R keras somewhat slower than python keras about keras HOT 4 CLOSED

Good one! On the same instance type w/ GPU I get 55 seconds for R and 40 seconds for Python. I have one pretty good theory about what might be going on (it has to do with how we convert R matrices to NumPy arrays and if I'm correct it will be straightforward to resolve). More soon!

from keras.

Cool, thanks for looking into it.

from keras.

Thanks again for reporting this, fix is here: 1d6c9f2

The problem was caused by the fact that when R matrices and arrays are marshaled to Python we convert them to NumPy arrays that map directly to the R managed memory (i.e. no copy is made and NumPy just uses the already allocated R array/matrix). This implies that the NumPy array will be Fortran-ordered (since that's the default for R) rather than C-ordered (the default for NumPy).

In the case of Keras training, between shuffling, drawing batches, copying to the GPU, and 10 epochs there are a ton of accesses to the Fortran-ordered array which add up to a ~ 40% performance penalty for training this dataset.

This change ensures that we do a single re-ordering of arrays/matrices to C-column order before we pass them to Keras, which eliminates the penalty we were paying for the accesses to the Fortran-ordered array.

@bwlewis Do you think the reticulate behavior of preserving R Fortran-ordered array memory when marshaling to NumPy is an anti-pattern? Alternatively it could be that this particular scenario of hundreds of accesses + GPU copying is problematic but that for most cases it's a performance win. Let me know what you think.

from keras.

@jjallaire Nice :)

from keras.