dig.oodgraph¶

Graph OOD (GOOD) Dataset interfaces under dig.oodgraph.

Please refer to the GOOD project for more details.

This module includes 8 GOOD datasets.

Graph prediction datasets: GOOD-HIV, GOOD-PCBA, GOOD-ZINC, GOOD-CMNIST, GOOD-Motif.
Node prediction datasets: GOOD-Cora, GOOD-Arxiv, GOOD-CBAS.

`GOODArxiv`	The GOOD-Arxiv dataset adapted from OGB benchmark.
`GOODCBAS`	The GOOD-CBAS dataset.
`GOODCMNIST`	The GOOD-CMNIST dataset following IRM paper.
`GOODCora`	The GOOD-Cora dataset.
`GOODHIV`	The GOOD-HIV dataset.
`GOODMotif`	The GOOD-Motif dataset motivated by Spurious-Motif.
`GOODPCBA`	The GOOD-PCBA dataset.
`GOODZINC`	The GOOD-ZINC dataset adapted from ZINC database.

class GOODArxiv(root: str, domain: str, shift: str = 'no_shift', transform=None, pre_transform=None, generate: bool = False)[source]¶

The GOOD-Arxiv dataset adapted from OGB benchmark.

Parameters

root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘degree’ and ‘time’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

download()[source]¶: Downloads the dataset to the self.raw_dir folder.

static load(dataset_root: str, domain: str, shift: str = 'no_shift', generate: bool = False)[source]¶

A staticmethod for dataset loading. This method instantiates dataset class, constructing train, id_val, id_test, ood_val (val), and ood_test (test) splits. Besides, it collects several dataset meta information for further utilization.

Parameters

dataset_root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘degree’ and ‘time’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

Returns

dataset or dataset splits. dataset meta info.

property processed_file_names¶: The name of the files in the self.processed_dir folder that must be present in order to skip processing.

class GOODCBAS(root: str, domain: str, shift: str = 'no_shift', transform=None, pre_transform=None, generate: bool = False)[source]¶

The GOOD-CBAS dataset. Modified from BA-Shapes.

Parameters

root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘color’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

download()[source]¶: Downloads the dataset to the self.raw_dir folder.

static load(dataset_root: str, domain: str, shift: str = 'no_shift', generate: bool = False)[source]¶

A staticmethod for dataset loading. This method instantiates dataset class, constructing train, id_val, id_test, ood_val (val), and ood_test (test) splits. Besides, it collects several dataset meta information for further utilization.

Parameters

dataset_root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘degree’ and ‘time’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

Returns

dataset or dataset splits. dataset meta info.

property processed_file_names¶: The name of the files in the self.processed_dir folder that must be present in order to skip processing.

class GOODCMNIST(root: str, domain: str, shift: str = 'no_shift', subset: str = 'train', transform=None, pre_transform=None, generate: bool = False)[source]¶

The GOOD-CMNIST dataset following IRM paper.

Parameters

root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘color’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
subset (str) – The split set. Allowed: ‘train’, ‘id_val’, ‘id_test’, ‘val’, and ‘test’. When shift=’no_shift’, ‘id_val’ and ‘id_test’ are not applicable.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

download()[source]¶: Downloads the dataset to the self.raw_dir folder.

static load(dataset_root: str, domain: str, shift: str = 'no_shift', generate: bool = False)[source]¶

A staticmethod for dataset loading. This method instantiates dataset class, constructing train, id_val, id_test, ood_val (val), and ood_test (test) splits. Besides, it collects several dataset meta information for further utilization.

Parameters

dataset_root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘degree’ and ‘time’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

Returns

dataset or dataset splits. dataset meta info.

property processed_file_names¶: The name of the files in the self.processed_dir folder that must be present in order to skip processing.

class GOODCora(root: str, domain: str, shift: str = 'no_shift', transform=None, pre_transform=None, generate: bool = False)[source]¶

The GOOD-Cora dataset. Adapted from the full Cora dataset.

Parameters

root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘degree’ and ‘word’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

download()[source]¶: Downloads the dataset to the self.raw_dir folder.

static load(dataset_root: str, domain: str, shift: str = 'no_shift', generate: bool = False)[source]¶

A staticmethod for dataset loading. This method instantiates dataset class, constructing train, id_val, id_test, ood_val (val), and ood_test (test) splits. Besides, it collects several dataset meta information for further utilization.

