T/SZROBOT 004-2023 Deep learning network model compression and quantization algorithm evaluation specification

Scope

Scope: This standard proposes an evaluation index system and evaluation process for deep learning network model compression and quantization algorithms, as well as requirements-phase evaluation, design-phase evaluation, implementation-phase evaluation, and operation-phase evaluation. This standard is applicable to guiding organizations such as deep learning algorithm developers, users, and third parties in conducting reliability evaluations of deep learning network model compression and quantization algorithms. Main technical content: This standard proposes an evaluation index system and evaluation process for deep learning network model compression and quantization algorithms, as well as requirements-phase evaluation, design-phase evaluation, implementation-phase evaluation, and operation-phase evaluation. 1. Scope 2. Algorithm Introduction 2.1 Multi-Model Compression Algorithm 2.2 Quantization Algorithm 2.3 Pruning Algorithm 3. Terms and Definitions 3.1 Reliability 3.2 Reliability Evaluation 3.3 Algorithm Failure 3.4 Hazard 3.5 Hazard Severity 3.6 Precision Rate 3.7 Recall Rate 3.8 Accuracy Rate 3.9 Corresponding Time 3.10 Bitops 4. Evaluation Index System 4.1 Evaluation Index System Table 4.2 Correctness of Algorithm Function Implementation 4.3 Performance of Algorithm Compression Implementation 4.4 Impact of Quantized Training Dataset 4.5 Impact of Software and Hardware Platform Dependency 4.6 Impact of Environmental Data 5. Evaluation Process 6. Requirements-Phase Evaluation 7. Design-Phase Evaluation 8. Implementation-Phase Evaluation 9. Operation-Phase Evaluation Appendix A Appendix B

Development Information

Related Standards