I'm trying to implement what is explained in a paper on audio signal processing. The guys who wrote this paper tried a Convolutional Neural Network and here is how they explain it :
"The CNN are constructed using two different building blocks: block A consists of two layers with
32 3*3 filters, and block
B consists of two layers with
64 3*3 filters ; both in combination with batch normalization, and each followed by a
3*3 max-pooling layer and a drop-out layer."
They state that they apply this model to
84*9 images (actually it's data from a spectogram,
84 is a number of frequency bins,
9 is the number of frames)
I have some trouble to find out precisely the different layers to add to the model :
- Does the Batch Normalization layer occurs after each convolutional layer ?
- Does the max-pooling layer occurs after each convolutional layer ?
- Does the drop-out layer occurs after each convolutional layer ?
- Is there any implicit padding ? Implicit stride ?
My guess is : if they say they are using Batch Normalization, max-pooling and drop-out, then maybe it is implemented after each Convolutional layer (the four of them) (I've often seen that it's applied after each layer)
But with those layers it's impossible to have an input_shape of
(84, 9) with no padding ...
I'm quite stuck.
Any ideas ?
Thanks a lot :)