why are we randomly adding any concvolutional layer with random filter sizes there must a logic behind adding any layer at any point .There must be a logic before defining filter size . There must be a proper way of adding layer with logical input size rather than adding anything anywhere with randomly defined filter size.
What is that proper/descent way?
Convolution layer
Hey @Par1hsharma ,
Convolution layers are used to learn features from the image, so more the features more better we learn it, but to an extent.
Now to get these features , you would have directly on the first convolution stated the filters as very very high value to get them , but in that case the model won’t learn much good things, in fact it will just create relations that are worthless.
Hence, we slowly slowly increase those number of filters to learn more details about the image.
Hence , moving in a way like 32 - 64 -128 - 256 -512 -1024 and so on…
these values are generally kept as power of 2 , but there is no such restriction. You can keep anything. You just need to learn from images about them.