After training, the bbox cannot be detected

[forkcc]

hello,I am learning to use this framework
def train(): cfg = get_cfg() cfg.merge_from_file(model_zoo.get_config_file("COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml")) cfg.MODEL.WEIGHTS = download_model() cfg.DATASETS.TRAIN = ("self_coco_train",) cfg.DATASETS.TEST = ("self_coco_test",) cfg.TEST.EVAL_PERIOD = 1000 cfg.DATALOADER.NUM_WORKERS = 4 cfg.INPUT.MIN_SIZE_TRAIN = WIDTH # cfg.INPUT.MAX_SIZE_TRAIN = WIDTH cfg.INPUT.MIN_SIZE_TEST = WIDTH # cfg.INPUT.MAX_SIZE_TEST = WIDTH cfg.SOLVER.WARMUP_ITERS = 1000 cfg.MODEL.ROI_HEADS.NUM_CLASSES = int(os.environ.get("CLASSES")) + 1 cfg.SOLVER.GAMMA = 0.05 cfg.INPUT.MIN_SIZE_TRAIN_SAMPLING = "choice" MAX = int(ITEMS / IMS_PER_BATCH) * 3000 cfg.SOLVER.MAX_ITER = MAX cfg.SOLVER.IMS_PER_BATCH = IMS_PER_BATCH cfg.SOLVER.STEPS = (MAX - 500, MAX) cfg.SOLVER.BASE_LR = LR cfg.MODEL.DEVICE = "cuda" cfg.OUTPUT_DIR = OUTPUT os.makedirs(OUTPUT, exist_ok=True) print("训练开始") trainer = CocoTrainer(cfg) best = BestCheckpointer(eval_period=cfg.TEST.EVAL_PERIOD, checkpointer=trainer.checkpointer, val_metric="bbox/AP75") trainer.register_hooks([CallbackHook(after_step=after_step), best]) trainer.resume_or_load(resume=False) print(trainer.train()) print("训练完毕")
I used this py file for training, and now I encountered a strange question graph.
If I set weights = model_zoo.get_checkpoint_url("COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml")

| AP | AP50 | AP75 | APs | APm | APl |
| 77.344 | 94.826 | 91.785 | nan | 70.000 | 78.397 |
| category | AP | category | AP |
| lyf | nan | lyf | 77.344 |

Although the score is high, no objects can be detected
If I delete weights then everything is normal, so I want to know the meaning of setting weights during training