| 131 | }; |
| 132 | |
| 133 | int rbm_demo(bool /*console*/, int perc) { |
| 134 | printf("** ArrayFire RBM Demo **\n\n"); |
| 135 | |
| 136 | array train_images, test_images; |
| 137 | array train_target, test_target; |
| 138 | int num_classes, num_train, num_test; |
| 139 | |
| 140 | // Load mnist data |
| 141 | float frac = (float)(perc) / 100.0; |
| 142 | setup_mnist<true>(&num_classes, &num_train, &num_test, train_images, |
| 143 | test_images, train_target, test_target, frac); |
| 144 | |
| 145 | dim4 dims = train_images.dims(); |
| 146 | |
| 147 | int feature_size = train_images.elements() / num_train; |
| 148 | |
| 149 | // Reshape images into feature vectors |
| 150 | array train_feats = moddims(train_images, feature_size, num_train).T(); |
| 151 | array test_feats = moddims(test_images, feature_size, num_test).T(); |
| 152 | |
| 153 | train_target = train_target.T(); |
| 154 | test_target = test_target.T(); |
| 155 | |
| 156 | rbm network(train_feats.dims(1), 2000); |
| 157 | |
| 158 | network.train(train_feats, |
| 159 | 0.1, // learning rate |
| 160 | 15, // num epochs |
| 161 | 100, // batch size |
| 162 | 1, // k |
| 163 | true); |
| 164 | |
| 165 | // Test reconstructed images |
| 166 | for (int ii = 0; ii < 5; ii++) { |
| 167 | array in = test_feats(ii, span); |
| 168 | array res, tmp; |
| 169 | |
| 170 | network.gibbs_vhv(res, tmp, in); |
| 171 | |
| 172 | in = moddims(in, dims[0], dims[1]); |
| 173 | res = moddims(res, dims[0], dims[1]); |
| 174 | |
| 175 | in = round(in); |
| 176 | res = round(res); |
| 177 | |
| 178 | printf("Reconstructed Error for image %2d: %.4f\n", ii, |
| 179 | sum<float>(abs(in - res)) / feature_size); |
| 180 | } |
| 181 | |
| 182 | return 0; |
| 183 | } |
| 184 | |
| 185 | int main(int argc, char **argv) { |
| 186 | int device = argc > 1 ? atoi(argv[1]) : 0; |