Adam for c2st_mlp_prob(), preferring torch's C++ implementation

optim_ignite_adam() takes the whole step in C++ where optim_adam() loops over the parameters in R. The two produce identical trajectories, and this loop is thousands of steps over six small tensors, so the R-side overhead is most of what it costs: swapping one for the other halves the running time. optim_ignite_adam() arrived in torch 0.13.0 and DESCRIPTION allows 0.11.0, which is what the fallback is for.

Usage

c2st_adam(groups, lr)

Arguments

groups

Parameter groups, as optim_adam() takes them.

lr

Adam step size.