search

czczup / ViT-Adapter

[ICLR 2023 Spotlight] Vision Transformer Adapter for Dense Predictions
https://arxiv.org/abs/2205.08534
Apache License 2.0
1.26k stars 139 forks source link

The vit adapter exhibits very low performance on the potsdam dataset #38

Closed Akinpzx closed 2 years ago

Akinpzx commented 2 years ago

image

The miou and acc are very low but the loss is high

czczup commented 2 years ago

@Akinpzx Hi, thanks for your attention. Could you share the config?

Akinpzx commented 2 years ago

我觉得作者您应该是中国人我就用中文好了，就是我用的config就是您文件里的segmentation里的ade20k中配置文件并做出了一些修改，不知道这是否是最主要的原因呢？

czczup commented 2 years ago

@Akinpzx 我跑一下这个数据集试试，晚一些给你回复

Akinpzx commented 2 years ago

如果可以的话，如果您跑的结果还可以的话可以分享给我一下配置文件吗？

@Akinpzx 我跑一下这个数据集试试，晚一些给你回复

czczup commented 2 years ago

如果可以的话，如果您跑的结果还可以的话可以分享给我一下配置文件吗？

@Akinpzx 我跑一下这个数据集试试，晚一些给你回复

可以的

Akinpzx commented 2 years ago

如果可以的话，如果您跑的结果还可以的话可以分享给我一下配置文件吗？

@Akinpzx 我跑一下这个数据集试试，晚一些给你回复

可以的

请问作者您有什么进展了吗？不知结果如何呢？

czczup commented 2 years ago

如果可以的话，如果您跑的结果还可以的话可以分享给我一下配置文件吗？

@Akinpzx 我跑一下这个数据集试试，晚一些给你回复

可以的

请问作者您有什么进展了吗？不知结果如何呢？

最近卡比较紧张，刚排上训练，晚上可以给你一个初步结果

czczup commented 2 years ago

@Akinpzx 请问你用的mmseg是什么版本，我看potsdam数据集需要升级到v0.21以上

Akinpzx commented 2 years ago

@Akinpzx 请问你用的mmseg是什么版本，我看potsdam数据集需要升级到v0.21以上

我用的mmseg就是您在readme中的配置呢

czczup commented 2 years ago 
2022-08-30 17:43:17,895 - mmseg - INFO - Iter [650/80000]       lr: 6.162e-07, eta: 16:46:51, time: 0.734, data_time: 0.011, memory: 14739, decode.loss_cls: 1.1344, decode.loss_mask: 1.8733, decode.loss_dice: 2.7332, decode.d0.loss_cls: 3.9326, decode.d0.loss_mask: 1.8059, decode.d0.loss_dice: 2.8472, decode.d1.loss_cls: 1.0693, decode.d1.loss_mask: 1.7949, decode.d1.loss_dice: 2.6332, decode.d2.loss_cls: 0.8967, decode.d2.loss_mask: 1.8853, decode.d2.loss_dice: 2.6762, decode.d3.loss_cls: 0.8909, decode.d3.loss_mask: 1.8976, decode.d3.loss_dice: 2.6680, decode.d4.loss_cls: 0.9083, decode.d4.loss_mask: 1.9004, decode.d4.loss_dice: 2.6463, decode.d5.loss_cls: 0.9199, decode.d5.loss_mask: 1.9178, decode.d5.loss_dice: 2.6426, decode.d6.loss_cls: 0.9489, decode.d6.loss_mask: 1.9419, decode.d6.loss_dice: 2.6575, decode.d7.loss_cls: 1.0075, decode.d7.loss_mask: 1.9297, decode.d7.loss_dice: 2.6561, decode.d8.loss_cls: 1.0386, decode.d8.loss_mask: 1.9060, decode.d8.loss_dice: 2.6876, loss: 58.4478
2022-08-30 17:43:55,960 - mmseg - INFO - Iter [700/80000]       lr: 6.632e-07, eta: 16:46:14, time: 0.762, data_time: 0.010, memory: 14739, decode.loss_cls: 1.0483, decode.loss_mask: 1.7637, decode.loss_dice: 2.6110, decode.d0.loss_cls: 3.9144, decode.d0.loss_mask: 1.7102, decode.d0.loss_dice: 2.8163, decode.d1.loss_cls: 1.0022, decode.d1.loss_mask: 1.7091, decode.d1.loss_dice: 2.5921, decode.d2.loss_cls: 0.8398, decode.d2.loss_mask: 1.7653, decode.d2.loss_dice: 2.6079, decode.d3.loss_cls: 0.8310, decode.d3.loss_mask: 1.7782, decode.d3.loss_dice: 2.5665, decode.d4.loss_cls: 0.8056, decode.d4.loss_mask: 1.8115, decode.d4.loss_dice: 2.5830, decode.d5.loss_cls: 0.8257, decode.d5.loss_mask: 1.7929, decode.d5.loss_dice: 2.5601, decode.d6.loss_cls: 0.8534, decode.d6.loss_mask: 1.8186, decode.d6.loss_dice: 2.5559, decode.d7.loss_cls: 0.8931, decode.d7.loss_mask: 1.7994, decode.d7.loss_dice: 2.5683, decode.d8.loss_cls: 0.9442, decode.d8.loss_mask: 1.7963, decode.d8.loss_dice: 2.5709, loss: 55.7350
