Andrew Hoopes

You can think of the HyperConvFromDense layer as a functional equivalent to a regular Conv layer, with the main exception that its internal weights are predicted based on the values...

Yes, if we could see some example code it would help in replicating the problem. So, it seems like every thing works okay until you specify validation data? In which...

Yes, probably the case. What are your command line arguments (and path to training images)?

The `add_identity` parameter has been removed in recent versions since it's now assumed that the affine is always 'over' the identity. If you remove that from your code, do things...

At the moment, `load_pheno_csv` assumes the inputs have been pre-normalized and doesn't have any way to differentiate between categorical vs continuous parameters. This is definitely a feature we should add,...

@kvttt, @tenres sorry for losing track of this yes, but yes, I would generally recommend normalizing in all cases, including age, for stability purposes. @adalca do you agree?

thanks @kousu this is super useful! you might have noticed we just uploaded a new version (0.2) to pypi, but we're happy to include this for an upcoming 0.3 release....

In vxm utilities, the affine matrix shape should always be `(N, N+1)`, where `N = ndim`, for consistency. So you can crop your 2D input affine matrix to shape (2,...

@ixy40 originally, the model files used in the tutorial were just pure weights-only files - they didn't contain any information about model structure. That's why in the tutorial, the model...

If you plan to use your own data on pre-trained models, then you should ensure that your inputs have been preprocessed in the same manner as the images used for...