训练denet网络精度问题

[MiniBullLab]

denet在水果上面的训练精度异常:

1/19 Batch 0... Done. (0.759s)
2/19 Batch 1... Done. (0.989s)
3/19 Batch 2... Done. (1.005s)
4/19 Batch 3... Done. (0.832s)
5/19 Batch 4... Done. (0.983s)
6/19 Batch 5... Done. (0.875s)
7/19 Batch 6... Done. (1.024s)
8/19 Batch 7... Done. (1.090s)
9/19 Batch 8... Done. (1.092s)
10/19 Batch 9... Done. (1.054s)
11/19 Batch 10... Done. (1.052s)
12/19 Batch 11... Done. (0.998s)
13/19 Batch 12... Done. (1.163s)
14/19 Batch 13... Done. (1.059s)
15/19 Batch 14... Done. (0.874s)
16/19 Batch 15... Done. (0.761s)
17/19 Batch 16... Done. (0.958s)
18/19 Batch 17... Done. (1.025s)
19/19 Batch 18... Done. (0.806s)
Mean AP = 0.0006
~~~~~~~~
Results:
apple: 0.000
pear: 0.002
potato: 0.000
orange: 0.000

[lpj0822]

已经修改

[MiniBullLab]

develop训练结果： poch:99 | mAP = 0.4896 | pear: 0.000 apple: 0.958 orange: 0.000 potato: 1.000 edge_tools训练结果： poch: 99 | mAP: 0.976 | apple: 0.958 potato: 1.000 orange: 1.000 pear: 0.944

[kingwangxiang]

问题调试中

[foww-0001]

大致定位到loss层上一层的卷积模型参数给的不一样。

[foww-0001]

替换为main分支上的loss函数，仍然有两类无法检测出来。

Epoch: 99 | mAP: 0.490 | apple: 0.958 orange: 0.000 potato: 1.000 pear: 0.001

[foww-0001]

定位到是config中的参数不同，detect2d_class和dataset中的detect2d_class中的不同导致训练结果偏差，具体需要进一步测试。

[MiniBullLab]

有问题尽快沟通

[foww-0001]

mAP: 0.980 | orange: 1.000 potato: 1.000 pear: 0.963 apple: 0.958 最新的develop分支精度正常。