search

XiaofeiWang2018 / DR

9 stars 3 forks source link

about the number of epochs #1

Closed QueenZK520 closed 3 years ago

QueenZK520 commented 3 years ago

您好，请问在第一阶段训练中，epoch数真的是30000吗？

XiaofeiWang2018 commented 3 years ago

Thanks for your attention. A good question. In fact, you can train as many epoches as you wish till the model converges to the preferred performance. So, definitely 30000 is an option, but not mandatory.

XiaofeiWang2018 commented 3 years ago

Thanks for your attention. A good question. In fact, you can train as many epoches as you wish till the model converges to the preferred performance. So, definitely 30000 is an option, but not mandatory.

QueenZK520 commented 3 years ago

I also want to know have you done any preprocessing for original images in DDR. I see that you load data from the folder named _imgafter2.

QueenZK520 commented 3 years ago

I find that I get very low IOU in each lesion, especially in MA. May you share with me the value of IOU or DICE you get with your method.

XiaofeiWang2018 commented 3 years ago

As for the preprocessing, certainly, proper preprocessing, e.g., black side cropping and illumination balance, is common and necessary for fundus image processing. As for the IoU, I haven't test for it, you can just refer to the paper that estblished the DDR. BTW, I think one possible reason that causes the low IoU may be the improper process of the GT maps of DDR, which I remember are grayscaled but not binarized. In fact, I am now out of office for some tests these weeks and am quite busy with my phd application and graduation recently.

QueenZK520 commented 3 years ago

Thank you very much! You are my idol! Thanks again.