Parameters

dataset_root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘degree’ and ‘time’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

Returns

dataset or dataset splits. dataset meta info.

property processed_file_names¶: The name of the files in the self.processed_dir folder that must be present in order to skip processing.

class GOODHIV(root: str, domain: str, shift: str = 'no_shift', subset: str = 'train', transform=None, pre_transform=None, generate: bool = False)[source]¶

The GOOD-HIV dataset. Adapted from MoleculeNet.

Parameters

root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘scaffold’ and ‘size’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
subset (str) – The split set. Allowed: ‘train’, ‘id_val’, ‘id_test’, ‘val’, and ‘test’. When shift=’no_shift’, ‘id_val’ and ‘id_test’ are not applicable.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

download()[source]¶: Downloads the dataset to the self.raw_dir folder.

static load(dataset_root: str, domain: str, shift: str = 'no_shift', generate: bool = False)[source]¶

A staticmethod for dataset loading. This method instantiates dataset class, constructing train, id_val, id_test, ood_val (val), and ood_test (test) splits. Besides, it collects several dataset meta information for further utilization.

Parameters

dataset_root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘degree’ and ‘time’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

Returns

dataset or dataset splits. dataset meta info.

property processed_file_names¶: The name of the files in the self.processed_dir folder that must be present in order to skip processing.

class GOODMotif(root: str, domain: str, shift: str = 'no_shift', subset: str = 'train', transform=None, pre_transform=None, generate: bool = False)[source]¶

The GOOD-Motif dataset motivated by Spurious-Motif.

Parameters

root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘basis’ and ‘size’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
subset (str) – The split set. Allowed: ‘train’, ‘id_val’, ‘id_test’, ‘val’, and ‘test’. When shift=’no_shift’, ‘id_val’ and ‘id_test’ are not applicable.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

download()[source]¶: Downloads the dataset to the self.raw_dir folder.

static load(dataset_root: str, domain: str, shift: str = 'no_shift', generate: bool = False)[source]¶

A staticmethod for dataset loading. This method instantiates dataset class, constructing train, id_val, id_test, ood_val (val), and ood_test (test) splits. Besides, it collects several dataset meta information for further utilization.

Parameters

dataset_root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘degree’ and ‘time’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

Returns

dataset or dataset splits. dataset meta info.

property processed_file_names¶: The name of the files in the self.processed_dir folder that must be present in order to skip processing.

class GOODPCBA(root: str, domain: str, shift: str = 'no_shift', subset: str = 'train', transform=None, pre_transform=None, generate: bool = False)[source]¶

The GOOD-PCBA dataset. Adapted from MoleculeNet.

Parameters

root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘scaffold’ and ‘size’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
subset (str) – The split set. Allowed: ‘train’, ‘id_val’, ‘id_test’, ‘val’, and ‘test’. When shift=’no_shift’, ‘id_val’ and ‘id_test’ are not applicable.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

download()[source]¶: Downloads the dataset to the self.raw_dir folder.

static load(dataset_root: str, domain: str, shift: str = 'no_shift', generate: bool = False)[source]¶

A staticmethod for dataset loading. This method instantiates dataset class, constructing train, id_val, id_test, ood_val (val), and ood_test (test) splits. Besides, it collects several dataset meta information for further utilization.

Parameters

dataset_root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘degree’ and ‘time’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

Returns

dataset or dataset splits. dataset meta info.

property processed_file_names¶: The name of the files in the self.processed_dir folder that must be present in order to skip processing.

class GOODZINC(root: str, domain: str, shift: str = 'no_shift', subset: str = 'train', transform=None, pre_transform=None, generate: bool = False)[source]¶

The GOOD-ZINC dataset adapted from ZINC database.

Parameters

root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘scaffold’ and ‘size’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
subset (str) – The split set. Allowed: ‘train’, ‘id_val’, ‘id_test’, ‘val’, and ‘test’. When shift=’no_shift’, ‘id_val’ and ‘id_test’ are not applicable.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

download()[source]¶: Downloads the dataset to the self.raw_dir folder.

static load(dataset_root: str, domain: str, shift: str = 'no_shift', generate: bool = False)[source]¶

A staticmethod for dataset loading. This method instantiates dataset class, constructing train, id_val, id_test, ood_val (val), and ood_test (test) splits. Besides, it collects several dataset meta information for further utilization.

Parameters

dataset_root (str) – The dataset saving root.
domain (str) – The domain selection. Allowed: ‘degree’ and ‘time’.
shift (str) – The distributional shift we pick. Allowed: ‘no_shift’, ‘covariate’, and ‘concept’.
generate (bool) – The flag for regenerating dataset. True: regenerate. False: download.

Returns

dataset or dataset splits. dataset meta info.

property processed_file_names¶: The name of the files in the self.processed_dir folder that must be present in order to skip processing.