Does Xavier initializations makes sense for 3d data?

[Roger]

Hi, 
I am not sure if Xavier initialization does not make sense for 3D data,
but if it does, 
There will be this error:
Check failed: num_axes() <= 4 (5 vs. 4) Cannot use legacy accessors on Blobs with > 4 axes.

Which happens because in the filler we use the legacy shape (but we only use num and channels which would be shape[0] and shape[1])

Is xavier initiallization only for 2d data, or can it just use the new shape accessors?

Thanks!

[Ernst Schwartz]

Hi!

Any updates on this? I'm having the same issue ...