使用isic数据集跑出来效果很奇怪

[erlingzz]

你好，我使用isic数据集训练模型，大约跑了20w个step，测试时num_ensemble设置为1，部分生成的分割图很奇怪，如下（左为gt，右为测试结果）：

请问这是正常的情况嘛？

[GuoHuike7]

20w轮训了多久呀，我这采样一张2分钟

[erlingzz]

20w轮训了多久呀，我这采样一张2分钟

我在3090上训练大概训练了2天，我采样也差不多这么慢，或许我们可以交流一下嘛？我的邮箱是523291461@qq.com

[GuoHuike7]

加你qq了，方便通过一下嘛

[WuJunde]

don't sample 1 time, ensemble is very important for the performance. but this result is unexpectedly bad for even 1 sample.

[erlingzz]

don't sample 1 time, ensemble is very important for the performance. but this result is unexpectedly bad for even 1 sample.

Thank you for your answer. Could you please tell me how many steps you ran on the ISIC dataset, and what the final iou and dice are?

[WuJunde]

20w step about 89% dice in my workplace

[ValeriaLiu]

20w step about 89% dice in my workplace

I work on my own dataset and I find that as the steps increase, the train loss will not change significantly but the prediction's accuracy of the model will increase. Have you ever encountered such a similar situation? Look forward to your reply.

[WuJunde]

yes, it happens. train more until it stuck