Regression testing
dblalock opened this issue · comments
dblalock commented
We should automatically do full training runs for all of the benchmarks sometimes to detect accuracy regressions. This will be more challenging than regular testing since we'll have to:
- Launch multi-GPU jobs
- Monitor the results
- Track previous results as a reference
This should actually be pretty good MCLI dogfooding.
Daniel King commented
Closing as we're not tracking this here.