:return: mean_IoU, IoU_array, pixel_acc, mean_acc
(self)
| 186 | |
| 187 | |
| 188 | def evaluate(self): |
| 189 | """ |
| 190 | |
| 191 | :return: mean_IoU, IoU_array, pixel_acc, mean_acc |
| 192 | """ |
| 193 | |
| 194 | if self._distributed: |
| 195 | link.synchronize() |
| 196 | |
| 197 | conf_matrix_list = self.all_gather(self._conf_matrix) |
| 198 | # self._predictions = self.all_gather(self._predictions) |
| 199 | # self._predictions = list(itertools.chain(*self._predictions)) |
| 200 | if link.get_rank() != 0: |
| 201 | return |
| 202 | |
| 203 | self._conf_matrix = np.zeros_like(self._conf_matrix) |
| 204 | for conf_matrix in conf_matrix_list: |
| 205 | self._conf_matrix += conf_matrix |
| 206 | |
| 207 | # if self._output_dir: |
| 208 | # os.makedirs(self._output_dir, exist_ok=True) |
| 209 | # file_path = os.path.join(self._output_dir, "humam_parsing_predictions.json") |
| 210 | # with open(file_path, "w") as f: |
| 211 | # f.write(json.dumps(self._predictions)) |
| 212 | |
| 213 | acc = np.full(self._num_classes, np.nan, dtype=np.float) |
| 214 | iou = np.full(self._num_classes, np.nan, dtype=np.float) |
| 215 | tp = self._conf_matrix.diagonal().astype(np.float) |
| 216 | pos_gt = np.sum(self._conf_matrix, axis=0).astype(np.float) |
| 217 | # class_weights = pos_gt / np.sum(pos_gt) |
| 218 | pos_pred = np.sum(self._conf_matrix, axis=1).astype(np.float) |
| 219 | acc_valid = pos_gt > 0 |
| 220 | acc[acc_valid] = tp[acc_valid] / pos_gt[acc_valid] |
| 221 | iou_valid = (pos_gt + pos_pred) > 0 |
| 222 | union = pos_gt + pos_pred - tp |
| 223 | iou[acc_valid] = tp[acc_valid] / union[acc_valid] |
| 224 | macc = np.sum(acc[acc_valid]) / np.sum(acc_valid) |
| 225 | miou = np.sum(iou[acc_valid]) / np.sum(iou_valid) |
| 226 | # fiou = np.sum(iou[acc_valid] * class_weights[acc_valid]) |
| 227 | pacc = np.sum(tp) / np.sum(pos_gt) |
| 228 | |
| 229 | res = {} |
| 230 | res["mIoU"] = 100 * miou |
| 231 | # res["fwIoU"] = 100 * fiou |
| 232 | for i, name in enumerate(self._class_names): |
| 233 | res["IoU-{}".format(name)] = 100 * iou[i] |
| 234 | res["mACC"] = 100 * macc |
| 235 | res["pACC"] = 100 * pacc |
| 236 | for i, name in enumerate(self._class_names): |
| 237 | res["ACC-{}".format(name)] = 100 * acc[i] |
| 238 | |
| 239 | if self._output_dir: |
| 240 | file_path = os.path.join(self._output_dir, "human_parsing_evaluation.pth") |
| 241 | with open(file_path, "wb") as f: |
| 242 | torch.save(res, f) |
| 243 | results = OrderedDict({"human_parsing": res}) |
| 244 | self._logger.info(results) |
| 245 | return results |
nothing calls this directly
no test coverage detected