weird scores during training

[wmh1203]

Thank you for providing this wonderful work! I'm trying to reproduce it on the provided GZMH dataset. After constructing VOC format dataset, I successfully start training the model. However, the scores (mAP, recall, ap, ...) are all 0 even after several epochs. I'm wondering whether there is something wrong with my dataset design. The following is my setting.

For my annotation xml file, since there are no sample files, I construct them like this.

Since I can run the code, I'm wondering why the metric scores are all zero. It seems that the model doesn't learn anything. Could you please give some insights for this? Thanks a lot.

[wmh1203]

Here is the screenshot during training. Thanks for your help.

[hewg2008]

Please add my WeChat: 18666957143.

[antifen]

  Thank you for your interest in our paper.   For the dataset design, JPEG Images store pictures with .jpg suffix (that is, all the pictures contained in test.txt and train.txt) ,Refer to the data.png file attached to the email.   As for the xml file stored in Annotations, it is necessary to cut the HPF with the size of 2084×2084 into a patch with the size of 224×224, so that the network can learn the features better. See annotation.png for details.   According to the fact that the number of mitotic nuclei in ground truth is 0, the most likely reason is that the network has not read the dataset, because the number of mitotic nuclei in ground truth is fixed in the GZMH dataset and does not need to be detected by the network.

  Best regards.

 

------------------ 原始邮件 ------------------ 发件人: "antifen/mitosis-nuclei-detection" @.>; 发送时间: 2024年2月24日(星期六) 下午3:51 @.>; @.***>; 主题: [antifen/mitosis-nuclei-detection] weird scores during training (Issue #3)

Thank you for providing this wonderful work! I'm trying to reproduce it on the provided GZMH dataset. After constructing VOC format dataset, I successfully start training the model. However, the scores (mAP, recall, ap, ...) are all 0 even after several epochs. I'm wondering whether there is something wrong with my dataset design. The following is my setting.

image.png (view on web)

For my annotation xml file, since there are no sample files, I construct them like this.

image.png (view on web)

Since I can run the code, I'm wondering why the metric scores are all zero. It seems that the model doesn't learn anything. Could you please give some insights for this? Thanks a lot.

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

[wmh1203]

  Thank you for your interest in our paper.   For the dataset design, JPEG Images store pictures with .jpg suffix (that is, all the pictures contained in test.txt and train.txt) ,Refer to the data.png file attached to the email.   As for the xml file stored in Annotations, it is necessary to cut the HPF with the size of 2084×2084 into a patch with the size of 224×224, so that the network can learn the features better. See annotation.png for details.   According to the fact that the number of mitotic nuclei in ground truth is 0, the most likely reason is that the network has not read the dataset, because the number of mitotic nuclei in ground truth is fixed in the GZMH dataset and does not need to be detected by the network.   Best regards.   …

Thanks for the reply and sorry for the late response since I'm also working on other projects.🙏 I'll try this again later and follow up soon. By the way, it seems that the attached files are not included. Could you please resend it? My email is bill.wang@connect.ust.hk Thanks again for the help!

[shbkukuk]

Hi @wmh1203 , Can you solve the problem that obtain the scores all zeros ?