kmeans(means, clusters, data, k) data: input, 1D or 2D (range > [0-1]) k: input, # desired means (k > 1) means: output, vector of means
| 59 | // k: input, # desired means (k > 1) |
| 60 | // means: output, vector of means |
| 61 | void kmeans(array &means, array &clusters, const array in, int k, |
| 62 | int iter = 100) { |
| 63 | unsigned n = in.dims(0); // Num of data points |
| 64 | unsigned d = in.dims(2); // Num of features (will only be 1 in spider image example) |
| 65 | |
| 66 | // reshape input |
| 67 | array data = in * 0; |
| 68 | |
| 69 | // re-center and scale down data to [0, 1] |
| 70 | array minimum = min(in); |
| 71 | array maximum = max(in); |
| 72 | |
| 73 | gfor(seq ii, d) { |
| 74 | data(span, span, ii) = |
| 75 | (in(span, span, ii) - minimum(ii).scalar<float>()) / |
| 76 | maximum(ii).scalar<float>(); |
| 77 | } |
| 78 | |
| 79 | // Initial guess of means |
| 80 | means = randu(1, k, d); |
| 81 | array curr_clusters = constant(0, data.dims(0)) - 1; |
| 82 | array prev_clusters; |
| 83 | |
| 84 | // Stop updating after specified number of iterations |
| 85 | for (int i = 0; i < iter; i++) { |
| 86 | // Store previous cluster ids |
| 87 | prev_clusters = curr_clusters; |
| 88 | |
| 89 | // Get cluster ids for current means |
| 90 | curr_clusters = clusterize(data, means); |
| 91 | |
| 92 | // Break early if clusters not changing |
| 93 | unsigned num_changed = count<unsigned>(prev_clusters != curr_clusters); |
| 94 | |
| 95 | if (num_changed < (n / 1000) + 1) break; |
| 96 | |
| 97 | // Update current means for new clusters |
| 98 | means = new_means(data, curr_clusters, k); |
| 99 | } |
| 100 | |
| 101 | // Scale up means |
| 102 | gfor(seq ii, d) { |
| 103 | means(span, span, ii) = |
| 104 | maximum(ii) * means(span, span, ii) + minimum(ii); |
| 105 | } |
| 106 | |
| 107 | clusters = prev_clusters; |
| 108 | } |
| 109 | |
| 110 | // K-Means image recoloring. |
| 111 | // Shifts the hues of an image to the k mean hues. |
no test coverage detected