A deep-learning program can produce the right answer while wasting time or memory. DeepPerf connects a study of real performance problems with reproducible examples and checks that developers can use on their own code.
I am the first author of the ESEC/FSE 2022 paper. I studied 224 performance problems, built a 58-case reproducible benchmark, and developed DeepPerf, a rule-based static checker.
From reported problems to usable checks
The empirical study examined how performance problems arise and the patterns behind them. The benchmark turns a subset of those reports into runnable examples, making the problems easier to investigate and compare.
DeepPerf examines source code without running the program. Its rules target recurring performance problems in TensorFlow and Keras code, including inefficient data processing. This makes the checker useful for finding specific known patterns while keeping its coverage explicit.
Findings and fixes
Evaluation on 1,108 open-source repositories identified 488 new issues, of which 27 were fixed. The paper and project artifacts describe the study, benchmark, and checker.
I also contributed data-pipeline parallelisation fixes derived from the findings. These were merged into Keras #15295 and TensorFlow Agents #650, connecting the research to changes adopted by open-source maintainers.