Closed rprueckl closed 2 years ago
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?
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...
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."
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,
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
execute quicknat
Thanks! I'll try it!
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.
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!