conv neural network - How to understand the convolution parameters in tensorflow? -

when reading chapter of "deep mnist expert" in tensorflow tutorial.

there give below function weight of first layer. can't understand why patch size 5*5 , why features number 32, random numbers can pick or rules must followed? , whether features number "32" “convolution kernel”？

w_conv1 = weight_variable([5, 5, 1, 32])

first convolutional layer

we can implement our first layer. consist of convolution, followed max pooling. convolutional compute 32 features each 5x5 patch. weight tensor have shape of [5, 5, 1, 32]. first 2 dimensions patch size, next number of input channels, , last number of output channels. have bias vector component each output channel.

the patch size , number of features network hyper-parameters, therefore arbitrary.

there rules of thumb, way, follow in order define working , performing network. kernel size should small, due equivalence between application of multiple small kernels , lower number of big kernels (it's image processing topic , it's explained in vgg paper). in addiction, operations small filters way faster execute.

the number of features extract (32 in example) arbitrary , find right number somehow art.