2022-08-30 17:44:35,093 - mmseg - INFO - Iter [750/80000]       lr: 7.102e-07, eta: 16:47:28, time: 0.782, data_time: 0.010, memory: 14739, decode.loss_cls: 0.9884, decode.loss_mask: 1.7916, decode.loss_dice: 2.5596, decode.d0.loss_cls: 3.9064, decode.d0.loss_mask: 1.7048, decode.d0.loss_dice: 2.7288, decode.d1.loss_cls: 0.9465, decode.d1.loss_mask: 1.7243, decode.d1.loss_dice: 2.5194, decode.d2.loss_cls: 0.7897, decode.d2.loss_mask: 1.7706, decode.d2.loss_dice: 2.5032, decode.d3.loss_cls: 0.7638, decode.d3.loss_mask: 1.8169, decode.d3.loss_dice: 2.4766, decode.d4.loss_cls: 0.7462, decode.d4.loss_mask: 1.8093, decode.d4.loss_dice: 2.4968, decode.d5.loss_cls: 0.7585, decode.d5.loss_mask: 1.8079, decode.d5.loss_dice: 2.4886, decode.d6.loss_cls: 0.7973, decode.d6.loss_mask: 1.8149, decode.d6.loss_dice: 2.5059, decode.d7.loss_cls: 0.8530, decode.d7.loss_mask: 1.8230, decode.d7.loss_dice: 2.5218, decode.d8.loss_cls: 0.9020, decode.d8.loss_mask: 1.8219, decode.d8.loss_dice: 2.5423, loss: 54.6799
2022-08-30 17:45:11,404 - mmseg - INFO - Iter [800/80000]       lr: 7.572e-07, eta: 16:43:49, time: 0.726, data_time: 0.009, memory: 14739, decode.loss_cls: 0.9292, decode.loss_mask: 1.7690, decode.loss_dice: 2.4137, decode.d0.loss_cls: 3.8851, decode.d0.loss_mask: 1.7016, decode.d0.loss_dice: 2.6155, decode.d1.loss_cls: 0.8755, decode.d1.loss_mask: 1.7264, decode.d1.loss_dice: 2.3946, decode.d2.loss_cls: 0.7130, decode.d2.loss_mask: 1.7955, decode.d2.loss_dice: 2.3777, decode.d3.loss_cls: 0.7156, decode.d3.loss_mask: 1.8411, decode.d3.loss_dice: 2.3394, decode.d4.loss_cls: 0.6896, decode.d4.loss_mask: 1.8178, decode.d4.loss_dice: 2.3553, decode.d5.loss_cls: 0.6966, decode.d5.loss_mask: 1.8421, decode.d5.loss_dice: 2.3723, decode.d6.loss_cls: 0.7374, decode.d6.loss_mask: 1.8207, decode.d6.loss_dice: 2.3756, decode.d7.loss_cls: 0.7677, decode.d7.loss_mask: 1.8179, decode.d7.loss_dice: 2.3912, decode.d8.loss_cls: 0.8384, decode.d8.loss_mask: 1.7937, decode.d8.loss_dice: 2.3839, loss: 52.7929
2022-08-30 17:45:51,429 - mmseg - INFO - Iter [850/80000]       lr: 8.040e-07, eta: 16:46:17, time: 0.800, data_time: 0.009, memory: 14739, decode.loss_cls: 0.8088, decode.loss_mask: 1.6895, decode.loss_dice: 2.3235, decode.d0.loss_cls: 3.8720, decode.d0.loss_mask: 1.6149, decode.d0.loss_dice: 2.5795, decode.d1.loss_cls: 0.7906, decode.d1.loss_mask: 1.6328, decode.d1.loss_dice: 2.3551, decode.d2.loss_cls: 0.5908, decode.d2.loss_mask: 1.7062, decode.d2.loss_dice: 2.3469, decode.d3.loss_cls: 0.5752, decode.d3.loss_mask: 1.7307, decode.d3.loss_dice: 2.3066, decode.d4.loss_cls: 0.5777, decode.d4.loss_mask: 1.7321, decode.d4.loss_dice: 2.2958, decode.d5.loss_cls: 0.5975, decode.d5.loss_mask: 1.7426, decode.d5.loss_dice: 2.3051, decode.d6.loss_cls: 0.6213, decode.d6.loss_mask: 1.7161, decode.d6.loss_dice: 2.3165, decode.d7.loss_cls: 0.6618, decode.d7.loss_mask: 1.7041, decode.d7.loss_dice: 2.3114, decode.d8.loss_cls: 0.7209, decode.d8.loss_mask: 1.7046, decode.d8.loss_dice: 2.3104, loss: 50.2410
