Using an existing task for model evaluation

rnaglib’s task module provides you with readymade dataset splits for your model evaluation in just a few lines of code.

0.) Generate necessary index files:

$ rnaglib_index

1.) Choose the task appropriate to your model. Here, we chose RNA-Site, a task instance called LigandBindindSite for illustration.

When instantiating the task, custom splitters or other arguments can be passed if needed.:

from rnaglib.tasks import BindingSiteDetection
from rnaglib.representations import GraphRepresentation

task = BindingSiteDetection(root='tutorial')
#You can pass arguments to use a custom splitter or dataset etc. if desired.

2.) Add the representation used by your model to the task object. Voxel grid or point cloud are also possible representations; here we use a graph representation in the pytorch-geometric framework.:

representation = GraphRepresentation('pyg')

task.dataset.add_representation(representation)

3.) Lastly, split your task dataset.:

train_ind, val_ind, test_ind = task.split()

train_set = task.dataset.subset(train_ind)

val_set = task.dataset.subset(val_ind)

test_set = task.dataset.subset(test_ind)

Here you go, these splits are now ready to be used by your model of choice and can for example be passed to a DataLoader. For an example of a model trainig on these splits have a look at this simple model.