GPU support

[topepo]

No description provided.

[topepo]

I'm getting messages that say

[W MPSFallback.mm:11] Warning: The operator 'aten::sgn.out' is not currently supported on the MPS backend and will fall back to run on the CPU. This may have performance implications. (function operator())

I'm doing some initial testing and the gpu is taking 2x-4x more time than the cpu version

tibble::tribble(
 ~seed,    ~n, ~extra_cols, ~hidden_units, ~epochs,   ~cpu,    ~gpu, ~gpu_to_cpu,
   836, 1e+06,         100,            10,     100,  85.83,  224.07,        2.61,
   836, 1e+06,         100,            10,    1000, 489.62,  934.22,        1.91,
   836, 1e+06,         100,            10,     500, 299.11,  582.33,        1.95,
   836, 1e+06,         100,           100,     100, 441.41, 1977.09,        4.48
 )

Here is a reprex. I'm on an m1 max macstudio (Mac13,1)

library(tidymodels)
library(brulee) # pak::pak(c("tidymodels/brulee#76"), ask = FALSE)

# ------------------------------------------------------------------------------

tidymodels_prefer()
theme_set(theme_bw())
options(pillar.advice = FALSE, pillar.min_title_chars = Inf)

# ------------------------------------------------------------------------------

set.seed(836)
sim_tr <-
 sim_regression(1e+06) %>%
 bind_cols(sim_noise(1e+06, num_vars = 100))

# ------------------------------------------------------------------------------

cpu_time <-
 system.time({
  cpu_fit <-
   brulee_mlp(outcome ~ ., data = sim_tr,
              hidden_units = 10, epochs = 100, stop_iter = 100)
 })

# ------------------------------------------------------------------------------

gpu_time <-
 system.time({
  gpu_fit <-
   brulee_mlp(outcome ~ ., data = sim_tr,
              hidden_units = 10, epochs = 100, stop_iter = 100,
              device = "mps")
 })