2022-08-30 17:46:31,371 - mmseg - INFO - Iter [900/80000]       lr: 8.509e-07, eta: 16:48:17, time: 0.799, data_time: 0.063, memory: 14739, decode.loss_cls: 0.7840, decode.loss_mask: 1.7451, decode.loss_dice: 2.3271, decode.d0.loss_cls: 3.8630, decode.d0.loss_mask: 1.6362, decode.d0.loss_dice: 2.5202, decode.d1.loss_cls: 0.7439, decode.d1.loss_mask: 1.6850, decode.d1.loss_dice: 2.3303, decode.d2.loss_cls: 0.5860, decode.d2.loss_mask: 1.7597, decode.d2.loss_dice: 2.3119, decode.d3.loss_cls: 0.5553, decode.d3.loss_mask: 1.8050, decode.d3.loss_dice: 2.2938, decode.d4.loss_cls: 0.5440, decode.d4.loss_mask: 1.8150, decode.d4.loss_dice: 2.3085, decode.d5.loss_cls: 0.5553, decode.d5.loss_mask: 1.8015, decode.d5.loss_dice: 2.3181, decode.d6.loss_cls: 0.5917, decode.d6.loss_mask: 1.7841, decode.d6.loss_dice: 2.2859, decode.d7.loss_cls: 0.6297, decode.d7.loss_mask: 1.7732, decode.d7.loss_dice: 2.3119, decode.d8.loss_cls: 0.6862, decode.d8.loss_mask: 1.7670, decode.d8.loss_dice: 2.3097, loss: 50.4283
Akinpzx commented 2 years ago 
2022-08-30 17:43:17,895 - mmseg - INFO - Iter [650/80000]       lr: 6.162e-07, eta: 16:46:51, time: 0.734, data_time: 0.011, memory: 14739, decode.loss_cls: 1.1344, decode.loss_mask: 1.8733, decode.loss_dice: 2.7332, decode.d0.loss_cls: 3.9326, decode.d0.loss_mask: 1.8059, decode.d0.loss_dice: 2.8472, decode.d1.loss_cls: 1.0693, decode.d1.loss_mask: 1.7949, decode.d1.loss_dice: 2.6332, decode.d2.loss_cls: 0.8967, decode.d2.loss_mask: 1.8853, decode.d2.loss_dice: 2.6762, decode.d3.loss_cls: 0.8909, decode.d3.loss_mask: 1.8976, decode.d3.loss_dice: 2.6680, decode.d4.loss_cls: 0.9083, decode.d4.loss_mask: 1.9004, decode.d4.loss_dice: 2.6463, decode.d5.loss_cls: 0.9199, decode.d5.loss_mask: 1.9178, decode.d5.loss_dice: 2.6426, decode.d6.loss_cls: 0.9489, decode.d6.loss_mask: 1.9419, decode.d6.loss_dice: 2.6575, decode.d7.loss_cls: 1.0075, decode.d7.loss_mask: 1.9297, decode.d7.loss_dice: 2.6561, decode.d8.loss_cls: 1.0386, decode.d8.loss_mask: 1.9060, decode.d8.loss_dice: 2.6876, loss: 58.4478
2022-08-30 17:43:55,960 - mmseg - INFO - Iter [700/80000]       lr: 6.632e-07, eta: 16:46:14, time: 0.762, data_time: 0.010, memory: 14739, decode.loss_cls: 1.0483, decode.loss_mask: 1.7637, decode.loss_dice: 2.6110, decode.d0.loss_cls: 3.9144, decode.d0.loss_mask: 1.7102, decode.d0.loss_dice: 2.8163, decode.d1.loss_cls: 1.0022, decode.d1.loss_mask: 1.7091, decode.d1.loss_dice: 2.5921, decode.d2.loss_cls: 0.8398, decode.d2.loss_mask: 1.7653, decode.d2.loss_dice: 2.6079, decode.d3.loss_cls: 0.8310, decode.d3.loss_mask: 1.7782, decode.d3.loss_dice: 2.5665, decode.d4.loss_cls: 0.8056, decode.d4.loss_mask: 1.8115, decode.d4.loss_dice: 2.5830, decode.d5.loss_cls: 0.8257, decode.d5.loss_mask: 1.7929, decode.d5.loss_dice: 2.5601, decode.d6.loss_cls: 0.8534, decode.d6.loss_mask: 1.8186, decode.d6.loss_dice: 2.5559, decode.d7.loss_cls: 0.8931, decode.d7.loss_mask: 1.7994, decode.d7.loss_dice: 2.5683, decode.d8.loss_cls: 0.9442, decode.d8.loss_mask: 1.7963, decode.d8.loss_dice: 2.5709, loss: 55.7350
2022-08-30 17:44:35,093 - mmseg - INFO - Iter [750/80000]       lr: 7.102e-07, eta: 16:47:28, time: 0.782, data_time: 0.010, memory: 14739, decode.loss_cls: 0.9884, decode.loss_mask: 1.7916, decode.loss_dice: 2.5596, decode.d0.loss_cls: 3.9064, decode.d0.loss_mask: 1.7048, decode.d0.loss_dice: 2.7288, decode.d1.loss_cls: 0.9465, decode.d1.loss_mask: 1.7243, decode.d1.loss_dice: 2.5194, decode.d2.loss_cls: 0.7897, decode.d2.loss_mask: 1.7706, decode.d2.loss_dice: 2.5032, decode.d3.loss_cls: 0.7638, decode.d3.loss_mask: 1.8169, decode.d3.loss_dice: 2.4766, decode.d4.loss_cls: 0.7462, decode.d4.loss_mask: 1.8093, decode.d4.loss_dice: 2.4968, decode.d5.loss_cls: 0.7585, decode.d5.loss_mask: 1.8079, decode.d5.loss_dice: 2.4886, decode.d6.loss_cls: 0.7973, decode.d6.loss_mask: 1.8149, decode.d6.loss_dice: 2.5059, decode.d7.loss_cls: 0.8530, decode.d7.loss_mask: 1.8230, decode.d7.loss_dice: 2.5218, decode.d8.loss_cls: 0.9020, decode.d8.loss_mask: 1.8219, decode.d8.loss_dice: 2.5423, loss: 54.6799
