Home Empirical Findings ML Libraries
Post

Empirical Findings ML Libraries

This table presents empirical findings of RQ3. Among the 577 defective APIs, we find clear clues to the root causes of inducing pipelines’ performance inconsistencies, crashes and NaN bugs for 270 APIs (46.8%), in tde corresponding library release notes, issues and code commits. We identified four types of essential root causes and discuss them in two scenarios: performance impacts induced by one library version and performance impacts induced by a combination of library versions.
Root CausesCrashExecution TimeCompetition ScoreMemory UsageNaN
#Variants#Pipelines#Variants#Pipelines#Variants#Pipelines#Variants#Pipelines#Variants#Pipelines
Performance Impacts induced by one library versionAPI OptimizationHardware support--92572121--
Data Processing--501114701453713--
Calculations--675184291125415--
Otders--41281153--
Sum--1,309369162881832--
Default Hypermeter Change--30525217792
Technical Debt--47291922--
API Breaking Change61935------23612
Performance Impacts induced by a combination of library versionsIndirect Dependency Interference317991152254211163
Data Flow Interference1731,11424661195011643
help