|
| 1 | +import os |
| 2 | +import random |
| 3 | +import string |
| 4 | + |
| 5 | +import pytest |
| 6 | + |
| 7 | +from torch_geometric.datasets import WebQSPDataset |
| 8 | +from torch_geometric.datasets.web_qsp_dataset import KGQABaseDataset |
| 9 | +from torch_geometric.testing import ( |
| 10 | + onlyFullTest, |
| 11 | + onlyOnline, |
| 12 | + onlyRAG, |
| 13 | + withPackage, |
| 14 | +) |
| 15 | + |
| 16 | + |
| 17 | +@pytest.mark.skip(reason="Times out") |
| 18 | +@onlyOnline |
| 19 | +@onlyFullTest |
| 20 | +def test_web_qsp_dataset(tmp_path): |
| 21 | + dataset = WebQSPDataset(root=tmp_path) |
| 22 | + # Split for this dataset is 2826 train | 246 val | 1628 test |
| 23 | + # default split is train |
| 24 | + assert len(dataset) == 2826 |
| 25 | + assert str(dataset) == "WebQSPDataset(2826)" |
| 26 | + |
| 27 | + dataset_train = WebQSPDataset(root=tmp_path, split="train") |
| 28 | + assert len(dataset_train) == 2826 |
| 29 | + assert str(dataset_train) == "WebQSPDataset(2826)" |
| 30 | + |
| 31 | + dataset_val = WebQSPDataset(root=tmp_path, split="val") |
| 32 | + assert len(dataset_val) == 246 |
| 33 | + assert str(dataset_val) == "WebQSPDataset(246)" |
| 34 | + |
| 35 | + dataset_test = WebQSPDataset(root=tmp_path, split="test") |
| 36 | + assert len(dataset_test) == 1628 |
| 37 | + assert str(dataset_test) == "WebQSPDataset(1628)" |
| 38 | + |
| 39 | + |
| 40 | +class MockSentenceTransformer: |
| 41 | + def __init__(self, *args, **kwargs): |
| 42 | + pass |
| 43 | + |
| 44 | + def to(self, device): |
| 45 | + return self |
| 46 | + |
| 47 | + def eval(self): |
| 48 | + return self |
| 49 | + |
| 50 | + def encode(self, sentences, batch_size=None, output_device=None): |
| 51 | + import torch |
| 52 | + |
| 53 | + def string_to_tensor(s: str) -> torch.Tensor: |
| 54 | + return torch.ones(1024).float() |
| 55 | + |
| 56 | + if isinstance(sentences, str): |
| 57 | + return string_to_tensor(sentences) |
| 58 | + return torch.stack([string_to_tensor(s) for s in sentences]) |
| 59 | + |
| 60 | + |
| 61 | +def create_mock_graphs(tmp_path: str, train_size: int, val_size: int, |
| 62 | + test_size: int, num_nodes: int, num_edge_types: int, |
| 63 | + num_trips: int, seed: int = 42): |
| 64 | + random.seed(seed) |
| 65 | + strkeys = string.ascii_letters + string.digits |
| 66 | + qa_strkeys = string.ascii_letters + string.digits + " " |
| 67 | + |
| 68 | + def create_mock_triplets(num_nodes: int, num_edges: int, num_trips: int): |
| 69 | + nodes = list( |
| 70 | + {"".join(random.sample(strkeys, 10)) |
| 71 | + for i in range(num_nodes)}) |
| 72 | + edges = list( |
| 73 | + {"".join(random.sample(strkeys, 10)) |
| 74 | + for i in range(num_edges)}) |
| 75 | + triplets = [] |
| 76 | + |
| 77 | + for i in range(num_trips): |
| 78 | + h = random.randint(0, num_nodes - 1) |
| 79 | + t = random.randint(0, num_nodes - 1) |
| 80 | + r = random.randint(0, num_edge_types - 1) |
| 81 | + triplets.append((nodes[h], edges[r], nodes[t])) |
| 82 | + return triplets |
| 83 | + |
| 84 | + train_triplets = [ |
| 85 | + create_mock_triplets(num_nodes, num_edge_types, num_trips) |
| 86 | + for _ in range(train_size) |
| 87 | + ] |
| 88 | + val_triplets = [ |
| 89 | + create_mock_triplets(num_nodes, num_edge_types, num_trips) |
| 90 | + for _ in range(val_size) |
| 91 | + ] |
| 92 | + test_triplets = [ |
| 93 | + create_mock_triplets(num_nodes, num_edge_types, num_trips) |
| 94 | + for _ in range(test_size) |
| 95 | + ] |
| 96 | + |
| 97 | + train_questions = [ |
| 98 | + "".join(random.sample(qa_strkeys, 10)) for _ in range(train_size) |
| 99 | + ] |
| 100 | + val_questions = [ |
| 101 | + "".join(random.sample(qa_strkeys, 10)) for _ in range(val_size) |
| 102 | + ] |
| 103 | + test_questions = [ |
| 104 | + "".join(random.sample(qa_strkeys, 10)) for _ in range(test_size) |
| 105 | + ] |
| 106 | + |
