Remove option for preprocess_bijection as it could potentially cause confusion. For example, a user may expect to be able to run train_flow twice in order to train for more epochs (e.g. rerunning a cell containing train_flow). However this would not be the case when using a preprocess_bijection, as the flow outputted from the first call to train_flow would already be transformed by Invert(preprocess_bijection), and the output from the second call would have the inverse transformation applied twice.
Instead, an example is added to the FAQ for scaling/preprocessing, which is more explicit.
Remove option for
preprocess_bijectionas it could potentially cause confusion. For example, a user may expect to be able to runtrain_flowtwice in order to train for more epochs (e.g. rerunning a cell containingtrain_flow). However this would not be the case when using apreprocess_bijection, as the flow outputted from the first call totrain_flowwould already be transformed byInvert(preprocess_bijection), and the output from the second call would have the inverse transformation applied twice.Instead, an example is added to the FAQ for scaling/preprocessing, which is more explicit.