2022-08-30 17:45:11,404 - mmseg - INFO - Iter [800/80000]       lr: 7.572e-07, eta: 16:43:49, time: 0.726, data_time: 0.009, memory: 14739, decode.loss_cls: 0.9292, decode.loss_mask: 1.7690, decode.loss_dice: 2.4137, decode.d0.loss_cls: 3.8851, decode.d0.loss_mask: 1.7016, decode.d0.loss_dice: 2.6155, decode.d1.loss_cls: 0.8755, decode.d1.loss_mask: 1.7264, decode.d1.loss_dice: 2.3946, decode.d2.loss_cls: 0.7130, decode.d2.loss_mask: 1.7955, decode.d2.loss_dice: 2.3777, decode.d3.loss_cls: 0.7156, decode.d3.loss_mask: 1.8411, decode.d3.loss_dice: 2.3394, decode.d4.loss_cls: 0.6896, decode.d4.loss_mask: 1.8178, decode.d4.loss_dice: 2.3553, decode.d5.loss_cls: 0.6966, decode.d5.loss_mask: 1.8421, decode.d5.loss_dice: 2.3723, decode.d6.loss_cls: 0.7374, decode.d6.loss_mask: 1.8207, decode.d6.loss_dice: 2.3756, decode.d7.loss_cls: 0.7677, decode.d7.loss_mask: 1.8179, decode.d7.loss_dice: 2.3912, decode.d8.loss_cls: 0.8384, decode.d8.loss_mask: 1.7937, decode.d8.loss_dice: 2.3839, loss: 52.7929
2022-08-30 17:45:51,429 - mmseg - INFO - Iter [850/80000]       lr: 8.040e-07, eta: 16:46:17, time: 0.800, data_time: 0.009, memory: 14739, decode.loss_cls: 0.8088, decode.loss_mask: 1.6895, decode.loss_dice: 2.3235, decode.d0.loss_cls: 3.8720, decode.d0.loss_mask: 1.6149, decode.d0.loss_dice: 2.5795, decode.d1.loss_cls: 0.7906, decode.d1.loss_mask: 1.6328, decode.d1.loss_dice: 2.3551, decode.d2.loss_cls: 0.5908, decode.d2.loss_mask: 1.7062, decode.d2.loss_dice: 2.3469, decode.d3.loss_cls: 0.5752, decode.d3.loss_mask: 1.7307, decode.d3.loss_dice: 2.3066, decode.d4.loss_cls: 0.5777, decode.d4.loss_mask: 1.7321, decode.d4.loss_dice: 2.2958, decode.d5.loss_cls: 0.5975, decode.d5.loss_mask: 1.7426, decode.d5.loss_dice: 2.3051, decode.d6.loss_cls: 0.6213, decode.d6.loss_mask: 1.7161, decode.d6.loss_dice: 2.3165, decode.d7.loss_cls: 0.6618, decode.d7.loss_mask: 1.7041, decode.d7.loss_dice: 2.3114, decode.d8.loss_cls: 0.7209, decode.d8.loss_mask: 1.7046, decode.d8.loss_dice: 2.3104, loss: 50.2410
2022-08-30 17:46:31,371 - mmseg - INFO - Iter [900/80000]       lr: 8.509e-07, eta: 16:48:17, time: 0.799, data_time: 0.063, memory: 14739, decode.loss_cls: 0.7840, decode.loss_mask: 1.7451, decode.loss_dice: 2.3271, decode.d0.loss_cls: 3.8630, decode.d0.loss_mask: 1.6362, decode.d0.loss_dice: 2.5202, decode.d1.loss_cls: 0.7439, decode.d1.loss_mask: 1.6850, decode.d1.loss_dice: 2.3303, decode.d2.loss_cls: 0.5860, decode.d2.loss_mask: 1.7597, decode.d2.loss_dice: 2.3119, decode.d3.loss_cls: 0.5553, decode.d3.loss_mask: 1.8050, decode.d3.loss_dice: 2.2938, decode.d4.loss_cls: 0.5440, decode.d4.loss_mask: 1.8150, decode.d4.loss_dice: 2.3085, decode.d5.loss_cls: 0.5553, decode.d5.loss_mask: 1.8015, decode.d5.loss_dice: 2.3181, decode.d6.loss_cls: 0.5917, decode.d6.loss_mask: 1.7841, decode.d6.loss_dice: 2.2859, decode.d7.loss_cls: 0.6297, decode.d7.loss_mask: 1.7732, decode.d7.loss_dice: 2.3119, decode.d8.loss_cls: 0.6862, decode.d8.loss_mask: 1.7670, decode.d8.loss_dice: 2.3097, loss: 50.4283

