check that when we unroll, we unroll with a minimim value of 1. removed the two useless generation fct that generate the unroll only by batch or by kern. We have a fct that unroll by both of them. Set the unroll value to one to have the equivalent fct.

added a new unrolled verion of ConvOp(not used by default). It unroll the batch and the kernel at the same time. This give the biggest speed up in my test. Also modified the test_multilayer_conf fct to allow testing the different parameter for the unroll.

added the unroll version by kern. If you put the ConvOp option unroll_batch or unroll_kern>0, we will those version with the provided unrool size.

added option to ConvOp that allow using the unroll version of the code. The default is to don't use this version of the code.

added a function to generate an unrolled version of the c code for the ConvOp. The code is not used.

scipy.signal.convolve2d causes errors on atexit. i moved imports to make it possible to use our convolve op without triggering this problem

self). Also did some housecleaning

* typo in c_code (comparing type instead of type.dtype)
* added test case for 3D input tensor in test_ConvOpGrad

Raise a NotImplementedError.

This is because the gradient on the weights is obtained by convoling the visible
layer with the hidden layer (as the kernel). Problem is the convolution code
does not support kernels larger than the input image (which is the case in full
mode). Either fix ConvOp to allow this, or maybe convolve the hidden with the
visible as the kernel, then DimShuffle at the end.

multiple input features