Python Deep Learning - Second Edition

Inception v2 and v3

Inception v2 and v3 were released together and propose several improvements over the original inception block. The first is the factorization of the 5 x 5 convolution in two stacked 3 x 3 convolutions. We discussed the advantages of this in the VGG section. We can see the new inception block in the following diagram:

Inception block A

The next improvement is the factorization of an nxn convolution in two stacked asymmetrical 1xn and nx1 convolutions. For example, we can split a single 3 x 3 convolution into two 1 x 3 and 3 x 1 convolutions, where the 3 x 1 convolution is applied over the output of the 1 x 3 convolution. ...