请问miou，acc啥的呢，loss的话我现在打不开我的log文件有点不记得了，但是我记得开始160多到最后稳定收敛到了20多，感觉表现还是比较不太好？

czczup commented 2 years ago

@Akinpzx 这是8000/80000 iters的mIoU，完整的训完还需要16小时

2022-08-30 19:21:49,013 - mmseg - INFO - Iter [8000/80000]      lr: 1.292e-06, eta: 16:00:51, time: 1.284, data_time: 0.010, memory: 14739, decode.loss_cls: 0.1364, decode.loss_mask
: 1.0678, decode.loss_dice: 1.2613, decode.d0.loss_cls: 1.1983, decode.d0.loss_mask: 1.0618, decode.d0.loss_dice: 1.2902, decode.d1.loss_cls: 0.1540, decode.d1.loss_mask: 1.0613, de
code.d1.loss_dice: 1.2739, decode.d2.loss_cls: 0.1577, decode.d2.loss_mask: 1.0582, decode.d2.loss_dice: 1.2563, decode.d3.loss_cls: 0.1552, decode.d3.loss_mask: 1.0524, decode.d3.l
oss_dice: 1.2437, decode.d4.loss_cls: 0.1612, decode.d4.loss_mask: 1.0587, decode.d4.loss_dice: 1.2411, decode.d5.loss_cls: 0.1574, decode.d5.loss_mask: 1.0585, decode.d5.loss_dice:
 1.2433, decode.d6.loss_cls: 0.1523, decode.d6.loss_mask: 1.0612, decode.d6.loss_dice: 1.2478, decode.d7.loss_cls: 0.1546, decode.d7.loss_mask: 1.0591, decode.d7.loss_dice: 1.2467,
decode.d8.loss_cls: 0.1480, decode.d8.loss_mask: 1.0647, decode.d8.loss_dice: 1.2537, loss: 25.7369
[>>>>>>>>>>>>>>>>>>>>>>>>>>>] 2016/2016, 8.5 task/s, elapsed: 237s, ETA:     0s

2022-08-30 19:25:46,489 - mmseg - INFO - per class results:
2022-08-30 19:25:46,498 - mmseg - INFO -
+--------------------+-------+-------+
|       Class        |  IoU  |  Acc  |
+--------------------+-------+-------+
| impervious_surface | 87.92 | 94.09 |
|      building      | 94.29 | 97.25 |
|   low_vegetation   | 78.19 | 89.95 |
|        tree        | 80.55 | 87.61 |
|        car         | 91.87 | 97.72 |
|      clutter       | 42.05 | 52.84 |
+--------------------+-------+-------+
2022-08-30 19:25:46,498 - mmseg - INFO - Summary:
2022-08-30 19:25:46,498 - mmseg - INFO -
+------+-------+-------+
| aAcc |  mIoU |  mAcc |
+------+-------+-------+
| 91.2 | 79.14 | 86.58 |
+------+-------+-------+
2022-08-30 19:26:08,547 - mmseg - INFO - Now best checkpoint is saved as best_mIoU_iter_8000.pth.
2022-08-30 19:26:08,558 - mmseg - INFO - Best mIoU is 0.7914 at 8000 iter.
2022-08-30 19:26:08,574 - mmseg - INFO - Exp name: mask2former_beit_adapter_large_512_80k_potsdam_ss.py
Akinpzx commented 2 years ago

@Akinpzx 这是8000/80000 iters的mIoU，完整的训完还需要16小时

2022-08-30 19:21:49,013 - mmseg - INFO - Iter [8000/80000]      lr: 1.292e-06, eta: 16:00:51, time: 1.284, data_time: 0.010, memory: 14739, decode.loss_cls: 0.1364, decode.loss_mask
: 1.0678, decode.loss_dice: 1.2613, decode.d0.loss_cls: 1.1983, decode.d0.loss_mask: 1.0618, decode.d0.loss_dice: 1.2902, decode.d1.loss_cls: 0.1540, decode.d1.loss_mask: 1.0613, de
code.d1.loss_dice: 1.2739, decode.d2.loss_cls: 0.1577, decode.d2.loss_mask: 1.0582, decode.d2.loss_dice: 1.2563, decode.d3.loss_cls: 0.1552, decode.d3.loss_mask: 1.0524, decode.d3.l
oss_dice: 1.2437, decode.d4.loss_cls: 0.1612, decode.d4.loss_mask: 1.0587, decode.d4.loss_dice: 1.2411, decode.d5.loss_cls: 0.1574, decode.d5.loss_mask: 1.0585, decode.d5.loss_dice:
 1.2433, decode.d6.loss_cls: 0.1523, decode.d6.loss_mask: 1.0612, decode.d6.loss_dice: 1.2478, decode.d7.loss_cls: 0.1546, decode.d7.loss_mask: 1.0591, decode.d7.loss_dice: 1.2467,
decode.d8.loss_cls: 0.1480, decode.d8.loss_mask: 1.0647, decode.d8.loss_dice: 1.2537, loss: 25.7369
[>>>>>>>>>>>>>>>>>>>>>>>>>>>] 2016/2016, 8.5 task/s, elapsed: 237s, ETA:     0s

