSDL). More details please visit our official documents, dsdl datasets can be downloaded from our platform OpenDataLab.

Steps

  • install dsdl and opendatalab:

    pip install dsdl
pip install opendatalab

  • install mmseg and pytorch: please refer this installation documents.

  • prepare dsdl dataset (take voc2012 as an example)

    • dowaload dsdl dataset (you will need an opendatalab account to do so. register one now)

      cd data

odl login
odl get PASCAL_VOC2012

      usually, dataset is compressed on opendatalab platform, the downloaded voc 2012 dataset should be like this:

      data/
├── PASCAL_VOC2012
│   ├── dsdl
│   │   ├── dsdl_Det_full.zip
│   │   └── dsdl_SemSeg_full.zip
│   ├── raw
│   │   ├── VOC2012test.tar
│   │   ├── VOCdevkit_18-May-2011.tar
│   │   └── VOCtrainval_11-May-2012.tar
│   └── README.md
└── ...

    • decompress dataset

      cd dsdl
unzip dsdl_SemSeg_full.zip

      as we do not need detection dsdl files, we only decompress the semantic segmentation files here.

      cd ../raw
tar -xvf VOCtrainval_11-May-2012.tar
tar -xvf VOC2012test.tar

cd ../../

  • change traning config

    open the voc config file and set some file paths as below:

    data_root = 'data/PASCAL_VOC2012'
img_prefix = 'raw/VOCdevkit/VOC2012'
train_ann = 'dsdl/dsdl_SemSeg_full/set-train/train.yaml'
val_ann = 'dsdl/dsdl_SemSeg_full/set-val/val.yaml'

    as dsdl datasets with one task using one dataloader, we can simplly change these file paths to train a model on a different dataset.

  • train:

    • using single gpu:
    python tools/train.py {config_file}
    • using slrum:
    ./tools/slurm_train.sh {partition} {job_name} {config_file} {work_dir} {gpu_nums}

Test Results

Datasets Model mIoU(%) Config
voc2012 model 76.73 config
cityscapes model 79.01 config