Mobile Net dimensions

Hello! The following is the architecture for “MobileNet”:
mobile_net

The input size DOES get halved when going from the row 4 conv_dw / s2 to the row 5 conv / s1 layer (112 x 112 x 64) to (56 x 56 x 64).

I don’t understand why the input size for the last conv / s1 layer is 7 x 7 x 1024 even though the layer before it (conv_dw / s2 layer) has a stride of 2. Shouldn’t the input size for the last conv / s1 layer be
3 x 3 x 1024 instead?

Based on the stride of 2 I would assume the same. However, since an avg pooling layer with a kernel size of 7x7 is used afterwards, I guess this particular layer should use a stride of 1?

Yeah, it works with a stride of 1. Probably a typo in the original MobileNet-v1 paper I suppose.