2022-08-30 19:25:46,489 - mmseg - INFO - per class results:
2022-08-30 19:25:46,498 - mmseg - INFO -
+--------------------+-------+-------+
|       Class        |  IoU  |  Acc  |
+--------------------+-------+-------+
| impervious_surface | 87.92 | 94.09 |
|      building      | 94.29 | 97.25 |
|   low_vegetation   | 78.19 | 89.95 |
|        tree        | 80.55 | 87.61 |
|        car         | 91.87 | 97.72 |
|      clutter       | 42.05 | 52.84 |
+--------------------+-------+-------+
2022-08-30 19:25:46,498 - mmseg - INFO - Summary:
2022-08-30 19:25:46,498 - mmseg - INFO -
+------+-------+-------+
| aAcc |  mIoU |  mAcc |
+------+-------+-------+
| 91.2 | 79.14 | 86.58 |
+------+-------+-------+
2022-08-30 19:26:08,547 - mmseg - INFO - Now best checkpoint is saved as best_mIoU_iter_8000.pth.
2022-08-30 19:26:08,558 - mmseg - INFO - Best mIoU is 0.7914 at 8000 iter.
2022-08-30 19:26:08,574 - mmseg - INFO - Exp name: mask2former_beit_adapter_large_512_80k_potsdam_ss.py

wow，希望您能分享一下配置文件呢，肯定是我配置错了导致数据那么低的，我学习语义分割不是很久，还不是很清楚

czczup commented 2 years ago

@Akinpzx 我更新了config，你可以试试看。另外，麻烦确认一下mmseg的版本，如果是v0.20.2，应该是没法直接跑potsdam这个数据集的，因为他是v0.21才加进来的

2022-08-30 21:08:50,453 - mmseg - INFO - Exp name: mask2former_beit_adapter_large_512_80k_potsdam_ss.py                                                                     [15/1870]
2022-08-30 21:08:50,459 - mmseg - INFO - Iter [16000/80000]     lr: 1.149e-06, eta: 14:15:07, time: 1.269, data_time: 0.064, memory: 14739, decode.loss_cls: 0.1133, decode.loss_mask
: 0.8411, decode.loss_dice: 1.0110, decode.d0.loss_cls: 0.3088, decode.d0.loss_mask: 0.8615, decode.d0.loss_dice: 1.0595, decode.d1.loss_cls: 0.1243, decode.d1.loss_mask: 0.8513, de
code.d1.loss_dice: 1.0372, decode.d2.loss_cls: 0.1146, decode.d2.loss_mask: 0.8482, decode.d2.loss_dice: 1.0249, decode.d3.loss_cls: 0.1102, decode.d3.loss_mask: 0.8433, decode.d3.l
oss_dice: 1.0234, decode.d4.loss_cls: 0.1071, decode.d4.loss_mask: 0.8428, decode.d4.loss_dice: 1.0266, decode.d5.loss_cls: 0.1121, decode.d5.loss_mask: 0.8383, decode.d5.loss_dice:
 1.0219, decode.d6.loss_cls: 0.1120, decode.d6.loss_mask: 0.8406, decode.d6.loss_dice: 1.0128, decode.d7.loss_cls: 0.1144, decode.d7.loss_mask: 0.8400, decode.d7.loss_dice: 1.0127,
decode.d8.loss_cls: 0.1059, decode.d8.loss_mask: 0.8423, decode.d8.loss_dice: 1.0103, loss: 20.0125
[>>>>>>>>>>>>>>>>>>>>>>>>>>>] 2016/2016, 58.8 task/s, elapsed: 34s, ETA:     0s

2022-08-30 21:09:25,874 - mmseg - INFO - per class results:
2022-08-30 21:09:25,883 - mmseg - INFO -
+--------------------+-------+-------+
|       Class        |  IoU  |  Acc  |
+--------------------+-------+-------+
| impervious_surface | 88.24 | 93.48 |
|      building      | 94.79 | 98.11 |
|   low_vegetation   | 78.63 | 90.54 |
|        tree        | 81.08 | 88.79 |
|        car         | 92.88 | 96.93 |
|      clutter       |  43.8 | 52.93 |
+--------------------+-------+-------+
2022-08-30 21:09:25,883 - mmseg - INFO - Summary:
2022-08-30 21:09:25,883 - mmseg - INFO -
+-------+------+------+
|  aAcc | mIoU | mAcc |
+-------+------+------+
| 91.54 | 79.9 | 86.8 |
+-------+------+------+
2022-08-30 21:09:26,071 - mmseg - INFO - The previous best checkpoint /mnt/petrelfs/chenzhe1/workspace/ViT-Adapter/segmentation/work_dirs/mask2former_beit_adapter_large_512_80k_pot$dam_ss/best_mIoU_iter_8000.pth was removed
2022-08-30 21:09:49,308 - mmseg - INFO - Now best checkpoint is saved as best_mIoU_iter_16000.pth.
2022-08-30 21:09:49,322 - mmseg - INFO - Best mIoU is 0.7990 at 16000 iter.
2022-08-30 21:09:49,338 - mmseg - INFO - Exp name: mask2former_beit_adapter_large_512_80k_potsdam_ss.py
Akinpzx commented 2 years ago

