03:30 <rcurtin[m]> so the good news is that I'm able to get comparable performance to PyTorch once I make sure the initializations are the same... but the bad news is that those results only hold for a batch size of 1 :)

03:30 <rcurtin[m]> so I must have some bug somewhere related to larger batch sizes

03:32 <rcurtin[m]> I'll work on getting that solved soon---once that's done, then I can move on to adapting the rest of the layers that are used by any of our examples in the `examples` repository, and then we can initially merge the ann-vtable support

03:32 <rcurtin[m]> (I'll need to make a list of the layers we need to adapt for the initial merge; it will include some of the RNN layers too I think)