Python Deep Learning - Second Edition

Inception v1

Here, we can see the first version of the inception block, part of the GoogLeNet network architecture. GoogLeNet contains nine such inception blocks; we can see them in the following diagram:

Inception v1 block

The v1 block has four paths:

• 1 x 1 convolution, which acts as a kind of repeater to the input
• 1 x 1 convolution, followed by a 3 x 3 convolution
• 1 x 1 convolution, followed by a 5 x 5 convolution
• 3 x 3 max pooling with a stride of 1

The layers in the block use padding in such a way that the input and the output have the same shape (but different depths). The padding is also necessary, because each path would produce output ...