How to enable CUDA acceleration and benchmark metric of single image pair

[jakubMitura14]

Hello I am comparing Monai metrics as gold standard to some other tools - for this purpose I want to run the metrics like Dice metric ... with CUDA acceleration on given images - as an example i took whole body bone segmentation where I compare original to translated image. Metric evaluates but I do not see how to switch on CUDA acceleration for it and synchronize GPU in order to get valid benchmark results.

I had seen this example https://github.com/Project-MONAI/tutorials/blob/master/acceleration/fast_training_tutorial.ipynb where optimizers and transforms are done with CUDA acceleration, but I am not sure what about metrics.

import monai
from monai.metrics import Cumulative
dice_metric = DiceMetric(include_background=True, reduction="mean")
dice_metric(y_pred=not_translated[0, :], y=translated[0, :])  # callable to add metric to the buffer
metric = dice_metric.aggregate().item() 

[Nic-Ma]

Hi @jakubMitura14 ,

Thanks for your interest and experiments here.
Actually, when you compute metrics for the GPU Tensor data(image, label), it executes on GPU.

Thanks.

[jakubMitura14]

Thank You do you have an example how to switch normal tensor to gpu tensor?

[Nic-Ma]

Hi @jakubMitura14 ,

That's PyTorch feature, just call something like:

a = torch.Tensor([XXX])
a.to(torch.device("cuda"))

Thanks.

[jakubMitura14]

Thanks it works !