Adding support for (1) rectangular images and (2) user-defined linking and feeding weight matrices

[madratman]

(Maybe this should be broken into two issues)

• To use GRID_FOUR or GRID_EIGHT type connectivity, one needs to restrict to a perfect square number of (pcnn at least) oscillators.
  This restrics the scope of the library(at least pcnn, which I am using) to (cropped) square images.
  Here's the relevant code block from network.cpp:

void network::create_grid_four_connections() {
    const double conv_side_size = std::sqrt((double) num_osc);
    if (conv_side_size - std::floor(conv_side_size) > 0) {
        throw std::runtime_error("Invalid number of oscillators in the network for the grid structure");
    }

And of course, making type GRID_EIGHT calls the previous method only. code block

https://github.com/annoviko/pyclustering/blob/master/ccore/ccore/network.cpp#L236

• The feeding and linking weight "matrix" in pcnn (haven't checked out other networks) is defined just as a single integer. Hence one can only have an array of equal weights.
  This should be improved to allow users to define custom linkages like gaussian, or inverse of distance.
  [code block 1](https://github.com/annoviko/pyclustering/blob/master/ccore/ccore/pcnn.h#L34 -L35)

typedef struct pcnn_parameters {
...
    double W = 1.0;
    double M = 1.0;

code block 2

        for (std::vector<unsigned int>::const_iterator iter = neighbors.begin(); iter != neighbors.end(); iter++) {
            const double output_neighbor = m_oscillators[(*iter)].output;

            feeding_influence += output_neighbor * m_params.M;
            linking_influence += output_neighbor * m_params.W;
        }

[annoviko]

I think the first part of the issue has been resolved:
eab634c

Thank you for your pull request. It has been merged with some changes: added the same support to other oscillatory networks, added unit-tests.

[annoviko]

Anyway such feature (the first part) should be also added to the pyclustering (python) implementation. I will take care about it.

[annoviko]

Support of rectangular structures and images by base class network and PCNN is implemented: d9dde07 .

[annoviko]

Support of weight matrices is added. Adjacency class collection can be used for that purpose:

• adjacency_weight_list
• adjacency_matrix

Both implement adajcency interface.