Skip to content

Latest commit

 

History

History
79 lines (61 loc) · 4.51 KB

dev_guide_en.md

File metadata and controls

79 lines (61 loc) · 4.51 KB
Raw

English | 简体中文

Developer's Guide

This document introduces the advanced content of this toolkit and is helpful for readers who want to contribute their code.

0 Contents

1 Directory Structure

PanopticSeg
├── configs: Configuration files. Organized by model.
├── deploy: Code related to model deployment.
│   └── python: Python deployment code.
├── docs: Documents.
├── paddlepanseg:
│   ├── core: Core APIs that supports model training, evaluation, and inference.
│   ├── cvlibs: Core data structures.
│   │   ├── config.py: Manage configurations.
│   │   ├── info_dicts.py: Data containers for prediction results, samples, etc.
│   │   └── manager.py: Register and manage components.
│   ├── datasets: Interfaces for dataset IO.
│   ├── deploy: Deployment APIs.
│   ├── models: Implementations of models.
│   │   ├── backbones: Implementations of backbone networks.
│   │   ├── losses: Implementations of loss functions.
│   │   ├── ops: Implementations of external operators.
│   │   ├── common.py: Common classes and functions.
│   │   └── param_init.py: Utility functions for weight initialization.
│   ├── postprocessors: Postprocessors.
│   ├── runners: Runners.
│   ├── transforms: Data transformation operators.
│   │   └── generate_targets: Transformation operators to generate training targets.
│   └── utils: Utility functions and classes.
├── test_tipc: TIPC scripts.
└── tools: Useful tools for users.
    ├── data: Tools for preparing public datasets.
    ├── analyze_model.py: Analyze the parameter count and computational complexity of a panoptic segmentation model.
    ├── export.py: Export model to static graph.
    ├── predict.py: Make inference with a pre-trained model and produce visualization results.
    ├── train.py: Train a model on a specific dataset.
    └── val.py: Evaluate the performance of a model on a specific dataset.

2 Data Structures

2.1 InfoDict

InfoDict objects are the containers for predictions, samples, etc. used in this toolkit. InfoDict is one of the most important data structures of this toolkit and does not exist in PaddleSeg. The InfoDict class is defined in paddlepanseg/cvlibs/info_dicts.py.

Basically we have four built-in InfoDict types: SampleDict, NetOutDict, PPOutDict, and MetricDict. They are all subclasses of InfoDict.

  • SampleDict: Designed to contain information and meta-information of data samples.
  • NetOutDict: Designed to contain network outputs.
  • PPOutDict: Designed to contain output of a postprocessor.
  • MetricDict: Designed to contain output of calling the evaluate() method of an Evaluator object.

The factory function build_info_dict() defined in paddlepanseg/cvlibs/info_dicts.py is the suggested way to construct an InfoDict object. By passing the _type_ argument you specify the type of InfoDict (one of 'sample', 'net_out', 'pp_out', and 'metric').

3 Encoding Panoptic Segmentation Labels

Please find more details here.

4 Dataset Customization

The recommended way of customizing your own dataset is to organize the dataset in MS COCO format and write a new dataset interface inheriting from paddlepanseg.datasets.base_dataset.COCOStylePanopticDataset.

Adding a new subclass of COCOStylePanopticDataset usually involves the following steps:

  • Store the meta-information of the dataset in a list of dicts. See CITYSCAPES_CATEGORIES in paddlepanseg/datasets/cityscapes.py for an example.
  • Define a Python class that inherits from paddlepanseg.datasets.base_dataset.COCOStylePanopticDataset. Set the class attribute CATEGORY_META_INFO to the list of dicts you defined in the first step, and set the class attribute NUM_CLASSES to the number of classes in the dataset.
  • Rewrite the static method _get_image_id(), which accepts the path of the image and returns a unique identifier of that image.
  • Wrap the new dataset class with the decorator paddlepanseg.cvlibs.manager.DATASETS.add_component(). See paddlepanseg/datasets/cityscapes.py for an example.