Couldn't understand how to test on one instance

[billel-a]

Hello i was trying to implement this model on a small application im having two problems:
1- is there a function that i d give the model and the niftii of one patient or the multiple niftii of one patient and it returns a segmented nifti something like test( OneData,Model) : segmentation or isn't there anything like that in this git.
2- Can this run on cpu only or do i need a gpu

Thanks in advance
Great work

[himashi92]

You can use test.py https://github.com/himashi92/vizviva_brats_2021/blob/main/test.py to make predictions for the given patient data. In this repo, I primarily focus on brain tumor segmentation and patient data comprises multiple modalities.

Yes, you can use a CPU, but I recommend using GPU for both training and inference time.

[billel-a]

Since there is the pretrained model I was thinking if im not doing a retraining but a simple one data segmentation it ll be fast even on gpu the problem i had working with your test.py is :
line 85 - 86 you have this :
if ngpus == 0:
raise RuntimeWarning("This will not be able to run on CPU only")

and to use test.py do I have to put data in the folder dataset ? cause I was looking to just load them by path each time

[himashi92]

Remove that code fragment. Try using device = torch.device("cpu"), and load model to cpu instead of gpu like this model.to(device).

You have to give your folder dataset path for input_dir parameter https://github.com/himashi92/vizviva_brats_2021/blob/main/config.py#L6

[billel-a]

yeah but this is used to do a bunch of predictions like all the test data or train data or ...
Sorry to bother you but what im looking for is just one data prediction im working on a niftii visualizer where he ll chose from the GUI the file he wants and i want based on the file he chose i do the prediction and segmentation just for that.
Can you tell me what i have to do ?
And thanks so much for your help

[himashi92]

What is the dataset you gonna use? Is it this brats dataset?
So basically, this model works for multi-modal nifty images. Therefore, when you choose a nifty file you have to choose the folder which has four modalities in it, not a single nifty file. For example, if you choose the patient_01's folder (that is BraTS2021_00000), the file structure looks like this.

BraTS2021_00000
|--------------BraTS2021_00000_flair.nii.gz
|--------------BraTS2021_00000_t1.nii.gz
|--------------BraTS2021_00000_t1ce.nii.gz
|--------------BraTS2021_00000_t2.nii.gz
|--------------BraTS2021_00000_seg.nii.gz

So the chosen input dir from GUI should be BraTS2021_00000 folder path.

[billel-a]

It shouldnt use the file : BraTS2021_00000_seg.nii.gz isn't this the result file what the model should give as output ?
and for the rest yes i ll be giving it :
|--------------BraTS2021_00000_flair.nii.gz
|--------------BraTS2021_00000_t1.nii.gz
|--------------BraTS2021_00000_t1ce.nii.gz
|--------------BraTS2021_00000_t2.nii.gz

[himashi92]

yes, you are right. But this will be handled by code.

[billel-a]

Thanks so much