EfficientNet： when compound coefficient d is demical, what the nerual network looks like?

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

I can understand when when the w and r is demical ,the network channel and resolution will multiple by w and r respectively .But I am a little confused, when d is demical , what does it mean.
(after reading the paper “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks”, I still can not understand)
Thanks

2

Hi xiaobaihe,

Thanks for your question. I am not sure if I understand what you mean with the word ‘demical’. Could you maybe rephrase your question?

3

Hi reinoudbosch,
what I mean is that :
if the d(Scaling Dimensions Depth) is integer，for example d = 2，then it will scale network by double all the Operator Fi in stage i .But when d=1.2， how to scale the each stage of network.

4

Hi xiaobaihe,

That is a great question. As it is hard to find any discussion concerning this point, my interpretation is that the base model allows for a multiplication by 1.2 to still arrive at a depth that can be expressed as an integer (but not times the base model). E.g. the base model consists of 5 units of components sequentially. This can then be adjusted to 6 units of components sequentially (which is 1.2 times the base model).

Does this make sense to you?

5

this makes sense. thanks.
And it seems that the one unit which is added , could be the first units of original components sequentially