I want to try a very simple case of learning using CNN (let's say digit recognition with 5000 pictures database). Each time I read a tutorial they speak about brand new GPU or linux server which I don't have. I juste have a simple windows laptop.
So my question is simple : what is the best way to run a "simple" CNN on basic hardware?
For simple examples like the one you mention, you should be able to train only on CPU and for that you can use Keras in Windows as its in Python.
Now for more involved dataset, either you buy a GPU (and install linux) or you use a cloud instance.
They are easy to set up, cheap if you just want to try stuff before making a decision about buying a GPU. Once they are set up you log in with ssh or interact with a notebook. In AWS, you can use g2 instance for cheap hardware or p2 instance for more expensive one (instance types). To set up all drivers, theano etc, you can use this image.
Some neural network libraries can be used with Microsoft Windows:
Use a library agnostic to the hardware.
Theano, for instance, will automatically switch from a CPU to a GPU if one is available. So, the code written will be exactly the same and you have nothing more to do than install the required libraries.
You can find examples of theano implementations of neural networks here : http://deeplearning.net/tutorial/lenet.html
You can use docker containers. However, feasibility depends on the windows version you are using.
