OWOD多个task训练后测试结果问题

[wl654655902]

烟火数据task1，5类，人车数据集task2,25类， "task_list": [0, 5, 30],
训练task1过程中：

smoke                                   [00]: AP50=11.782, #pred=5124
fire                                    [01]: AP50=4.394, #pred=2116
cloud                                   [02]: AP50=2.165, #pred=68
fog                                     [03]: AP50=0.000, #pred=190
chimney                                 [04]: AP50=0.394, #pred=378
unknown                                 [05]: AR50=74.162, #pred=524823
01/16 00:56:48 - mmengine - INFO - known classes has 7876 predictions.
01/16 00:56:48 - mmengine - INFO - unknown classes has 524823 predictions.
01/16 00:56:48 - mmengine - INFO - Wilderness Impact: {50: 0.0034278180619644037}
01/16 00:56:48 - mmengine - INFO - Absolute OSE (total_num_unk_det_as_known): {50: 27.0}
01/16 00:56:48 - mmengine - INFO - total_num_unk 21515
01/16 00:56:48 - mmengine - INFO - Current class AP50: 3.7471883620144704
01/16 00:56:48 - mmengine - INFO - Current class Precisions50: 9.631729353553066
01/16 00:56:48 - mmengine - INFO - Current class Recall50: 13.304210068743606
01/16 00:56:48 - mmengine - INFO - Known AP50: 3.7471883620144704
01/16 00:56:48 - mmengine - INFO - Known Precisions50: 9.631729353553066
01/16 00:56:48 - mmengine - INFO - Known Recall50: 13.304210068743606
01/16 00:56:48 - mmengine - INFO - Unknown AP50: 26.99864069336281
01/16 00:56:48 - mmengine - INFO - Unknown Precisions50: 3.040263098225497
01/16 00:56:48 - mmengine - INFO - Unknown Recall50: 74.16221237547654

训练task2过程中：

smoke                                   [00]: AP50=11.809, #pred=5123
fire                                    [01]: AP50=4.392, #pred=2119
cloud                                   [02]: AP50=2.165, #pred=68
fog                                     [03]: AP50=0.000, #pred=189
chimney                                 [04]: AP50=0.390, #pred=378
person                                  [05]: AP50=18.859, #pred=11307
car                                     [06]: AP50=49.922, #pred=17710
pickup truck                            [07]: AP50=12.473, #pred=157
minibus                                 [08]: AP50=26.351, #pred=1842
motorhome                               [09]: AP50=0.000, #pred=0
truck                                   [10]: AP50=41.992, #pred=10958
forklift                                [11]: AP50=2.349, #pred=104
pushdozer                               [12]: AP50=23.951, #pred=860
crane                                   [13]: AP50=34.171, #pred=273
road roller                             [14]: AP50=9.965, #pred=65
excavator                               [15]: AP50=44.151, #pred=959
mixer truck                             [16]: AP50=34.719, #pred=659
oil tank truck                          [17]: AP50=16.583, #pred=235
watering cart                           [18]: AP50=18.266, #pred=312
mixer truck                             [19]: AP50=0.000, #pred=0
fire fighting truck                     [20]: AP50=20.766, #pred=88
garbage truck                           [21]: AP50=0.000, #pred=70
bus                                     [22]: AP50=31.923, #pred=420
tractors                                [23]: AP50=10.000, #pred=9
bicycle                                 [24]: AP50=38.277, #pred=460
electric scooter                        [25]: AP50=0.000, #pred=2
motorbike                               [26]: AP50=37.552, #pred=2617
tricycle                                [27]: AP50=19.135, #pred=1405
train                                   [28]: AP50=69.948, #pred=128
other                                   [29]: AP50=4.534, #pred=2274
unknown                                 [30]: AR50=0.000, #pred=356875
01/16 01:37:43 - mmengine - INFO - known classes has 60791 predictions.
01/16 01:37:43 - mmengine - INFO - unknown classes has 356875 predictions.
01/16 01:37:43 - mmengine - INFO - Wilderness Impact: {50: 0.0}
01/16 01:37:43 - mmengine - INFO - Absolute OSE (total_num_unk_det_as_known): {50: 0.0}
01/16 01:37:43 - mmengine - INFO - total_num_unk 0
01/16 01:37:43 - mmengine - INFO - Prev class AP50: 3.7514281310398987
01/16 01:37:43 - mmengine - INFO - Prev class Precisions50: 9.6304934163019
01/16 01:37:43 - mmengine - INFO - Prev class Recall50: 13.31480328902648
01/16 01:37:43 - mmengine - INFO - Current class AP50: 22.635590289466926
01/16 01:37:43 - mmengine - INFO - Current class Precisions50: 17.38094251472993
01/16 01:37:43 - mmengine - INFO - Current class Recall50: 39.9013804206519
01/16 01:37:43 - mmengine - INFO - Known AP50: 19.48822992972909
01/16 01:37:43 - mmengine - INFO - Known Precisions50: 16.08920099832526
01/16 01:37:43 - mmengine - INFO - Known Recall50: 35.47028423204767
01/16 01:37:43 - mmengine - INFO - Unknown AP50: 0.0
01/16 01:37:43 - mmengine - INFO - Unknown Precisions50: 0.0
01/16 01:37:43 - mmengine - INFO - Unknown Recall50: 0.0

在训练task2过程中，task1 5类的评测指标没有变过。task2的指标有改变。

模型训练好后，我分别用yolo_uniow_l_lora_bn_1e-3_20e_8gpus_koala_koala_train_task1和yolo_uniow_l_lora_bn_1e-3_20e_8gpus_koala_koala_train_task2里的best模型评测烟火数据（task1），task2里best模型评测效果有一些下降。
但是我用这两个模型推理人车数据（task2）时，yolo_uniow_l_lora_bn_1e-3_20e_8gpus_koala_koala_train_task1里的best模型能推理出烟火5类和unknow（人车和其他目标都是unknow），这个能理解，重点是用yolo_uniow_l_lora_bn_1e-3_20e_8gpus_koala_koala_train_task2预测人车数据集，也只能推理出烟火5类，人车根本就检测不到了，少许目标结果是第六类，但是标签也是不对的。

[wl654655902]

我问题解决了，在推理task2人车数据集时，需要将数据配置文件中的owod_task = {{'$TASK:1'}} 改成owod_task = {{'$TASK:2'}}