Registered user since Fri 19 Mar 2021
PhD student at SMU (:
Contributions
Research Papers
Thu 13 Oct 2022 16:40 - 17:00 at Ballroom C East - Technical Session 29 - AI for SE II Chair(s): Tim MenziesAlthough large pre-trained models of code have delivered significant advancements in various code processing tasks, there is an impediment to the wide and fluent adoption of these powerful models in the daily workflow of software developers: these large models consume hundreds of megabytes of memory and run slowly especially on personal devices, which causes problems in model deployment and greatly degrades the user experience.
It motivates us to propose Compressor, a novel approach that can compress the pre-trained models of code into extremely small models with negligible performance sacrifice. Our proposed method formulates the design of tiny models as simplifying the pre-trained model architecture: searching for a significantly smaller model that follows an architectural design similar to the original pre-trained model. To tackle this problem, Compressor proposes a genetic algorithm (GA)-based strategy to guide the simplification process. Prior studies found that a model with higher computational cost tends to be more powerful. Inspired by this insight, the GA algorithm is designed to maximize a model’s Giga floating-point operations (GFLOPs), an indicator of the model computational cost, to satisfy the constraint of the target model size. Then, we use the knowledge distillation technique to train the small model: unlabelled data is fed into the large model and the outputs are used as labels to train the small model. We evaluate Compressor with two state-of-the-art pre-trained models, i.e., CodeBERT and GraphCodeBERT, on two important tasks, i.e, vulnerability prediction and clone detection. We use the proposed method to compress models to a size (3 MB), which is only 0.6% of the original model size. The results show that compressed CodeBERT and GraphCodeBERT reduce the inference latency by 70.75% and 79.21%, respectively. More importantly, they maintain 96.15% and 97.74% of the original performance on the vulnerability prediction task. They even maintain higher ratios (99.20% and 97.52%) of the original performance on the clone detection task.
DOI Pre-print