I have some trouble understanding the total_size parameter and there is no mention of it in the docs. From the examples I have seen so far the parameter should be set to the total size of the training data when doing mini-batch training. This is simple to interpret if the data is just a 1d array but what should I put in there if I’m training a model on views of 2d data (subsampling in both dimensions)? Should I then use total_size=data.shape[0]*data.shape[1]?
From the gitter conversation it seems that total_size for subsampling in more then 1 dimensions needs some extra support on the pymc3 side. @ferrine suggested the following API:
total_size = int # for shape[0] subsampling
total_size = [int, None, int] # for subsampling [shape[0], shape[2]]
total_size = [int, Ellipsis, int] # for subsampling [shape[0], shape[-1]]
In my 2d case I would write total_size = [data.shape[0], data.shape[1]].