eval mIoU func
(args)
| 146 | |
| 147 | |
| 148 | def eval(args): |
| 149 | """ |
| 150 | eval mIoU func |
| 151 | """ |
| 152 | # DataLoader need run on cpu |
| 153 | paddle.set_device("cpu") |
| 154 | devices = paddle.device.get_device().split(":")[0] |
| 155 | |
| 156 | val_loader = build_dataloader(all_config, "Eval", devices, logger) |
| 157 | post_process_class = build_post_process(all_config["PostProcess"], global_config) |
| 158 | eval_class = build_metric(all_config["Metric"]) |
| 159 | model_type = global_config["model_type"] |
| 160 | |
| 161 | predictor, rerun_flag = load_predictor(args) |
| 162 | |
| 163 | if rerun_flag: |
| 164 | eval_dataset = find_images_with_bounding_size(val_loader.dataset) |
| 165 | batch_sampler = paddle.io.BatchSampler( |
| 166 | eval_dataset, batch_size=1, shuffle=False, drop_last=False |
| 167 | ) |
| 168 | val_loader = paddle.io.DataLoader( |
| 169 | eval_dataset, batch_sampler=batch_sampler, num_workers=4, return_list=True |
| 170 | ) |
| 171 | |
| 172 | input_names = predictor.get_input_names() |
| 173 | input_handle = predictor.get_input_handle(input_names[0]) |
| 174 | output_names = predictor.get_output_names() |
| 175 | output_handle = predictor.get_output_handle(output_names[0]) |
| 176 | sample_nums = len(val_loader) |
| 177 | predict_time = 0.0 |
| 178 | time_min = float("inf") |
| 179 | time_max = float("-inf") |
| 180 | print("Start evaluating ( total_iters: {}).".format(sample_nums)) |
| 181 | |
| 182 | for batch_id, batch in enumerate(val_loader): |
| 183 | images = np.array(batch[0]) |
| 184 | |
| 185 | batch_numpy = [] |
| 186 | for item in batch: |
| 187 | batch_numpy.append(np.array(item)) |
| 188 | |
| 189 | # ori_shape = np.array(batch_numpy).shape[-2:] |
| 190 | input_handle.reshape(images.shape) |
| 191 | input_handle.copy_from_cpu(images) |
| 192 | start_time = time.time() |
| 193 | |
| 194 | predictor.run() |
| 195 | preds = output_handle.copy_to_cpu() |
| 196 | |
| 197 | end_time = time.time() |
| 198 | timed = end_time - start_time |
| 199 | time_min = min(time_min, timed) |
| 200 | time_max = max(time_max, timed) |
| 201 | predict_time += timed |
| 202 | |
| 203 | if model_type == "det": |
| 204 | preds_map = {"maps": preds} |
| 205 | post_result = post_process_class(preds_map, batch_numpy[1]) |
no test coverage detected
searching dependent graphs…