You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Is it normal that almost res reports are the same？ I also find that the more epoches finished , the same reports are less . But after 20 epoched finished , they are also so many same reports.
the example_datas are npz,and the full datas are nii.gz, i use #img =nib.load(path) #img_data=img.get_fdata() to read the nii.gz files,i think it is correct. or are there other methods to read the nii.gz files.
May I ask how much data volume you are using for reproduction, I am generating the same report with default hyperparameters using a data size of 3000 patients
Thank you all for bringing up this issue. The problem of generating similar or identical reports is often related to the model's temperature setting during generation. Adjusting the temperature parameter can help introduce more variability in the outputs. I recommend trying to set the temperature to 1.5 and see if that resolves the issue.
Additionally, the problem might also be related to the size of the training dataset. If you're working with a smaller dataset, the model may not have enough diversity in the data to produce varied reports. If possible, try increasing the dataset size to see if it improves the results.
@ibrahimethemhamamci Thanks for your reply, I have set the temperature parameter to different values, but the report generated in the verification stage is still the same. I checked the relevant problems, which seem to occur in visual extraction, and found that the features extracted by the model are very repetitive for different images, which makes it difficult for the model to distinguish the differences between different images. This is just my inference, I do not know whether it is correct, I would like to ask if you used the pre-training model to fine-tune your project, or did you start training from scratch?
@ibrahimethemhamamci Thanks for your reply, I have set the temperature parameter to different values, but the report generated in the verification stage is still the same. I checked the relevant problems, which seem to occur in visual extraction, and found that the features extracted by the model are very repetitive for different images, which makes it difficult for the model to distinguish the differences between different images. This is just my inference, I do not know whether it is correct, I would like to ask if you used the pre-training model to fine-tune your project, or did you start training from scratch?
Hi , thank you for your insight , I would like to ask, which report.xlsx did you use to train your model?... and how many images did you use ?
@hari3100 Hi I am using my own small 3D medical image data, and I am trying to extract the image features using the new vision extractor
I feel that is were your issue is originating from, at least as much as I understand the CT2Rep model , it trains on the CT scans in .npz format and also each of those images are associated with a medical report using the accession number, you can try looking the structure in the example dataset given , and the Full dataset on CT-RATE page that requires us to create the structure.
So you used the data set mentioned in the paper? Could you please show me the text description you generated for the validation sample with this data set during the model validation phase?thank you @
@hari3100 Hi I am using my own small 3D medical image data, and I am trying to extract the image features using the new vision extractor
I feel that is were your issue is originating from, at least as much as I understand the CT2Rep model , it trains on the CT scans in .npz format and also each of those images are associated with a medical report using the accession number, you can try looking the structure in the example dataset given , and the Full dataset on CT-RATE page that requires us to create the structure.
So you used the data set mentioned in the paper? Could you please show me the text description you generated for the validation sample with this data set during the model validation phase?thank you @
@hari3100 Hi I am using my own small 3D medical image data, and I am trying to extract the image features using the new vision extractor
I feel that is were your issue is originating from, at least as much as I understand the CT2Rep model , it trains on the CT scans in .npz format and also each of those images are associated with a medical report using the accession number, you can try looking the structure in the example dataset given, and the Full dataset on CT-RATE page that requires us to create the structure.
Hi, Sadly I have not yet trained the model using the mentioned dataset, Me and my team are still structuring the dataset, and figuring out how to manage such a huge dataset with each file being this big around 150MB each.
But I have just tried training the model using the example dataset and the provided example_report.xlsx. The results are useless as that dataset has only a few images and three reports.
感谢您的回复，我已将 temperature 参数设置为不同的值，但在验证阶段生成的报告仍然相同。我检查了相关问题，这些问题似乎发生在视觉提取中，发现模型提取的特征对于不同的图像来说非常重复，这使得模型很难区分不同图像之间的差异。这只是我的推论，不知道是否正确，我想问一下你是用预训练模型来微调你的项目，还是从头开始训练的？
Have you already solved this problem? Can you share your method?
Activity
1benv commentedon Jul 19, 2024
the example_datas are npz,and the full datas are nii.gz, i use #img =nib.load(path) #img_data=img.get_fdata() to read the nii.gz files,i think it is correct. or are there other methods to read the nii.gz files.
jonathanzxshi commentedon Aug 1, 2024
I also find this problem, so I also use nib.load and get_fdata, but i dont know how to generate the report after training. Can you help me?
Andyfever123 commentedon Aug 11, 2024
I encountered the same issue. Has anyone found a solution? I'm not quite sure what the cause is."
hugowang010406 commentedon Aug 12, 2024
May I ask how much data volume you are using for reproduction, I am generating the same report with default hyperparameters using a data size of 3000 patients
sunnyyy17 commentedon Aug 13, 2024
Im also getting similar results with res_reports being almost the same even with the full dataset provided in CT-RATE huggingface repository.
ibrahimethemhamamci commentedon Aug 20, 2024
Thank you all for bringing up this issue. The problem of generating similar or identical reports is often related to the model's temperature setting during generation. Adjusting the temperature parameter can help introduce more variability in the outputs. I recommend trying to set the temperature to 1.5 and see if that resolves the issue.
Additionally, the problem might also be related to the size of the training dataset. If you're working with a smaller dataset, the model may not have enough diversity in the data to produce varied reports. If possible, try increasing the dataset size to see if it improves the results.
Please let me know if these suggestions help!
Andyfever123 commentedon Sep 26, 2024
@ibrahimethemhamamci Thanks for your reply, I have set the temperature parameter to different values, but the report generated in the verification stage is still the same. I checked the relevant problems, which seem to occur in visual extraction, and found that the features extracted by the model are very repetitive for different images, which makes it difficult for the model to distinguish the differences between different images. This is just my inference, I do not know whether it is correct, I would like to ask if you used the pre-training model to fine-tune your project, or did you start training from scratch?
hari3100 commentedon Sep 27, 2024
Hi , thank you for your insight , I would like to ask, which report.xlsx did you use to train your model?... and how many images did you use ?
Andyfever123 commentedon Sep 27, 2024
@hari3100 Hi
I am using my own small 3D medical image data, and I am trying to extract the image features using the new vision extractor
hari3100 commentedon Sep 27, 2024
I feel that is were your issue is originating from, at least as much as I understand the CT2Rep model , it trains on the CT scans in .npz format and also each of those images are associated with a medical report using the accession number, you can try looking the structure in the example dataset given , and the Full dataset on CT-RATE page that requires us to create the structure.
Andyfever123 commentedon Sep 27, 2024
So you used the data set mentioned in the paper? Could you please show me the text description you generated for the validation sample with this data set during the model validation phase?thank you @
hari3100 commentedon Sep 27, 2024
Hi, Sadly I have not yet trained the model using the mentioned dataset, Me and my team are still structuring the dataset, and figuring out how to manage such a huge dataset with each file being this big around 150MB each.
But I have just tried training the model using the example dataset and the provided example_report.xlsx. The results are useless as that dataset has only a few images and three reports.
Coldermyy commentedon Mar 9, 2025