T/SZROBOT 0004-2023 Deep learning network model compression and quantization algorithm evaluation specification

Scope

Scope: This standard proposes an evaluation index system and evaluation process for compression and quantization algorithms of artificial intelligence deep learning network models, as well as content such as evaluation in the requirements phase, design phase, implementation phase, and operation phase. This standard is applicable to guiding relevant organizations, including deep learning algorithm developers, users, and third parties, in conducting reliability evaluations of compression and quantization algorithms for deep learning network models. Note: When there are relevant dedicated product standards, product standards take precedence over this standard. Main technical content: 1. Scope 2. Algorithm introduction 2.1 Multi-model compression algorithms 2.2 Quantization algorithms 2.3 Pruning algorithms 3. Terms and definitions 3.1 Reliability 3.2 Reliability evaluation 3.3 Algorithm failure 3.4 Danger 3.5 Danger severity 3.6 Precision rate 3.7 Recall rate 3.8 Accuracy rate 3.9 Response time 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 datasets 4.5 Impact of software and hardware platform dependencies 4.6 Impact of environmental data 5. Evaluation process 6. Evaluation in the requirements phase 7. Evaluation in the design phase 8. Evaluation in the implementation phase 9. Evaluation in the operation phase Appendix A Appendix B

Development Information

Related Standards