Currently for the 3D case I take a random 3D patch from an entire 3D sample, and this counts as one datapoint (which will be different every time). That means epochs are very very short, while in the 2D case one 3D sample has a datapoint for every horizontal slice.
Should probably base things on number of data points rather than epochs, and make sure that for example the frequency of validation makes sense in both cases.
Currently for the 3D case I take a random 3D patch from an entire 3D sample, and this counts as one datapoint (which will be different every time). That means epochs are very very short, while in the 2D case one 3D sample has a datapoint for every horizontal slice.
Should probably base things on number of data points rather than epochs, and make sure that for example the frequency of validation makes sense in both cases.