| 107 | + train_answers = [ |
| 108 | + "".join(random.sample(qa_strkeys, 10)) for _ in range(train_size) |
| 109 | + ] |
| 110 | + val_answers = [ |
| 111 | + "".join(random.sample(qa_strkeys, 10)) for _ in range(val_size) |
| 112 | + ] |
| 113 | + test_answers = [ |
| 114 | + "".join(random.sample(qa_strkeys, 10)) for _ in range(test_size) |
| 115 | + ] |
| 116 | + |
| 117 | + train_graphs = { |
| 118 | + "graph": train_triplets, |
| 119 | + "question": train_questions, |
| 120 | + "answer": train_answers |
| 121 | + } |
| 122 | + val_graphs = { |
| 123 | + "graph": val_triplets, |
| 124 | + "question": val_questions, |
| 125 | + "answer": val_answers |
| 126 | + } |
| 127 | + test_graphs = { |
| 128 | + "graph": test_triplets, |
| 129 | + "question": test_questions, |
| 130 | + "answer": test_answers |
| 131 | + } |
| 132 | + |
| 133 | + from datasets import Dataset, DatasetDict, load_from_disk |
| 134 | + |
| 135 | + ds_train = Dataset.from_dict(train_graphs, split="train") |
| 136 | + ds_val = Dataset.from_dict(val_graphs, split="validation") |
| 137 | + ds_test = Dataset.from_dict(test_graphs, split="test") |
| 138 | + |
| 139 | + ds = DatasetDict({ |
| 140 | + "train": ds_train, |
| 141 | + "validation": ds_val, |
| 142 | + "test": ds_test |
| 143 | + }) |
| 144 | + |
| 145 | + def mock_load_dataset(name: str): |
| 146 | + # Save the dataset and then load it to emulate downloading from HF |
| 147 | + DATASET_CACHE_DIR = os.path.join(tmp_path, |
| 148 | + ".cache/huggingface/datasets", name) |
| 149 | + os.makedirs(DATASET_CACHE_DIR, exist_ok=True) |
| 150 | + |
| 151 | + ds.save_to_disk(DATASET_CACHE_DIR) |
| 152 | + dataset_remote = load_from_disk(DATASET_CACHE_DIR) |
| 153 | + return dataset_remote |
| 154 | + |
| 155 | + return mock_load_dataset, ds |
| 156 | + |
| 157 | + |
| 158 | +@onlyRAG |
| 159 | +@withPackage("datasets", "pandas") |
| 160 | +def test_kgqa_base_dataset(tmp_path, monkeypatch): |
| 161 | + |
| 162 | + num_nodes = 500 |
| 163 | + num_edge_types = 25 |
| 164 | + num_trips = 5000 |
| 165 | + |
| 166 | + # Mock the dataset graphs |
| 167 | + mock_load_dataset_func, expected_result = create_mock_graphs( |
| 168 | + tmp_path, train_size=10, val_size=5, test_size=5, num_nodes=num_nodes, |
| 169 | + num_edge_types=num_edge_types, num_trips=num_trips) |
| 170 | + |
| 171 | + import datasets |
| 172 | + |
| 173 | + monkeypatch.setattr(datasets, "load_dataset", mock_load_dataset_func) |
| 174 | + |
| 175 | + # Mock the SentenceTransformer |
| 176 | + import torch_geometric.datasets.web_qsp_dataset |
| 177 | + monkeypatch.setattr(torch_geometric.datasets.web_qsp_dataset, |
| 178 | + "SentenceTransformer", MockSentenceTransformer) |
| 179 | + |
| 180 | + dataset_train = KGQABaseDataset(root=tmp_path, dataset_name="TestDataset", |
| 181 | + split="train", use_pcst=False) |
| 182 | + assert len(dataset_train) == 10 |
| 183 | + assert str(dataset_train) == "KGQABaseDataset(10)" |
| 184 | + for graph in dataset_train: |
| 185 | + assert graph.x.shape == (num_nodes, 1024) |
| 186 | + assert graph.edge_index.shape == (2, num_trips) |
| 187 | + assert graph.edge_attr.shape == ( |
| 188 | + num_trips, 1024) # Reminder: edge_attr encodes the entire triplet |
| 189 | + |
| 190 | + dataset_val = KGQABaseDataset(root=tmp_path, dataset_name="TestDataset", |
| 191 | + split="val", use_pcst=False) |
| 192 | + assert len(dataset_val) == 5 |
| 193 | + assert str(dataset_val) == "KGQABaseDataset(5)" |
| 194 | + |
| 195 | + dataset_test = KGQABaseDataset(root=tmp_path, dataset_name="TestDataset", |
| 196 | + split="test", use_pcst=False) |
| 197 | + assert len(dataset_test) == 5 |
| 198 | + assert str(dataset_test) == "KGQABaseDataset(5)" |
| 199 | + |
| 200 | + # TODO(zaristei): More rigorous tests to validate that values are correct |
| 201 | + # TODO(zaristei): Proper tests for PCST and CWQ |
0 commit comments