I can't reproduce the results of segmentation in the paper

[YuHe0108]

I cloned this project and configured the environment, dataset and parameters. However, when starting the training and using python main.py segmentation, I found that no matter how many times the training was conducted, the iou of small and large was always 0. Moreover, the mask effect predicted by the model is very poor. I have no idea where the problem lies.

100%|██████████| 6/6 [00:35<00:00, 5.94s/it, 0.3247 (34218.1797) / 0.3145 0.1385, 0.0000, 0.0000] Epoch #11 100%|██████████| 29/29 [02:14<00:00, 4.65s/it, 0.2912 (840568.0000) / 0.3135 0.1346, 0.0000, 0.0000] 100%|██████████| 6/6 [00:34<00:00, 5.80s/it, 1.7307 (29785.9453) / 0.3145 0.1385, 0.0000, 0.0000] Epoch #12 100%|██████████| 29/29 [02:21<00:00, 4.88s/it, 0.2919 (987476.0000) / 0.3135 0.1346, 0.0000, 0.0000] 100%|██████████| 6/6 [00:35<00:00, 6.00s/it, 2.4828 (23808.0781) / 0.3145 0.1385, 0.0000, 0.0000] Epoch #13 100%|██████████| 29/29 [02:14<00:00, 4.63s/it, 0.2972 (1279292.5000) / 0.3135 0.1346, 0.0000, 0.0000] 67%|██████▋ | 4/6 [00:21<00:06, 3.24s/it, 53.7378 (255833280.0000) / 0.3125 0.1362, 0.0000, 0.0000]