Access edge data in Pytorch Geometric¶

This tutorial will show you how to access various datasets and their corresponding edgelists in tgb

The logic for PyG data is stored in dataset_pyg.py in tgb/linkproppred and tgb/nodeproppred folders

This tutorial requires Pytorch and PyG, refer to README.md for installation instructions

In [1]:

from tgb.linkproppred.dataset_pyg import PyGLinkPropPredDataset

from tgb.linkproppred.dataset_pyg import PyGLinkPropPredDataset

specifying the name of the dataset

In [2]:

name = "tgbl-wiki"

name = "tgbl-wiki"

Process and load the dataset¶

if the dataset has been processed, it will be loaded from disc for fast access

if the dataset has not been downloaded, it will be processed automatically

In [3]:

dataset = PyGLinkPropPredDataset(name=name, root="datasets")
type(dataset)

dataset = PyGLinkPropPredDataset(name=name, root="datasets") type(dataset)

file found, skipping download
Dataset directory is  /mnt/f/code/TGB/tgb/datasets/tgbl_wiki
loading processed file

Out[3]:

tgb.linkproppred.dataset_pyg.PyGLinkPropPredDataset

Access edge data from TemporalData object¶

You can retrieve torch_geometric.data.temporal.TemporalData directly from PyGLinkPropPredDataset

In [4]:

data = dataset.get_TemporalData()
type(data)

data = dataset.get_TemporalData() type(data)

Out[4]:

torch_geometric.data.temporal.TemporalData

In [5]:

type(data.src)
type(data.dst)
type(data.t)
type(data.msg)

type(data.src) type(data.dst) type(data.t) type(data.msg)

Out[5]:

torch.Tensor

Directly access edge data as Pytorch tensors¶

the edge data can be easily accessed via the property of the method, these are converted into pytorch tensors (from PyGLinkPropPredDataset)

In [6]:

type(dataset.src)  #same as src from above
type(dataset.dst)  #same as dst
type(dataset.ts)  #same as t
type(dataset.edge_feat) #same as msg
type(dataset.edge_label) #same as label used in tgn

type(dataset.src) #same as src from above type(dataset.dst) #same as dst type(dataset.ts) #same as t type(dataset.edge_feat) #same as msg type(dataset.edge_label) #same as label used in tgn

Out[6]:

torch.Tensor

Accessing the train, test, val split¶

the masks for training, validation, and test split can be accessed directly from the dataset as well

In [7]:

train_mask = dataset.train_mask
val_mask = dataset.val_mask
test_mask = dataset.test_mask

type(train_mask)
type(val_mask)
type(test_mask)

train_mask = dataset.train_mask val_mask = dataset.val_mask test_mask = dataset.test_mask type(train_mask) type(val_mask) type(test_mask)

Out[7]:

torch.Tensor

In [ ]: