'Predict' VRAM usage for prediction.

[ExtReMLapin]

Hello, working on R-50 models, (detection/segmentation).

Is there any more-or-less reliable way to "guess"/predict how much memory a model will use, based on the number of images sent thru it.

I get around 380-417mb of VRAM used per images sent inside the batch (for a 1024102424bits image)

Thanks

[ExtReMLapin]

pls no pun with the title