@Akinpzx 我更新了config，你可以试试看。另外，麻烦确认一下mmseg的版本，如果是v0.20.2，应该是没法直接跑potsdam这个数据集的，因为他是v0.21才加进来的

2022-08-30 21:08:50,453 - mmseg - INFO - Exp name: mask2former_beit_adapter_large_512_80k_potsdam_ss.py                                                                     [15/1870]
2022-08-30 21:08:50,459 - mmseg - INFO - Iter [16000/80000]     lr: 1.149e-06, eta: 14:15:07, time: 1.269, data_time: 0.064, memory: 14739, decode.loss_cls: 0.1133, decode.loss_mask
: 0.8411, decode.loss_dice: 1.0110, decode.d0.loss_cls: 0.3088, decode.d0.loss_mask: 0.8615, decode.d0.loss_dice: 1.0595, decode.d1.loss_cls: 0.1243, decode.d1.loss_mask: 0.8513, de
code.d1.loss_dice: 1.0372, decode.d2.loss_cls: 0.1146, decode.d2.loss_mask: 0.8482, decode.d2.loss_dice: 1.0249, decode.d3.loss_cls: 0.1102, decode.d3.loss_mask: 0.8433, decode.d3.l
oss_dice: 1.0234, decode.d4.loss_cls: 0.1071, decode.d4.loss_mask: 0.8428, decode.d4.loss_dice: 1.0266, decode.d5.loss_cls: 0.1121, decode.d5.loss_mask: 0.8383, decode.d5.loss_dice:
 1.0219, decode.d6.loss_cls: 0.1120, decode.d6.loss_mask: 0.8406, decode.d6.loss_dice: 1.0128, decode.d7.loss_cls: 0.1144, decode.d7.loss_mask: 0.8400, decode.d7.loss_dice: 1.0127,
decode.d8.loss_cls: 0.1059, decode.d8.loss_mask: 0.8423, decode.d8.loss_dice: 1.0103, loss: 20.0125
[>>>>>>>>>>>>>>>>>>>>>>>>>>>] 2016/2016, 58.8 task/s, elapsed: 34s, ETA:     0s

2022-08-30 21:09:25,874 - mmseg - INFO - per class results:
2022-08-30 21:09:25,883 - mmseg - INFO -
+--------------------+-------+-------+
|       Class        |  IoU  |  Acc  |
+--------------------+-------+-------+
| impervious_surface | 88.24 | 93.48 |
|      building      | 94.79 | 98.11 |
|   low_vegetation   | 78.63 | 90.54 |
|        tree        | 81.08 | 88.79 |
|        car         | 92.88 | 96.93 |
|      clutter       |  43.8 | 52.93 |
+--------------------+-------+-------+
2022-08-30 21:09:25,883 - mmseg - INFO - Summary:
2022-08-30 21:09:25,883 - mmseg - INFO -
+-------+------+------+
|  aAcc | mIoU | mAcc |
+-------+------+------+
| 91.54 | 79.9 | 86.8 |
+-------+------+------+
2022-08-30 21:09:26,071 - mmseg - INFO - The previous best checkpoint /mnt/petrelfs/chenzhe1/workspace/ViT-Adapter/segmentation/work_dirs/mask2former_beit_adapter_large_512_80k_pot$dam_ss/best_mIoU_iter_8000.pth was removed
2022-08-30 21:09:49,308 - mmseg - INFO - Now best checkpoint is saved as best_mIoU_iter_16000.pth.
2022-08-30 21:09:49,322 - mmseg - INFO - Best mIoU is 0.7990 at 16000 iter.
2022-08-30 21:09:49,338 - mmseg - INFO - Exp name: mask2former_beit_adapter_large_512_80k_potsdam_ss.py

好的好的，多谢您的帮助，我这就去试试

czczup commented 2 years ago

@Akinpzx 最后训完是80左右

2022-08-31 10:56:42,526 - mmseg - INFO - Iter [80000/80000]     lr: 1.795e-11, eta: 0:00:00, time: 1.268, data_time: 0.021, memory: 14739, decode.loss_cls: 0.0268, decode.loss_mask:
 0.4506, decode.loss_dice: 0.5791, decode.d0.loss_cls: 0.0967, decode.d0.loss_mask: 0.4644, decode.d0.loss_dice: 0.6125, decode.d1.loss_cls: 0.0379, decode.d1.loss_mask: 0.4511, dec
ode.d1.loss_dice: 0.5910, decode.d2.loss_cls: 0.0320, decode.d2.loss_mask: 0.4499, decode.d2.loss_dice: 0.5850, decode.d3.loss_cls: 0.0243, decode.d3.loss_mask: 0.4511, decode.d3.lo
ss_dice: 0.5793, decode.d4.loss_cls: 0.0291, decode.d4.loss_mask: 0.4510, decode.d4.loss_dice: 0.5734, decode.d5.loss_cls: 0.0285, decode.d5.loss_mask: 0.4506, decode.d5.loss_dice:
0.5794, decode.d6.loss_cls: 0.0260, decode.d6.loss_mask: 0.4507, decode.d6.loss_dice: 0.5725, decode.d7.loss_cls: 0.0252, decode.d7.loss_mask: 0.4511, decode.d7.loss_dice: 0.5835, d
ecode.d8.loss_cls: 0.0252, decode.d8.loss_mask: 0.4507, decode.d8.loss_dice: 0.5796, loss: 10.7083
[>>>>>>>>>>>>>>>>>>>>>>>>>>>] 2016/2016, 57.2 task/s, elapsed: 35s, ETA:     0s

2022-08-31 10:57:18,443 - mmseg - INFO - per class results:
2022-08-31 10:57:18,448 - mmseg - INFO -
+--------------------+-------+-------+
|       Class        |  IoU  |  Acc  |
+--------------------+-------+-------+
| impervious_surface | 88.18 | 94.11 |
|      building      | 94.59 | 98.03 |
|   low_vegetation   | 78.84 | 90.53 |
|        tree        | 81.02 | 88.54 |
|        car         | 93.33 | 96.91 |
|      clutter       | 43.91 | 50.91 |
+--------------------+-------+-------+
2022-08-31 10:57:18,448 - mmseg - INFO - Summary:
2022-08-31 10:57:18,449 - mmseg - INFO -
+-------+-------+-------+
|  aAcc |  mIoU |  mAcc |
+-------+-------+-------+
| 91.58 | 79.98 | 86.51 |
+-------+-------+-------+
2022-08-31 10:57:18,456 - mmseg - INFO - Exp name: mask2former_beit_adapter_large_512_80k_potsdam_ss.py
Akinpzx commented 2 years ago

@Akinpzx 最后训完是80左右

2022-08-31 10:56:42,526 - mmseg - INFO - Iter [80000/80000]     lr: 1.795e-11, eta: 0:00:00, time: 1.268, data_time: 0.021, memory: 14739, decode.loss_cls: 0.0268, decode.loss_mask:
 0.4506, decode.loss_dice: 0.5791, decode.d0.loss_cls: 0.0967, decode.d0.loss_mask: 0.4644, decode.d0.loss_dice: 0.6125, decode.d1.loss_cls: 0.0379, decode.d1.loss_mask: 0.4511, dec
ode.d1.loss_dice: 0.5910, decode.d2.loss_cls: 0.0320, decode.d2.loss_mask: 0.4499, decode.d2.loss_dice: 0.5850, decode.d3.loss_cls: 0.0243, decode.d3.loss_mask: 0.4511, decode.d3.lo
ss_dice: 0.5793, decode.d4.loss_cls: 0.0291, decode.d4.loss_mask: 0.4510, decode.d4.loss_dice: 0.5734, decode.d5.loss_cls: 0.0285, decode.d5.loss_mask: 0.4506, decode.d5.loss_dice:
0.5794, decode.d6.loss_cls: 0.0260, decode.d6.loss_mask: 0.4507, decode.d6.loss_dice: 0.5725, decode.d7.loss_cls: 0.0252, decode.d7.loss_mask: 0.4511, decode.d7.loss_dice: 0.5835, d
ecode.d8.loss_cls: 0.0252, decode.d8.loss_mask: 0.4507, decode.d8.loss_dice: 0.5796, loss: 10.7083
[>>>>>>>>>>>>>>>>>>>>>>>>>>>] 2016/2016, 57.2 task/s, elapsed: 35s, ETA:     0s

2022-08-31 10:57:18,443 - mmseg - INFO - per class results:
2022-08-31 10:57:18,448 - mmseg - INFO -
+--------------------+-------+-------+
|       Class        |  IoU  |  Acc  |
+--------------------+-------+-------+
| impervious_surface | 88.18 | 94.11 |
|      building      | 94.59 | 98.03 |
|   low_vegetation   | 78.84 | 90.53 |
|        tree        | 81.02 | 88.54 |
|        car         | 93.33 | 96.91 |
|      clutter       | 43.91 | 50.91 |
+--------------------+-------+-------+
2022-08-31 10:57:18,448 - mmseg - INFO - Summary:
2022-08-31 10:57:18,449 - mmseg - INFO -
+-------+-------+-------+
|  aAcc |  mIoU |  mAcc |
+-------+-------+-------+
| 91.58 | 79.98 | 86.51 |
+-------+-------+-------+
2022-08-31 10:57:18,456 - mmseg - INFO - Exp name: mask2former_beit_adapter_large_512_80k_potsdam_ss.py

多谢您的帮助！

Akinpzx commented 2 years ago

@czczup 我还用一个问题想请教一下就是断点续训在哪设置呢？

czczup commented 2 years ago

@czczup 我还用一个问题想请教一下就是断点续训在哪设置呢？

sh dist_train.sh <config> <gpu-num> --resume-from <checkpoint-path>

用resume-from指定checkpoint

Akinpzx commented 2 years ago

@czczup 出现了一个错误 image

Akinpzx commented 2 years ago

@czczup 我将sh替换为bash解决了问题，虽然不是很懂其中的原理