eval.py的结果全是0 训练了300批次这是怎么回事啊

[jun214384468]

eval.py的结果全是0 训练了300批次这是怎么回事啊

[Shualite]

I get the same situation! How do you fix it?

[Shualite]

@SakuraRiven

[Shualite]

@SakuraRiven 麻烦大佬指点一下

[jun214384468]

I get the same situation! How do you fix it?

其他网络中同样的情况下，有人建议调整学习率试试，但是我没解决掉

[Shualite]

@jun214384468 你指的是其他人复现的EAST，也会有这样的问题么？

[SakuraRiven]

loss 有正常收敛吗，detect一张图的结果正常吗

[victoic]

The same is happening to me, I'm calling eval.py after each epoch and the result is always all 0 after 89 epochs. Loss seems to be decreasing from the starting 3 at epoch 1, to around 1.3 at epoch 89. Is this normal?

[SakuraRiven]

应该是mask分类的问题，可以考虑（1）加入CE loss，（2）调整crop截断文字的策略，可以简单的ignore掉这些截断文字

[victoic]

@SakuraRiven Thank you for the fast reply and attention. Could you please elaborate on how can I accomplish those tasks?
Specifically, I am not sure on (1) where to add CE loss and (2) where in the code are the crops truncated.

Thank you again.

[SakuraRiven]

(1) the classify_loss in loss.py could be considered with DICE+CE loss for easier convergence;
(2) In dataset.py, the is_cross_text function is used in crop_img function to avoid cutting texts. However, we can directly crop the patch and ignore the cutted texts.

[RtekiX]

loss 有正常收敛吗，detect一张图的结果正常吗

我遇到了相同的问题，loss正常收敛，detect也正常

已解决。在Windows下需要使用7z打包命令，我在打包的时候指令有错误，导致submit.zip是空的