Power of two dimensions

[wanderine]

In the colab example for training a 3D GAN, will the conversion to TFrecord include resizing to power of 2 dimensions in each direction? Most datasets do not have volumes that are a power of 2 in each direction, so could be nice to automatically do that in the conversion to TFrecords. I suppose it is harder to train a 3D GAN that can synthesize volumes of any size.

[satra]

@wanderine - we can indeed automate some version of this. nibabel has a conform function that makes everything isometric, which we can reuse. we are using some freesurfer tools in other models to do this. but indeed we should be able to do this when creating the dataset.

[kaczmarj]

In the colab example for training a 3D GAN, will the conversion to TFrecord include resizing to power of 2 dimensions in each direction?

if i understand the question correctly, yes, that example will resize each scan to the desired dimension. the dimensions created by that notebook would be 8x8x8, 16x16x16, 32x32x32, etc. a set of tfrecord files is created for each dimension, and each tfrecord file has a few scans of the same dimension.

however that's a special case... we only resize images in the multi-resolution case. i agree with @satra that we should allow this. i agree that most datasets will not be a common size.