Code Structure Overview

This repository demonstrates the core data processing pipeline and the implementation of various operators, divided into Python and C++ layers.

Python Layer (code/python/)

• /python/ defines the interfaces for invoking the pipeline and core operators.
• /python/engine contains the pipeline-related logic.
• /python/audio, /python/text, and /python/vision define the operator interfaces for different data modalities (audio, text, and vision respectively).
• /python/transforms contains definitions of various operations designed to transform and process input data, including tensor type conversion and tensor padding.

C++ Layer (code/src/)

• /src/engine contains the core logic of the data processing pipeline.
• /src/api includes interface definitions and encapsulation logic for data preprocessing in various modalities (audio, vision, and text).
• The actual operator implementations and intermediate layers are located in the corresponding directories under /src/. For example:/src/audio/ir/kernels/ contains intermediate representations (IR) of audio-related operators.

Python-C++ Binding and Runtime Utilities

• code/src/api/python/pybind_conversion provides a set of utilities for data conversion between Python and C++, such as converting a Python list of DatasetNode objects to a C++ std::vector<std::shared_ptr<DatasetNode>>.
• code/src/api/python/python_mp defines a multi-process runtime management framework that enables thread and worker process binding.

The complete code is at https://github.com/mindspore-ai/mindspore.