| 2833 | // ------------------------------------------------------------------------------------ |
| 2834 | |
| 2835 | void img2col( |
| 2836 | matrix<float>& output, |
| 2837 | const tensor& data, |
| 2838 | long n, |
| 2839 | long filter_nr, |
| 2840 | long filter_nc, |
| 2841 | long stride_y, |
| 2842 | long stride_x, |
| 2843 | long padding_y, |
| 2844 | long padding_x |
| 2845 | ) |
| 2846 | { |
| 2847 | const auto d = data.host() + data.k()*data.nr()*data.nc()*n; |
| 2848 | const rectangle boundary = get_rect(data); |
| 2849 | |
| 2850 | const long out_nr = 1+(data.nr()+2*padding_y-filter_nr)/stride_y; |
| 2851 | const long out_nc = 1+(data.nc()+2*padding_x-filter_nc)/stride_x; |
| 2852 | |
| 2853 | output.set_size(out_nr*out_nc, |
| 2854 | data.k()*filter_nr*filter_nc); |
| 2855 | DLIB_CASSERT(output.size() != 0); |
| 2856 | float* t = &output(0,0); |
| 2857 | |
| 2858 | // now fill in the Toeplitz output matrix for the n-th sample in data. |
| 2859 | long cnt = 0; |
| 2860 | const long max_r = data.nr() + padding_y-(filter_nr-1); |
| 2861 | const long max_c = data.nc() + padding_x-(filter_nc-1); |
| 2862 | for (long r = -padding_y; r < max_r; r+=stride_y) |
| 2863 | { |
| 2864 | for (long c = -padding_x; c < max_c; c+=stride_x) |
| 2865 | { |
| 2866 | for (long k = 0; k < data.k(); ++k) |
| 2867 | { |
| 2868 | for (long y = 0; y < filter_nr; ++y) |
| 2869 | { |
| 2870 | for (long x = 0; x < filter_nc; ++x) |
| 2871 | { |
| 2872 | DLIB_ASSERT(cnt < output.size()); |
| 2873 | long xx = c+x; |
| 2874 | long yy = r+y; |
| 2875 | if (boundary.contains(xx,yy)) |
| 2876 | *t = d[(k*data.nr() + yy)*data.nc() + xx]; |
| 2877 | else |
| 2878 | *t = 0; |
| 2879 | ++t; |
| 2880 | ++cnt; |
| 2881 | } |
| 2882 | } |
| 2883 | } |
| 2884 | } |
| 2885 | } |
| 2886 | } |
| 2887 | |
| 2888 | void col2img( |
| 2889 | const matrix<float>& output, |