uni-medical / stu-net Goto Github PK

Dear authors,

I wonder if you could help with clarifying how to conduct direct evaluation using STUNet? I have tried on several MSD datasets but the results were clearly wrong. Did you use the original nnUNetplans from total segmentor dataset?

Thanks!

I have put your files into the specific folder of nnunet, but I got this error raise RuntimeError("Could not find trainer class in nnunet.training.network_training") RuntimeError: Could not find trainer class in nnunet.training.network_training. It seems program can not find the trainer class, how can I do?

Hello！
I used a huge model to do finetune training. I had 80g of gpu memory, and still reported errors exceeding gpu memory, but when I looked at the gpu memory usage, the peak gpu memory only reached 30. How to solve this problem? Thank you!
RuntimeError: CUDA out of memory. Tried to allocate 1.25 GiB (GPU 0; 44.56 GiB total capacity; 41.63 GiB already allocated; 217.56 MiB free; 42.35 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF Exception in thread Thread-6:

I am trying to reproduce your fine-tune results on the Amos 2022 dataset using STU-Net_small. What is the best hyper-parameter setting for this model and dataset. Much appreciated if you can provide!

Dear authors,

thanks a lot for your work. It is great to see the code and the weights for the proposed approach. May I ask if you could share weights for other approaches in the paper? E.g. the SWIN-UNETR models you trained on the TotalSegmentator dataset.

Dear author, thanks for your impressive work.
Could you please offer some commands for how to pertain the mode on the totalsegmentator dataset?

Dear author, thank you so much for your work.
I'm impressed by your work and try to infer directly with your pretrained models. but things didn't work out for me.

I have followed your instructions in direct_inference and set up the environment. Then I used the example command python direct_inference.py STUNetTrainer_small example/Task032_AMOS22_Task1 example/result just as is written in your instructions. At first I thought this process would be smooth because I saw the terminal printed out

'...starting preprocessing generator
starting prediction...'

but suddenly an error occured:
File "D:\anaconda3\envs\pyTorch39\lib\multiprocessing\reduction.py", line 60, in dump ForkingPickler(file, protocol).dump(obj) _pickle.PicklingError: Can't pickle <function <lambda> at 0x0000022D9F31155 0>: attribute lookup <lambda> on nnunet.utilities.nd_softmax failed [W CudaIPCTypes.cpp:16] Producer process has been terminated before all sha red CUDA tensors released. See Note [Sharing CUDA tensors]

File "D:\anaconda3\envs\pyTorch39\lib\multiprocessing\spawn.py", line 126 , in _main self = reduction.pickle.load(from_parent) EOFError: Ran out of input

the error message is quite long so I only paste the latest lines here.

I have searched for this error and it is said that

this error typically occurs when the pickle module in Python tries to load an empty file or a file that has been truncated. It means that the end-of-file was reached unexpectedly while there was still data expected to be read. This can happen if the file is empty, corrupted, or if there is a mismatch between how the data was written and how it is being read.

I have checked the RESULT_FOLDER's structure, but it was exactly the same as the structure in your instructions. and your pkl files must be fine. With all the possibilities excluded, now I have no clue why this happens. So I came to ask for your help.

I would appreciate it if you could offer some suggestions.
Thank you very much!

Thank you for sharing awesome work!

I did fine tuning for my dataset with pre-trained weight(small and base model)
It shows me great result.

Can you share the training script used to train the TotalsSegmentator dataset?
Or is it sufficient to use the method in this link?
https://github.com/wasserth/TotalSegmentator/blob/master/resources/train_nnunet.md

I would like to add more datasets to the pre-training model to maximise performance.

like Verse and CTPelvic1K or more,
Of course, I may have to change the information on their labels.

Hello Ziyan,

Thanks you very much for the repo.

I am beginner and I am trying to understand Continous learning models.

I have a few question, I have read that in Continuous Deep Learning models there is an issue with Catastrophic Forgetfulness. How are you able to over come this issue. I apologize if my question is naive.

Also I was not able to download weights for FLARE23 from Baidu, could you please upload them to gdrive. I would like to try training on TCIA NSCLC dataset.

Regards,
Anil

I saw that you have a nnUNet v2 folder but the README only talks about nnUNet v1. May I ask if the pretrained weights could be used for finetuning nnUNet v2 ?

Thank you

Thank you for sharing! Can I apply the pre-trained parameters and architecture of STU to a new dataset (not the previously mentioned 104 types of segmentation), as the features needed for segmentation are generic and I only need to fine-tune the output layer?

Hi ! I'm interesting in your excelent work. In STU-Net, the final 1×1×1 convolution layer for segmentation output has 105 channels, but the TotalSegmentator dataset contains 104 anatomical structures (104 classes), So is there a channel that contains a segmentation mask for all foreground ?

Hello, after executing run_finetuning.py, the program unpacked dataset card for several hours, is this a normal phenomenon?

I want to ask whether the fine-tune dataset you used in your work are just follow the preprocessing method provided by nnUNet_v1？Thank you very much if you can help me!

how to train in 2 gpus?

Hi,

Thanks for sharing your great work with us. I just parsed the .pkl file you uploaded a while ago. And I found that there were no params of HU normalization (it was stored with the key "intensityproperties" based on my previous experience of nnunet). Also, I haven't found anything related to the process of HU normalization in your paper. I'm wondering if you did such HU normalization for CT images during inference and how you did that.

I want to use totalsegmentor V2 for a new pre-training, but there is only run_trainning.py for nnUNet in the run folder. How do I start re-training?

Hello, I am truly sorry for interrupting your work, and I really apologize for the disturbance. May I kindly ask about the network you used during the ABUS competition? I am really sorry for any inconvenience caused.

I am trying to reproduce STU-NET results and would like to have the same evaluation scripts you used to compute the dice score for the WORDS and FLARE datasets. I would appreciate it if the authors could make these scripts available.
Best

The paper mentions the learning rate * 0.1 for the rest of the segmentation head. How is it reflected in the code? I can't find the corresponding part. If I want to make the learning rate of the segmentation head 0.01 and the rest *0.1, I should set self.initial_lr in STUNetTraniner_ft in STUNetTrainer.py to 0.01 or 0.001. My English may be so poor, please forgive me. Thank you very much!

Really nice work! Would it be possible for you to kindly provide the '<model_name>.model.pkl' file along with each '<model_name>.model' file, as it is required for running inference of your models? Thank you very much.

I wonder if the model Surpport 2D segmentation

i want run this model in DDP,how can i get it

Is there any overlap of labels for different categories in the TotalSegmentator dataset? I noticed that some pixels seem to belong to multiple categories.

Really nice work!
I'm trying to use my private dataset to finetune the network, but it required a pkl file. I'm trying to use base_ep4k.model.pkl, but it raised an issue "KeyError: 'plans_per_stage'". Which pkl file should I provide here?

crop size is patch size？ And what is the relationship between crop size and patch size?
Thanks.