CNN questions and discussion

rasmusbergpalm / DeepLearnToolbox

Matlab/Octave toolbox for deep learning. Includes Deep Belief Nets, Stacked Autoencoders, Convolutional Neural Nets, Convolutional Autoencoders and vanilla Neural Nets. Each method has examples to get you started.

BSD 2-Clause "Simplified" License

3.78k stars 2.28k forks source link

CNN questions and discussion #92

Open mawady opened 10 years ago

mawady commented 10 years ago

Hii

I'm using this code in my masters thesis and have some points to discuss:

Learning rate: initial is too high (1 as default) and doesn't decrease by epochs
i changed the code [images attached] in some how to deal with multi-channel data (RGB, satellite,... ) but i'm not sure about it
in cnnnumgradcheck.m file, the commented code of sampling layer can be implemented in proper way or it's not important at all

Thanks for your time, Mohamed Elawady Heriot-Watt University

cnntrain2 cnnff1 cnnsetup1 cnntrain1

taygunk commented 10 years ago

Using global variables is a bad practice. What you should do is to inspect the input shape, and come up with choosing correct batchsize and numberOfChannels from the input data.

mawady commented 10 years ago

Thanks @taygunk for your reply

the presented code is just quick modification for original modification ... i already organized my implementation using some coding standards! :+1:

my problem is not selecting batch size or number of channels but the way i modify the code to accept N-dimensional data (not technical aspect), reason of not using adaptive learning rate, and status of cnnnumgradcheck.m (not called by any m-file!)

I already met some former PhD students of hinton (Ruslan & Volod) last week and already discussed my conceptual ambiguity and got quite reasonable answers for selecting CNN parameters :)