gradient clipping for numerical stability

[topepo]

Closes #73

There are occasions where the model parameters lead to overflow when computing the loss function:

library(brulee)
library(modeldata)
#> 
#> Attaching package: 'modeldata'
#> The following object is masked from 'package:datasets':
#> 
#>     penguins

i <- 25691
set.seed(i)
data_tr <- sim_logistic(200, ~ .1 + 2 * A - 3 * B + 1 * A *B, corr = .7)

set.seed(i+1)
mlp_fit <- brulee_mlp(class ~ ., data = data_tr, hidden_units = 10, epochs = 35,
                      stop_iter = Inf, verbose = TRUE)
#> epoch: 1 learn rate 0.01 Loss: 0.581
#> epoch: 2 learn rate 0.01 Loss: 0.492
#> epoch: 3 learn rate 0.01 Loss: 0.482
#> epoch: 4 learn rate 0.01 Loss: 0.474
#> epoch: 5 learn rate 0.01 Loss: 0.47
#> epoch: 6 learn rate 0.01 Loss: 0.465
#> epoch: 7 learn rate 0.01 Loss: 0.466 ✖
#> epoch: 8 learn rate 0.01 Loss: 0.461
#> epoch: 9 learn rate 0.01 Loss: 0.45
#> epoch: 10 learn rate 0.01 Loss: 0.45
#> epoch: 11 learn rate 0.01 Loss: 0.45 ✖
#> epoch: 12 learn rate 0.01 Loss: 0.45 ✖
#> epoch: 13 learn rate 0.01 Loss: 0.45 ✖
#> epoch: 14 learn rate 0.01 Loss: 0.45 ✖
#> epoch: 15 learn rate 0.01 Loss: 0.45 ✖
#> epoch: 16 learn rate 0.01 Loss: 0.45 ✖
#> epoch: 17 learn rate 0.01 Loss: 0.45 ✖
#> epoch: 18 learn rate 0.01 Loss: 0.45 ✖
#> epoch: 19 learn rate 0.01 Loss: 0.45 ✖
#> epoch: 20 learn rate 0.01 Loss: 0.45 ✖
#> epoch: 21 learn rate 0.01 Loss: 0.45 ✖
#> epoch: 22 learn rate 0.01 Loss: 0.45 ✖
#> epoch: 23 learn rate 0.01 Loss: 0.449
#> epoch: 24 learn rate 0.01 Loss: 0.45 ✖
#> epoch: 25 learn rate 0.01 Loss: 0.448
#> epoch: 26 learn rate 0.01 Loss: 0.448 ✖
#> epoch: 27 learn rate 0.01 Loss: 0.449 ✖
#> epoch: 28 learn rate 0.01 Loss: 0.448
#> epoch: 29 learn rate 0.01 Loss: 0.449 ✖
#> epoch: 30 learn rate 0.01 Loss: 0.448 ✖
#> epoch: 31 learn rate 0.01 Loss: 0.448
#> Warning: Loss is NaN at epoch 32. Training is stopped.

^{Created on 2025-07-27 with reprex v2.1.1}

This PR adds two new parameters that can prevent the gradient values or their norms from becoming too large:

library(brulee)
library(modeldata)
#> 
#> Attaching package: 'modeldata'
#> The following object is masked from 'package:datasets':
#> 
#>     penguins

i <- 25691
set.seed(i)
data_tr <- sim_logistic(200, ~ .1 + 2 * A - 3 * B + 1 * A *B, corr = .7)

set.seed(i+1)
mlp_fit <- brulee_mlp(class ~ ., data = data_tr, hidden_units = 10, epochs = 35, 
                      grad_norm_clip = 1 / 50, stop_iter = Inf, verbose = TRUE)
#> epoch: 1 learn rate 0.01 Loss: 0.783
#> epoch: 2 learn rate 0.01 Loss: 0.78
#> epoch: 3 learn rate 0.01 Loss: 0.777
#> epoch: 4 learn rate 0.01 Loss: 0.775
#> epoch: 5 learn rate 0.01 Loss: 0.772
#> epoch: 6 learn rate 0.01 Loss: 0.769
#> epoch: 7 learn rate 0.01 Loss: 0.767
#> epoch: 8 learn rate 0.01 Loss: 0.764
#> epoch: 9 learn rate 0.01 Loss: 0.761
#> epoch: 10 learn rate 0.01 Loss: 0.759
#> epoch: 11 learn rate 0.01 Loss: 0.756
#> epoch: 12 learn rate 0.01 Loss: 0.753
#> epoch: 13 learn rate 0.01 Loss: 0.751
#> epoch: 14 learn rate 0.01 Loss: 0.748
#> epoch: 15 learn rate 0.01 Loss: 0.746
#> epoch: 16 learn rate 0.01 Loss: 0.743
#> epoch: 17 learn rate 0.01 Loss: 0.74
#> epoch: 18 learn rate 0.01 Loss: 0.738
#> epoch: 19 learn rate 0.01 Loss: 0.735
#> epoch: 20 learn rate 0.01 Loss: 0.733
#> epoch: 21 learn rate 0.01 Loss: 0.73
#> epoch: 22 learn rate 0.01 Loss: 0.728
#> epoch: 23 learn rate 0.01 Loss: 0.725
#> epoch: 24 learn rate 0.01 Loss: 0.723
#> epoch: 25 learn rate 0.01 Loss: 0.721
#> epoch: 26 learn rate 0.01 Loss: 0.718
#> epoch: 27 learn rate 0.01 Loss: 0.716
#> epoch: 28 learn rate 0.01 Loss: 0.713
#> epoch: 29 learn rate 0.01 Loss: 0.711
#> epoch: 30 learn rate 0.01 Loss: 0.709
#> epoch: 31 learn rate 0.01 Loss: 0.706
#> epoch: 32 learn rate 0.01 Loss: 0.704
#> epoch: 33 learn rate 0.01 Loss: 0.702
#> epoch: 34 learn rate 0.01 Loss: 0.7
#> epoch: 35 learn rate 0.01 Loss: 0.697

^{Created on 2025-07-27 with reprex v2.1.1}

[dfalbel]

LGTM!