add defaultSettings functions for models

[egillax]

Since deep learning models can be very complicated with many hyperparameters it might be a bit overwhelming to use the current settings function.

I will implement defaultSetting functions with some default sane hyperparameters to use if the user's only want to fit one model. These hyperparameters will be either found from literature if the model's are implemented from a paper or from testing by the package authors.

• setDefaultResnet
• setDefaultMLP
• setDefaultTransformer

[ChungsooKim]

It was quite overwhelming to me also, because the default setting of the setResnet() function contains a hundred combinations of parameters.
My test results were as below.

test server : Intel Xeon CPU E5 2620 v3 (2CPUs 6Cores 24 Threads) / RAM: 492G

Cohort Size: 105580
Covariates: 7512
Outcome: 1033
data split: 75(train):25(test)
fold : 3
device: CPU
for 1 epochs it tooks 2~3 min

the default setting was 100 combinations of parameters with 30 epochs each.
that means it tooks about 2min30epochs3 folds * 100 param pairs + 2min *30 epochs * 1 training (3 times size than each fold) + 2 min * 1 test = 303+hrs (of course it can be reduced significantly with early stopping)

if we have appropriate default settings for each algorithms, it would be greate for researchers for just testing their possibility with DNN algorithms.

[egillax]

Could you share what hyperparameters were chosen in the end? I could compare with mine and the paper I got the model from. If they are all very similar we could use those as default. Some of them could depend though on the dataset, like it's size or number of covariates.

[ChungsooKim]

Sure! I'm fitting the ResNet model with 50 random parameter sets now, after it is done, I'll share it here