[jorgemf]

well, the data is way smaller than the model (at least the current trend), and you probably still need random access for the weights of the model. I am not sure if it a gain worth to pursue.