fastsurferCNN eval.py memory usage about fastsurfer HOT 7 CLOSED

Hey,
I just find out a solution to decrease the memory spending. Since 3 variable：prediction_probability_axial，prediction_probability_coronal， prediction_probability_sagittal is no longer used after
_, prediction_image = torch.max(torch.add(torch.mul(torch.add(prediction_probability_axial, prediction_probability_coronal), 0.4), torch.mul(prediction_probability_sagittal, 0.2)), 3)
We can use
a = torch.add(prediction_probability_axial, prediction_probability_coronal) del prediction_probability_axial,prediction_probability_coronal
to save the memory used by the two variables. From my experience, this step save 5GB memory and the script will fit in well with 32GB ram system.

Best,
Boyang

from fastsurfer.

Hey,

I will have a look at it, but it is possible that the view aggregation does indeed need a lot of CPU memory due to the large number of feature maps (95 classes).

Best,
Leonie

from fastsurfer.

Thanks a lot. And may I ask how many memory at least is required to run fastsurferCNN/eval.py for one image with size like (179x256x256)?

from fastsurfer.

Your RAM should be sufficient (at least if you run step i CNN segmentation on the GPU, where you might need to reduce batch size, e.g. try 8). On our GPU it takes 40 sec with 8GB RAM (RTX 2080). For the part ii, 32 Gig CPU RAM should really be enough. Only the view aggregation seems to be very memory hungry (probably around 30Gig) so could be that that is the problem if other stuff is loaded.

from fastsurfer.

OK, thanks, I'll try that. Another thing is that does your network support nifti format input? We use nifti format to store our brain mri images.

from fastsurfer.

Hey,

yes, nifti format is supported as well (see also #20 ).

Best,
Leonie

from fastsurfer.

Hey Boyang,

thank you :). I added the change with commit 25609f2. In total, this makes the view aggregation slightly slower (2-4 s on average on my machine). I would however consider this as tolerable given that it assures usage on the 32GB ram system.

Best,
Leonie

from fastsurfer.