search

CSCYQJ / MICCAI23-ProtoContra-SFDA

This is the official code of MICCAI23 paper "Source-Free Domain Adaptation for Medical Image Segmentation via Prototype-Anchored Feature Alignment and Contrastive Learning"
26 stars 6 forks source link

CL阶段性能下降 #10

Open Mswang12398464635 opened 11 months ago

Mswang12398464635 commented 11 months ago

论文方向很有意义。但在代码运行时，CT域的预训练阶段和MRI域的PFA阶段效果都不错，可是CL阶段却只有0.84不到。如果直接将预训练模型用于CL阶段，甚至达不到0.5。可以请教一下这是怎么回事嘛，非常感谢！

Mswang12398464635 commented 11 months ago

我CT到MR方向CL阶段复现结果才83.47%，还没有你的高呢，不知道怎么回事。我PFA阶段也才84.08％，你是有修改什么吗？

---原始邮件--- 发件人: @.> 发送时间: 2023年12月18日(周一) 晚上9:01 收件人: @.>; 抄送: @.**@.>; 主题: Re: [CSCYQJ/MICCAI23-ProtoContra-SFDA] CL阶段性能下降 (Issue #10)

我也有这个疑问，我CT到MR方向CL阶段复现结果为84.26%，MR到CT的为71.07%，和论文结果还有一些差距。还需要注意其他什么因素的影响吗？

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

zqp1226358 commented 11 months ago

@Mswang12398464635 我CT到MR方向，源95.35%，PFA阶段是83.89%，CL是84.26%，也没动什么，和作者应该是差不多一样的设置，除了batchsize小一些。

Mswang12398464635 commented 11 months ago

batchsize具体是多少呀，你把每个阶段的batchsize都改了吗？这个复现还挺困难，我看评论里大家好像都这样，不能复现到预期结果。有点害怕😨

---原始邮件--- 发件人: @.> 发送时间: 2023年12月20日(周三) 上午10:54 收件人: @.>; 抄送: @.**@.>; 主题: Re: [CSCYQJ/MICCAI23-ProtoContra-SFDA] CL阶段性能下降 (Issue #10)

@Mswang12398464635 我CT到MR方向，源95.35%，PFA阶段是83.89%，CL是84.26%，也没动什么，和作者应该是差不多一样的设置，除了batchsize小一些。

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>

zqp1226358 commented 11 months ago

7 5 5，我显卡跑不了太大的batchsize

zqp1226358 commented 11 months ago

@Mswang12398464635 你有尝试另外一个方向的表现吗？

Mswang12398464635 commented 11 months ago

我还没尝试，因为我一开始以为两个方向肯定是大同小异，我把这个方向整好，另一个应该也差不多。但看到你的结果，我觉得得换一下了，我准备把batchsize改了之后看看结果，然后换另一个方向。

---原始邮件--- 发件人: @.> 发送时间: 2023年12月20日(周三) 中午11:28 收件人: @.>; 抄送: @.**@.>; 主题: Re: [CSCYQJ/MICCAI23-ProtoContra-SFDA] CL阶段性能下降 (Issue #10)

@Mswang12398464635 你有尝试另外一个方向的表现吗？

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>

qwerasdzxcvb commented 6 months ago

@zqp1226358 我想请问一下，你有尝试另一个方向吗， 在MRtoCT，我在PFA阶段只有0.76，结果很低是正常的吗 但是在CTtoMR方向PFA结果有0.83左右

BarY7 commented 6 months ago

Hi @qwerasdzxcvb , I am also trying to reproduce the results in the paper, so far on my side the runs are very unstable - never quite reaching the paper's results, at best about 3-4% down.

Do you want to share your preprocessing code? perhaps we could try debugging it together. I also tried using the data preproceseed in SIFA:

https://github.com/cchen-cc/SIFA?tab=readme-ov-file

But results remain unstable.

qwerasdzxcvb commented 6 months ago

@BarY7 我用的作者分享的preprocess代码，在作者提交的历史记录里边

BarY7 commented 6 months ago

@qwerasdzxcvb I now also used the preprocessing notebook CT to MR direction: After PFA, the model is at 86% . But, the model had high no adaptation score to start with - 75%.

MR to CT: Results are much worse, no adaptation score is very low - about 36-40%. PFA sometimes brings the results to 70 area, sometimes doesn't change the results much.