Currently ProbFlow uses just a single MC sample from variational posteriors per batch. Fitting will be much more stable if we can use more. In fact I'm pretty sure it's impossible to use mixture distributions as variational posteriors with just 1 MC sample...?
This'll require some expand_dims-ing of the input tensors/numpy arrays/pandas d... Hmm won't work with pandas dataframes 🤔 .
Also means user slicing code in __call__ methods could cause problems.... Maybe just have the default be 1 and tell users to handle it if they want >1 (but handle it in applications and modules like Dense / DenseNetwork).
[ ] add ability to use >1 MC sample ber batch (n_mc_samples kwarg to Model.fit?)
[ ] tests
[ ] update applications and modules to be compatible w/ >1 MC sample/batch
Currently ProbFlow uses just a single MC sample from variational posteriors per batch. Fitting will be much more stable if we can use more. In fact I'm pretty sure it's impossible to use mixture distributions as variational posteriors with just 1 MC sample...?
This'll require some
expand_dims-ing of the input tensors/numpy arrays/pandas d... Hmm won't work with pandas dataframes 🤔 .Also means user slicing code in
__call__methods could cause problems.... Maybe just have the default be 1 and tell users to handle it if they want >1 (but handle it in applications and modules like Dense / DenseNetwork).n_mc_sampleskwarg toModel.fit?)