MindSpore Golden Stick

MindSpore Golden Stick is a model compression algorithm set jointly designed and developed by Huawei's Noah team and Huawei's MindSpore team. The architecture diagram of MindSpore Golden Stick is shown in the figure below, which is divided into five parts:

  1. The underlying MindSpore Rewrite module provides the ability to modify the front-end network. Based on the interface provided by this module, algorithm developers can add, delete, query and modify the nodes and topology relationships of the MindSpore front-end network according to specific rules;

  2. Based on MindSpore Rewrite, MindSpore Golden Stick will provide various types of algorithms, such as SimQAT algorithm, SLB quantization algorithm, SCOP pruning algorithm, etc.;

  3. At the upper level of the algorithm, MindSpore Golden Stick also plans advanced technologies such as AMC (AutoML for Model Compression), NAS (Neural Architecture Search), and HAQ (Hardware-aware Automated Quantization). This feature will be provided in future;

  4. In order to facilitate developers to analyze and debug algorithms, MindSpore Golden Stick provides some tools, such as visualization tool, profiler tool, summary tool, etc. This feature will be provided in future;

  5. In the outermost layer, MindSpore Golden Stick encapsulates a set of concise user interface.

_images/golden-stick-arch.png

Note

The architecture diagram is the overall picture of MindSpore Golden Stick, which includes the features that have been implemented in the current version and the capabilities planned in RoadMap. Please refer to release notes for available features in current version.

Code repository address: <https://gitee.com/mindspore/golden-stick>

Installation and Deployment

Post-Training Quantization

Quant-Aware Quantization

Pruner

API References

RELEASE NOTES