Source code for torch_geometric.datasets.ogb_mag

from typing import Optional, Callable, List

import os
import shutil
import os.path as osp

import torch
import numpy as np

from torch_geometric.data import (InMemoryDataset, HeteroData, download_url,
                                  extract_zip)


[docs]class OGB_MAG(InMemoryDataset): r"""The ogbn-mag dataset from the `"Open Graph Benchmark: Datasets for Machine Learning on Graphs" <https://arxiv.org/abs/2005.00687>`_ paper. ogbn-mag is a heterogeneous graph composed of a subset of the Microsoft Academic Graph (MAG). It contains four types of entities — papers (736,389 nodes), authors (1,134,649 nodes), institutions (8,740 nodes), and fields of study (59,965 nodes) — as well as four types of directed relations connecting two types of entities. Each paper is associated with a 128-dimensional :obj:`word2vec` feature vector, while all other node types are not associated with any input features. The task is to predict the venue (conference or journal) of each paper. In total, there are 349 different venues. Args: root (string): Root directory where the dataset should be saved. preprocess (string, optional): Pre-processes the original dataset by adding structural features (:obj:`"metapath2vec", :obj:`"TransE") to featureless nodes. (default: :obj:`None`) transform (callable, optional): A function/transform that takes in an :obj:`torch_geometric.data.HeteroData` object and returns a transformed version. The data object will be transformed before every access. (default: :obj:`None`) pre_transform (callable, optional): A function/transform that takes in an :obj:`torch_geometric.data.HeteroData` object and returns a transformed version. The data object will be transformed before being saved to disk. (default: :obj:`None`) """ url = 'http://snap.stanford.edu/ogb/data/nodeproppred/mag.zip' urls = { 'metapath2vec': ('https://data.pyg.org/datasets/' 'mag_metapath2vec_emb.zip'), 'transe': ('https://data.pyg.org/datasets/' 'mag_transe_emb.zip'), } def __init__(self, root: str, preprocess: Optional[str] = None, transform: Optional[Callable] = None, pre_transform: Optional[Callable] = None): preprocess = None if preprocess is None else preprocess.lower() self.preprocess = preprocess assert self.preprocess in [None, 'metapath2vec', 'transe'] super().__init__(root, transform, pre_transform) self.data, self.slices = torch.load(self.processed_paths[0]) @property def num_classes(self) -> int: return int(self.data['paper'].y.max()) + 1 @property def raw_dir(self) -> str: return osp.join(self.root, 'mag', 'raw') @property def processed_dir(self) -> str: return osp.join(self.root, 'mag', 'processed') @property def raw_file_names(self) -> List[str]: file_names = [ 'node-feat', 'node-label', 'relations', 'split', 'num-node-dict.csv.gz' ] if self.preprocess is not None: file_names += [f'mag_{self.preprocess}_emb.pt'] return file_names @property def processed_file_names(self) -> str: if self.preprocess is not None: return f'data_{self.preprocess}.pt' else: return 'data.pt' def download(self): if not all([osp.exists(f) for f in self.raw_paths[:5]]): path = download_url(self.url, self.raw_dir) extract_zip(path, self.raw_dir) for file_name in ['node-feat', 'node-label', 'relations']: path = osp.join(self.raw_dir, 'mag', 'raw', file_name) shutil.move(path, self.raw_dir) path = osp.join(self.raw_dir, 'mag', 'split') shutil.move(path, self.raw_dir) path = osp.join(self.raw_dir, 'mag', 'raw', 'num-node-dict.csv.gz') shutil.move(path, self.raw_dir) shutil.rmtree(osp.join(self.raw_dir, 'mag')) os.remove(osp.join(self.raw_dir, 'mag.zip')) if self.preprocess is not None: path = download_url(self.urls[self.preprocess], self.raw_dir) extract_zip(path, self.raw_dir) os.remove(path) def process(self): import pandas as pd data = HeteroData() path = osp.join(self.raw_dir, 'node-feat', 'paper', 'node-feat.csv.gz') x_paper = pd.read_csv(path, compression='gzip', header=None, dtype=np.float32).values data['paper'].x = torch.from_numpy(x_paper) path = osp.join(self.raw_dir, 'node-feat', 'paper', 'node_year.csv.gz') year_paper = pd.read_csv(path, compression='gzip', header=None, dtype=np.int64).values data['paper'].year = torch.from_numpy(year_paper).view(-1) path = osp.join(self.raw_dir, 'node-label', 'paper', 'node-label.csv.gz') y_paper = pd.read_csv(path, compression='gzip', header=None, dtype=np.int64).values.flatten() data['paper'].y = torch.from_numpy(y_paper) if self.preprocess is None: path = osp.join(self.raw_dir, 'num-node-dict.csv.gz') num_nodes_df = pd.read_csv(path, compression='gzip') for node_type in ['author', 'institution', 'field_of_study']: data[node_type].num_nodes = num_nodes_df[node_type].tolist()[0] else: emb_dict = torch.load(self.raw_paths[-1]) for key, value in emb_dict.items(): if key != 'paper': data[key].x = value for edge_type in [('author', 'affiliated_with', 'institution'), ('author', 'writes', 'paper'), ('paper', 'cites', 'paper'), ('paper', 'has_topic', 'field_of_study')]: f = '___'.join(edge_type) path = osp.join(self.raw_dir, 'relations', f, 'edge.csv.gz') edge_index = pd.read_csv(path, compression='gzip', header=None, dtype=np.int64).values edge_index = torch.from_numpy(edge_index).t().contiguous() data[edge_type].edge_index = edge_index for f, v in [('train', 'train'), ('valid', 'val'), ('test', 'test')]: path = osp.join(self.raw_dir, 'split', 'time', 'paper', f'{f}.csv.gz') idx = pd.read_csv(path, compression='gzip', header=None, dtype=np.int64).values.flatten() idx = torch.from_numpy(idx) mask = torch.zeros(data['paper'].num_nodes, dtype=torch.bool) mask[idx] = True data['paper'][f'{v}_mask'] = mask if self.pre_transform is not None: data = self.pre_transform(data) torch.save(self.collate([data]), self.processed_paths[0]) def __repr__(self) -> str: return 'ogbn-mag()'