Convert several eager execution function to hybrid execution. We have some preliminary evidence that this improves the run-time performance of the models:
Test
Python version
TF version
Before accuracy
After accuracy
Before loss
After loss
Before elapsed time (s)
After elapsed time (s)
Speedup
neural_network
3.10.0
2.9.3
0.9624
0.9635
9.333428266
3.812376191
2.448191836
autoencoder
3.10.0
2.9.3
0.006999
0.007014
110.4210886
34.1057281
3.23761124
logistic_regression
3.10.0
2.9.3
0.8286328125
0.8316015625
0.918056736
0.9068725395
1.415692188
0.7934420485
1.784241446
bidirectional_rnn
3.10.0
2.9.3
0.85625
0.821875
0.5128818989
0.58627882
28.0902812
5.041457747
5.571856913
convolutional_network
3.10.0
2.9.3
0.9867734375
0.9869921875
1.48369342
1.483417908
31.07854785
17.71073562
1.754785827
dcgan
3.10.0
2.9.3
1.208782502
0.04901289759
78.12116778
36.0855548
2.164887535
dynamic_rnn
3.10.0
2.9.3
0.8580357143
0.8657738095
0.3000548454
0.285470572
48.15241052
8.490720483
5.671180745
recurrent_network
3.10.0
2.9.3
0.9375
0.93125
0.1873067699
0.2336050078
42.15870964
7.818362872
5.39226822
build_custom_layers
3.10.0
2.9.3
0.907109375
0.919921875
3.339067001
3.328515396
1.387739662
0.843345543
1.645517277
save_restore_model
3.10.0
2.9.3
0.8957291667
0.8922395833
107.3751221
110.7431885
4.18468971
1.885201852
2.219756842
tensorboard_example
3.10.0
2.9.3
0.872734375
0.8712239583
110.2372933
112.0809294
8.789215875
4.572643171
1.922130275
For dcgan, we believe that the difference in loss is due to a TF bug that is still present in 2.15.0. This test can be reverted if desired.
Convert several eager execution function to hybrid execution. We have some preliminary evidence that this improves the run-time performance of the models:
For
dcgan, we believe that the difference in loss is due to a TF bug that is still present in 2.15.0. This test can be reverted if desired.