Hi

Padding helps you hold the information on the border of the feature or input matrix and also causes the input and output layer to have the same size. This problem that you face may be caused by your network or your data because it seems that your model can not learn. So you should explain more about the model and the kind of data you use so I can help you.

thank you for your reply!

so I'll try to be as clear as possible:

From a geometrical point of view I have a matrix of size In=N1*N2*L and I want to obtain a a matrix of size Out=N1*N2 ( so one dimension less than In). In this case N1=N2=20, L=3.

The relation between input-output is that (Matlab):

for t=1:N2

Out(:,t)= function(In(:,t,:))

end

this means that each row of the output is generated considering all elements on the first and third dimension and cycling over the second dimension of the input matrix.

For the CNN structure, in order to have input/ouput of the same dimensions my output has been modified in Out=(N1*N2)*L. So maybe this transformation is something somewhat undesirable? I did it also to avoid too heavy nn.

My net is a modified ResNet with LeakyReLU actiation function and Adam optimization algorithm....

many thanks!

Alessandro Salusti

Article Enlarging smaller images before inputting into convolutional...

Alessandro Salusti

Hi

The number of filters can indicate the number of output channels.

For example, if you have a 20*20*3 input and you want the output to be 20*20, you can use the CNN layer with filter=1( Conv2D(1, kernel_size, strides=(1, 1), padding='same') ). You can also use the depthwise function(

SeparableConv2D) with the desirable filters so that every channel be analyzed separately.

Keras tutorial can be useful:

https://keras.io/layers/convolutional/