Skip to content

Latest commit

 

History

History
57 lines (37 loc) · 2.26 KB

custom_dataset.md

File metadata and controls

57 lines (37 loc) · 2.26 KB

Custom DataSet

In order to train on your own dataset, you should do the following three steps:

Data Preparation

Data Structure

To be simple, you should reorganize your dataset as follow.

├── your_dataset
│   ├── images
│   │   ├── training
│   │   │   ├── xxx{img_suffix}
│   │   │   ├── yyy{img_suffix}
│   │   │   ├── zzz{img_suffix}
│   │   ├── validation
│   ├── annotations
│   │   ├── training
│   │   │   ├── xxx{seg_map_suffix}
│   │   │   ├── yyy{seg_map_suffix}
│   │   │   ├── zzz{seg_map_suffix}
│   │   ├── validation

Images and labels are stored separately, and are part into training and testing set. The above four directory path will be specific in the script, in a relative way to the dataset root path, so you can rename them as you like.

Annotations Format

Only support for gray-scale image now, you should always transform your image_label into gray-scale image first if needed.

The pixel intensity means the class index, and you can set a ignore_index which doesn't attend the computation of metric.

Write A New Script

  1. copy a existed dataset script in src/datasets, and replace the origin dataset name to your dataset name.
  2. change the num_classes default parameter.
  3. override the init function to specific some parameters especially self.file_list, which is a list of Image/Label path correspondence
    • ADE20K scans all files in the image_dir, and replace img_suffix to seg_map_suffix in the filename.
    • Cityscapes scans files with the given suffix, and get the correspondence in a sorted way.
    • You can refer to the two method above to create your dataset Image/Label correspondence. Both method need to modify the img_suffix and seg_map_suffix in the correct place.

​ 4. In the init.py, add your own dataset into if-elif structure in the get_dataset function.

Create Yaml Config

  1. copy a existed yaml config

  2. change the following parameters in Data :DATASET, DATA_PATH，NUM_CLASSES

  3. change parameters as you need. To do this, you'd better have a good understanding about the meaning of parameters in config.

Now you can use the command to train on your own dataset.