boostcampaitech4lv23cv1 / level2_objectdetection_cv-level2-cv-05 Goto Github PK

주어진 파일 그대로 값수정안하고 돌렸을때 나오는 문제입니다.
박스가 밀려서 좌표 부분 값들에 2배씩 해줘야 정상적인 값들이 나오네요

현재 mmdetection에서 seed에 따른 재현성에 문제가 있는 것 같습니다.
config에서 seed 설정을 해도 동일한 코드에서 다른 결과가 나옵니다.

open-mmlab/mmdetection#2558

위에 따르면 set_random_seed(seed, deterministic=True)를 통해 seed를 고정할 수 있지만
cuda 문제로 작은 차이가 발생할 수 있다고 합니다.

2022-11-23 07:53:17,945 - mmdet - WARNING - No meta information found in the runner.

Exception Traceback (most recent call last)
/tmp/ipykernel_7165/4164586547.py in
1 # 모델 학습
----> 2 train_detector(model, datasets[0], cfg, distributed=False, validate=True)

~/baseline/mmdetection/mmdet/apis/train.py in train_detector(model, dataset, cfg, distributed, validate, timestamp, meta)
242 elif cfg.load_from:
243 runner.load_checkpoint(cfg.load_from)
--> 244 runner.run(data_loaders, cfg.workflow)

/opt/conda/envs/detection/lib/python3.7/site-packages/mmcv/runner/epoch_based_runner.py in run(self, data_loaders, workflow, max_epochs, **kwargs)
115 self.logger.info('workflow: %s, max: %d epochs', workflow,
116 self._max_epochs)
--> 117 self.call_hook('before_run')
118
119 while self.epoch < self._max_epochs:

/opt/conda/envs/detection/lib/python3.7/site-packages/mmcv/runner/base_runner.py in call_hook(self, fn_name)
315 """
316 for hook in self._hooks:
--> 317 getattr(hook, fn_name)(self)
318
319 def get_hook_info(self) -> str:

/opt/conda/envs/detection/lib/python3.7/site-packages/mmcv/runner/dist_utils.py in wrapper(*args, **kwargs)
142 rank, _ = get_dist_info()
143 if rank == 0:
--> 144 return func(*args, **kwargs)
145
146 return wrapper

~/baseline/mmdetection/mmdet/core/hook/wandblogger_hook.py in before_run(self, runner)
199 self._init_data_table()
200 # Add data to the data table
--> 201 self._add_ground_truth(runner)
202 # Log ground truth data
203 self._log_data_table()

~/baseline/mmdetection/mmdet/core/hook/wandblogger_hook.py in _add_ground_truth(self, runner)
383 is_poly_mask=True,
384 height=img_height,
--> 385 width=img_width)
386 else:
387 wandb_masks = None

~/baseline/mmdetection/mmdet/core/hook/wandblogger_hook.py in _get_wandb_masks(self, masks, labels, is_poly_mask, height, width)
534 if is_poly_mask:
535 if height is not None and width is not None:
--> 536 mask = polygon_to_bitmap(mask, height, width)
537 # Create composite masks for each class.
538 if label not in mask_label_dict.keys():

~/baseline/mmdetection/mmdet/core/mask/structures.py in polygon_to_bitmap(polygons, height, width)
1067 ndarray: the converted masks in bitmap representation
1068 """
-> 1069 rles = maskUtils.frPyObjects(polygons, height, width)
1070 rle = maskUtils.merge(rles)
1071 bitmap_mask = maskUtils.decode(rle).astype(np.bool)

pycocotools/_mask.pyx in pycocotools._mask.frPyObjects()

Exception: input type is not supported.

사진과 같이 클래스를 지정해줬는데도 잘못나오네요 이것만 해결하면 yolo는 바로 될 거 같습니다

lr 0.0002 까지 낮추니까 학습되네요. mAP 0.3811 나왔습니다
https://stackoverflow.com/questions/40472499/issue-nan-with-adam-solver

test

github - slack test