Keras: What Filters are used in Convolution Neural Network?

[bhaveshoswal]

I am not able get the "nb_filters" in CNN 2D. i am doing a image classification using CNN 2D so i have to nb_filters argument in it for convolution purpose so suppose filter size is 3 * 3 and i want know what values does it contains? how it is decided for all 'n' filters

[farizrahman4u]

nb_filters sands for number of filters. If you set it to 10 for instance, 10 3*3 filters will be learned.

[bhaveshoswal]

thanks. you are right but what each instance contains and how it is decided?

[keunwoochoi]

what each instance contains

Save weights of model, open it as a hdf, then open each layer, check the values.

how it is decided?

What do you mean? Is Backprop what you're asking?

[bhaveshoswal]

thanks No keunwoochoi I meant that the filters which we take in 'nb_filter' argument in Convolution Neural Network is like '32, 3, 3' so what that 32 filters contains and how it is decided

[keunwoochoi]

That sounds like I answered correctly.
32 means 32 filters each of which size of 3x3. The values are decided so that it can represent what it needs to do and by backprop algorithm. I know I'm almost repeating the same answer but that's what your questions are asking I think.

[KeironO]

Hey @bhaveshoswal,

I wrote a short introduction paper about ConvNets, that you can read here.

https://www.researchgate.net/publication/285164623_An_Introduction_to_Convolutional_Neural_Networks

If you have any further questions feel free to ask.

Keiron

[bhaveshoswal]

thanks but this pdf gives a better and good explanation about filter value initializations must look

[Chang-Hoon]

above mentioned pdf has been moved to http://cs231n.stanford.edu/slides/2016/winter1516_lecture7.pdf

[tispratik]

Are the 32 filters of the same type... just 32 instances? Or are they 32 different types of filters?
Example of "types" of a filters are "Sharpen", "Blur", "Edge detect", "Emboss", etc.

[amay1212]

@tispratik ,

so this is how it works , first your image gets divided into matrix , kernel 33 with 32 filters mean ,
the combination of filter maps you will get while multiplying the kernel size matrix i.e(33) with the matrix of your input let say image and let suppose stride=1 , so the filter maps will be consisting of the product of two with a difference of 1 stride(assumption).

Hope this helps!!