can1357/simple_cnn

Simple Convolutional Neural Network Library

simple_cnn

simple_cnn is ment to be an easy to read and easy to use convolutional neural network library.

simple_cnn is written in a mostly C-like manner behind the scenes, doesnt use virtual classes and avoids using std where its possible so that it is easier to convert to CUDA code when needed.

Example use on handwritten digit recognition (Youtube Video):

Building

On linux, run make.

MNIST digits taken from http://yann.lecun.com/exdb/mnist/

simple_cnn's Issues

Save network

Hi is there a possibility to save the trained nn (all the neuron layers and synapses) to a file?
I'm thinking about something like the fann library does.
So if you want to use it again you just load the file instead of having to do the whole training again.

There is no multiplication feature if two fully connected layers are there..

I mean if there are two fully connected layer with 12 and 5 neurons.
You have to multiply weights with fully connected layer, Let say 12 x 1 layer with 12 x 5 weight matrix or tensor.. to get next layer which is 5 x 1..

how can i compile it ?

can i compile it and run it like youtube demo ?
if can't ,how do i make it .
thanks

divergence issue

When the number of layers has been increased like 2 conv layer, the network diverges and no matter what the learning rate is. I also added learning rate decay but it diverged again. I think there is another problem somewhere.

Number of channels

Hello,
I have a question regarding how to handle number of channels with this CNN implementation. For e.g., RGB images would require three channels but I could not find any part to handle the number of channels.

Missing "test.ppm"

There is no "test.ppm" in the project. Can you share one with me? Thank you!

linux port

Hello,

I was just looking over your "simple_cnn" code and noticed that it seem to be set up to compile for Windows.

I was wondering if you could port this version, or have a version that will compile up on Linux (Ubuntu) so that I do not have to dig through all of the code trying to get it to work on my Linux box?

I wanted to use your code to learn more about CNN's and how they work in a simple CPU (non-GPU) version before moving to more complex libraries.

Cheers,
Lonnie