ModelCI-e: Enabling Continual Learning in Deep Learning Serving Systems Yizheng Huang∗ Huaizheng Zhang∗ Yonggang Wen Nanyang Technological University Nanyang Technological University Nanyang Technological University
[email protected] [email protected] [email protected] Peng Sun Nguyen Binh Duong TA SenseTime Group Limited Singapore Management University
[email protected] [email protected] ABSTRACT streamline model deployment and provide optimizing mechanisms MLOps is about taking experimental ML models to production, (e.g., batching) for higher efficiency, thus reducing inference costs. i.e., serving the models to actual users. Unfortunately, existing ML We note that existing DL serving systems are not able to handle serving systems do not adequately handle the dynamic environ- the dynamic production environments adequately. The main reason ments in which online data diverges from offline training data, comes from the concept drift [8, 10] issue - DL model quality is resulting in tedious model updating and deployment works. This tied to offline training data and will be stale soon after deployed paper implements a lightweight MLOps plugin, termed ModelCI-e to serving systems. For instance, spam patterns keep changing (continuous integration and evolution), to address the issue. Specif- to avoid detection by the DL-based anti-spam models. In another ically, it embraces continual learning (CL) and ML deployment example, as fashion trends shift rapidly over time, a static model techniques, providing end-to-end supports for model updating and will result in an inappropriate product recommendation. If a serving validation without serving engine customization. ModelCI-e in- system is not able to quickly adapt to the data shift, the quality of cludes 1) a model factory that allows CL researchers to prototype inference would degrade significantly.