The following files are involved with datasets.
driver/dataset.py defines a class FastDataset, where the loading of a dataset is handled by the class method from_path(). Typically, for an OGB (or PyG) dataset, it is first loaded by the corresponding API. Then, certain processing is done, such as converting the node feature matrix to row-major order and half-precision, and symmetrizing the graph adjacency matrix. To expedite the loading of the dataset in subsequent uses, we store pertinent data in .pt files so that subsequent uses do not need to perform the transformation again.
driver/main.py uses the class FastDataset in the function get_dataset().