Table 2.

Difference between the three inceptions.

Inception-v1	Inception-v2	Inception-v3
Increase the number of units at each stage and shielding the large number of input filters of the last stage to the next layers	Increase the learning rate, remove dropout and local response normalization, shuffle training examples more thoroughly, reduce the L2 weight regularization and the photometric distortions	Trained much faster compared to the other inception and method
Error rate = 6.67%	Error rate = 4.82%	Error rate = 3.5%