We converted several eager execution function to hybrid execution as part of an assessment of our refactoring tool. The transformations produced were completely automated. We have some preliminary evidence that this improves the run-time performance:
Test
Python version
TensorFlow version
Epochs
Before accuracy
After accuracy
Before loss
After loss
Before elapsed time (s)
After elapsed time (s)
Speedup
MusicTransformer-tensorflow2.0/train.py
3.10.0
2.9.3
5
0.0233113606
0.02381204049
4.956358294
4.923633542
1330.180665
919.3014389
1.446947224
The above analysis was repeated five times for each program version (before and after the refactoring) and the values were averaged. The increased speedup is ~1.45 with seemingly negligible loss of accuracy. We would appreciate any feedback you may have to help us as assess our refactoring approach. Thank you for your time!
We converted several eager execution function to hybrid execution as part of an assessment of our refactoring tool. The transformations produced were completely automated. We have some preliminary evidence that this improves the run-time performance:
The above analysis was repeated five times for each program version (before and after the refactoring) and the values were averaged. The increased speedup is ~1.45 with seemingly negligible loss of accuracy. We would appreciate any feedback you may have to help us as assess our refactoring approach. Thank you for your time!