search

sandialabs / bcnn

3D Bayesian Convolutional Neural Network (BCNN) for Credible Geometric Uncertainty. Code for the paper: https://arxiv.org/abs/1910.10793
MIT License
60 stars 19 forks source link

miss test_coords.pickle in 2D BCNN #3

Closed kai-zhang-er closed 4 years ago

kai-zhang-er commented 4 years ago

How to generate test_coords.pickle for 2D BCNN? Does it have any impact on the prediction procedure? Thank you.

tmlabonte commented 4 years ago

test_coords.pickle is a file containing the original coordinates of each "chunk" of a 3D volume inside the larger original volume. We predict on each chunk individually, then refer to test_coords.pickle to reconstruct the entire prediction on the original volume. For a 2D BCNN, test_coords.pickle will not be generated because you do not need to chunk 2D data (it is usually small enough to fit in memory already).

Are you running into an error?

kai-zhang-er commented 4 years ago

I have solved the problem. Thank you very much. Would you like to publish your pretrained model for the 2D BCNN? I want to use the uncertainty map to evaluate my experiment.

tmlabonte commented 4 years ago

Unfortunately we are unable to publish our pretrained models as they are trained on proprietary datasets. Feel free to use the instructions in the README to train on your own dataset.