I have a question regarding the neural net (or convolutional neural nets) architecture. There are many many parameters that can be changed, number of layers, number of node per layers, layer types (dropout, flatten, maxpooling, convolution etc..). How do we know what the architecture of a neural net should be or at least it's starting point?
there is no easy answer to this, but very recommended not to reinvent the wheel but start with some of the standard CNN architectures which have already proven to work well (and all the hyperparameters from there). for example ResNet18 or similar. here is some of our example code
in general, finding architectures automatically doesn't work so well and can be as bad as random search [1,2]
NN how to choose architecture
Hello,
I have a question regarding the neural net (or convolutional neural nets) architecture. There are many many parameters that can be changed, number of layers, number of node per layers, layer types (dropout, flatten, maxpooling, convolution etc..). How do we know what the architecture of a neural net should be or at least it's starting point?
Thank you for your help
3
there is no easy answer to this, but very recommended not to reinvent the wheel but start with some of the standard CNN architectures which have already proven to work well (and all the hyperparameters from there). for example ResNet18 or similar. here is some of our example code
in general, finding architectures automatically doesn't work so well and can be as bad as random search [1,2]
2
Add comment