Training neural network is not working on GPU

[sametsoekel]

I'm having trouble with training mlp() specification by brulee. I know that brulee uses torch and I checked my torch & gpu relationship, seems okay. But in the example below, training goes on CPU.

suppressPackageStartupMessages({
    library(tidymodels)
    library(torch)
})

torch::cuda_is_available()

#> TRUE

torch::cuda_device_count()

#> [1] 1

set.seed(1)


modspec <- mlp(hidden_units = tune(),
  penalty = tune(),
  epochs = tune(),
  activation = tune(),
  learn_rate = tune()) %>%
set_mode('classification') %>%
set_engine('brulee')

fk_param <- modspec %>%
extract_parameter_set_dials %>%
grid_max_entropy(size = 50) 

spl_obj <- initial_split(iris,.7)

cv_obj <- vfold_cv(training(spl_obj),5)

rcp <- recipe(formula = Species ~
Sepal.Width +
Sepal.Length +
Petal.Width +
Petal.Length,
data = training(spl_obj)) %>%
step_normalize(all_numeric_predictors())

wf <- workflow() %>%
add_model(modspec) %>%
add_recipe(rcp)

cv_fit <- wf %>%
tune_grid(resamples = cv_obj,grid = fk_param)

[juliasilge]

Hmmm, that sounds frustrating @sametsoekel!

• Can you train a "bare" torch model like this one or this one using your GPU?
• Can you train an unwrapped (non-parsnip) brulee model like this one using your GPU?

[sametsoekel]

Hmmm, that sounds frustrating @sametsoekel!

• Can you train a "bare" torch model like this one or this one using your GPU?
• Can you train an unwrapped (non-parsnip) brulee model like this one using your GPU?

Hi @juliasilge I tried to train the first example you attached. At the beginning of training %4 of my GPU is used, about 5 minutes later I saw %15 of my GPU is being used. Shortly, I can say yes. I'm going to let you know if unwrapped brulee model still using my GPU.

Update :

I tried unwrapped brulee model you attached also, and set the epochs 200. While training, my GPU is never used.

Thanks for interest

[juliasilge]

Thank you @sametsoekel! That sounds like it is an issue with brulee then. 👍

[jonthegeek]

I suspect this is because brulee pre-dates luz. It doesn't look like brulee does anything to auto-switch between cpu and gpu. For example, there would have to be code to switch between cpu and gpu as needed in here: https://github.com/tidymodels/brulee/blob/main/R/mlp-fit.R#L635

I think this might require a pretty big refactor to implement, although you might be able to do the switching based on cuda_is_available (I've already forgotten how to do that cleanly without luz, I'm spoiled!).

[dfalbel]

@sametsoekel Thanks for filing the issue. We currently don't support GPU's in brulee, we'll add support for it soon and let you know here.

[sametsoekel]

Thank you all for the precious attention, I'll be looking forward to the feature being added.