用百度开源的印章数据集，训练到50多个epoch,loss还是没降下去 about pgnet.pytorch HOT 1 OPEN

用百度开源的印章数据集（8千张训练集），训练到50多个epoch,loss还是没降下去, 训练日志如下：

53/600/300/333 | loss:103.55 | score_loss:0.03 | border_loss:0.01 | direction_loss:0.01 | ctc_loss:20.70 | | Lr:0.00100000 53/600/305/333 | loss:96.91 | score_loss:0.03 | border_loss:0.01 | direction_loss:0.00 | ctc_loss:19.37 | | Lr:0.00100000 53/600/310/333 | loss:95.08 | score_loss:0.03 | border_loss:0.01 | direction_loss:0.00 | ctc_loss:19.01 | | Lr:0.00100000 53/600/315/333 | loss:165.42 | score_loss:0.03 | border_loss:0.01 | direction_loss:0.00 | ctc_loss:33.08 | | Lr:0.00100000 53/600/320/333 | loss:162.01 | score_loss:0.03 | border_loss:0.01 | direction_loss:0.01 | ctc_loss:32.39 | | Lr:0.00100000 53/600/325/333 | loss:162.18 | score_loss:0.03 | border_loss:0.01 | direction_loss:0.00 | ctc_loss:32.43 | | Lr:0.00100000 53/600/330/333 | loss:161.29 | score_loss:0.03 | border_loss:0.01 | direction_loss:0.00 | ctc_loss:32.25 | | Lr:0.00100000 epoch_loss:148.26 | epoch_score_loss:0.03 | epoch_border_loss:0.01 | epoch_dire_loss:0.00 | epoch_ctc_loss:29.64 54/600/0/333 | loss:94.88 | score_loss:0.03 | border_loss:0.01 | direction_loss:0.00 | ctc_loss:18.97 | | Lr:0.00100000 54/600/5/333 | loss:96.44 | score_loss:0.03 | border_loss:0.01 | direction_loss:0.01 | ctc_loss:19.28 | | Lr:0.00100000 54/600/10/333 | loss:93.79 | score_loss:0.02 | border_loss:0.01 | direction_loss:0.00 | ctc_loss:18.75 | | Lr:0.00100000 54/600/15/333 | loss:161.79 | score_loss:0.02 | border_loss:0.01 | direction_loss:0.00 | ctc_loss:32.35 | | Lr:0.00100000 54/600/20/333 | loss:162.57 | score_loss:0.03 | border_loss:0.01 | direction_loss:0.00 | ctc_loss:32.51 | | Lr:0.00100000 54/600/25/333 | loss:165.98 | score_loss:0.03 | border_loss:0.01 | direction_loss:0.00 | ctc_loss:33.19 | | Lr:0.00100000 54/600/30/333 | loss:164.87 | score_loss:0.03 | border_loss:0.01 | direction_loss:0.00 | ctc_loss:32.96 | | Lr:0.00100000 54/600/35/333 | loss:161.13 | score_loss:0.03 | border_loss:0.01 | direction_loss:0.00 | ctc_loss:32.22 | | Lr:0.00100000 54/600/40/333 | loss:159.28 | score_loss:0.03 | border_loss:0.01 | direction_loss:0.00 | ctc_loss:31.85 | | Lr:0.00100000 54/600/45/333 | loss:164.96 | score_loss:0.03 | border_loss:0.01 | direction_loss:0.00 | ctc_loss:32.98 | | Lr:0.00100000

这个是英文的，类别变了需要改一部分代码，具体我也没试过，但是loss那里肯定是要改的，你可以自己完善下

from pgnet.pytorch.