Model Training
Model training is implemented in the train_frm function.
It computes chemical descriptors for all molecules, trains a final model
on the full dataset and estimates its performance by cross-validation.
Hyperparameters such as the Lasso penalty or the number of XGBoost
boosting rounds are chosen by internal cross-validation; for XGBoost, a
grid search over further hyperparameters can be enabled via argument
tune_grid. To avoid data leakage, descriptor filtering and
the generation of polynomial and interaction terms are repeated inside
each cross-validation fold. For further details see the function
documentation of train_frm.
Reactive Graph
FastRet is a reactive web app, that means most of its calculations are triggered as reaction to user inputs.
Note that the figures below predate the GUI rework of FastRet 1.4.0 and have not been updated since, so some node names and connections may differ from the current implementation.
The reactive graphs for FastRet’s four modes are shown below, with
- Action buttons shown as red rectangles
- Upload buttons shown as orange rectangles
- Other inputs shown as green rectangles
- Intermediate calculations shown as violet rectangles
- Outputs shown as blue rectangles
- Synchronous functions shown as white rectangles with two lines on the left and right side
- Asynchronous functions shown as hexagons
- Red arrows from one node to another indicating that the first node triggers the second node
- Green arrows from one node to another indicating that the first node holds an input value for the second node
- Blue arrows from one node to another indicating that the first node writes an output value to the second node
