Out of Memory

[rprueckl]

Hi,

I tried to run QuickNAT on a dataset that I have preprocessed using Freesurfer as described in your documentation. My computer is equipped with a NVIDIA GeForce GTX 1650 (4GB RAM). I am using python run.py --mode=eval_bulk to start execution. I have set the batch size to 1, however I am still getting this:

RuntimeError: CUDA out of memory. Tried to allocate 16.00 MiB (GPU 0; 4.00 GiB total capacity; 1.67 GiB already allocated; 0 bytes free; 2.74 GiB reserved in total by PyTorch)

Is it possible to work around this problem somehow or are 4GB RAM simply not enough for executing QuickNAT?

Thanks for your time!

[rprueckl]

Hi again,

today I tried with an RTX 2080 (8GB) with a similar result:

RuntimeError: CUDA out of memory. Tried to allocate 2.06 GiB (GPU 0; 8.00 GiB total capacity; 4.21 GiB already allocated; 0 bytes free; 6.28 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

I think in the paper a GPU with 12GB was used?

[fhfhfh999]

Hi, Seems you have run the code. But I cannot understand how the 'converth5.py' works especially the labels. Could you give me a hint? I think the author of this paper will never answer any questions...

[fhfhfh999]

The paper mentioned in 2.3: "We use a constant weight decay of 0.0001. Batch size is set to 4, limited by the 12GB RAM of the NVIDIA TITAN X Pascal GPU."

[fhfhfh999]

The paper mentioned that they use FreeSurfer to handle IXI Dataset. But when I started to learn FreeSurfer, I found that FreeSurfer will not give a single "Auxiliary label". The output of FreeSurfer contains many files including a folder named "label". And I think the folder is not the "label" in "convert_h5.py". So, how to start training?

[rprueckl]

Hi,

I never executed training, only segmentation. I preprocessed my niftis with mri_convert --conform <input.nii> <output.nii>

Regarding executing QuickNAT, my steps are as follows:

• install quicknat

  • install cuda 11.3
  • install python 3.7.9 (x64)
  • make sure the correct python is in path
  • create a folder D:/quicknat_test
  • copy the folder called after the github commit hash to D:/quicknat_test/src
  • edit "settings_eval.ini" - change the following: data_dir = "D:/quicknat_test/nifti/process" directory_struct = "Linear" estimate_uncertainty = "True"
  • start cmd with admin permissions
  • install virtualenv for python (if not already done) pip install virtualenv
  • create a virtual environment virtualenv D:/quicknat_test/env
  • activate the virtual environment D:\quicknat_test\env\Scripts\activate.bat
  • go to the folder D:\quicknat_test\src\4e4e97e912b9f75f9c299065922009da737c4ef9
  • install correct torch pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu113
  • install the rest of the dependencies python -m pip install -r requirements.txt

• execute quicknat

  • copy the preprocessed nifti files into D:/quicknat_test/nifti/process
  • edit "test_list.txt" and enter the filenames in the data_dir you want to process
  • start cmd with admin permissions
  • activate the virtual environment D:\quicknat_test\env\Scripts\activate.bat
  • go to the src folder: cd D:\quicknat_test\src\4e4e97e912b9f75f9c299065922009da737c4ef9
  • start processing python run.py --mode=eval_bulk
  • results under: D:\quicknat_test\src\4e4e97e912b9f75f9c299065922009da737c4ef9\ixi_test_seg\one_view

[fhfhfh999]

Thanks! I'll try it!

[arickm]

The paper mentioned that they use FreeSurfer to handle IXI Dataset. But when I started to learn FreeSurfer, I found that FreeSurfer will not give a single "Auxiliary label". The output of FreeSurfer contains many files including a folder named "label". And I think the folder is not the "label" in "convert_h5.py". So, how to start training?

Hi, for training using FreeSurfer segmentations you can use the segmentation file: mri/aseg.mgz which contains the segmentation of subcortical structures used in QuickNat, and the mri volume: mri/orig.mgz. In utils/preprocessor.py is a function remap_labels that shows which of the classes were used.