void matlab::read_cast_matrix(mxArray *var, idx<T> &m) {
#ifdef __MATLAB__
// allocate a temporary matrix with same type as original matrix type
idx<Tmatlab> tmp(m.get_idxdim());
// load data
void *data = mxGetData(var);
// copy to idx
memcpy(m.idx_ptr(), (Tmatlab*) data, m.nelements() * sizeof (Tmatlab));
// copy-cast
idx_copy(tmp, m);
#endif
}
generic_conv_net(parameter &trainableParam,
intg output_size)
: layers_n<state_idx>(true) { // owns modules
cout << "Initializing ConvNet..." << endl;
//! Define the number of feature maps per layer (C0, C1, C2)
intg featureMaps0 = 6;
intg featureMaps1 = 12;
intg featureMaps2 = 40;
//! Define tables of connections between layers.
//! These two are fully connected layer, i.e. each feature map in a layer
//! is connected to every feature map in the previous layer
table0 = full_table(1, featureMaps0); //! from input to C0
table2 = full_table(featureMaps1, featureMaps2); //! from S1 to C2
//! ... whereas the connections there are sparse (S0 to C1):
table1 = idx<intg>(44, 2); //! from S0 to C1
intg tbl[44][2] =
{{0, 0}, {1, 0}, {2, 0}, //! 0,1,2 in S0 connected to 0 in C1
{1, 1}, {2, 1}, {3, 1}, //! and so on...
{2, 2}, {3, 2}, {4, 2},
{3, 3}, {4, 3}, {5, 3},
{4, 4}, {5, 4}, {0, 4},
{5, 5}, {0, 5}, {1, 5},
{0, 6}, {1, 6}, {2, 6}, {3, 6},
{1, 7}, {2, 7}, {3, 7}, {4, 7},
{2, 8}, {3, 8}, {4, 8}, {5, 8},
{3, 9}, {4, 9}, {5, 9}, {0, 9},
{4, 10}, {5, 10}, {0, 10}, {1, 10},
{0, 11}, {1, 11}, {2, 11}, {3, 11}, {4, 11}, {5, 11}};
memcpy(table1.idx_ptr(), tbl, table1.nelements() * sizeof (intg));
//! Finally we initialize the architecture of the ConvNet.
//! In this case we create a CSCSCF network.
//! It's easy to change the architecture, by simply removing/adding a call
//! to addModule(...)
//! C0 Layer
add_module(new nn_layer_convolution(trainableParam, //! Shared weights
7, 7, //! Dim of kernel
1, 1, //! size of subsampling
table0), //! Conx btwn input layer and C0
//! state_idx holds the feature maps of C0
new state_idx(featureMaps0,1,1));
//! S0 Layer
add_module(new nn_layer_subsampling(trainableParam,
2, 2, //! Dim of stride
2, 2, //! Dim of subsampling mask
featureMaps0),
new state_idx(featureMaps0,1,1));
//! C1 Layer
add_module(new nn_layer_convolution(trainableParam,
7, 7,
1, 1,
table1),
new state_idx(featureMaps1,1,1));
//! S1 Layer
add_module(new nn_layer_subsampling(trainableParam,
2, 2,
2, 2,
featureMaps1),
new state_idx(featureMaps1,1,1));
//! C2 Layer
add_module(new nn_layer_convolution(trainableParam,
7, 7,
1, 1,
table2),
new state_idx(featureMaps2,1,1));
//! F Layer
add_last_module(new nn_layer_full(trainableParam,
featureMaps2,
output_size));
}