-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdataloader.py
223 lines (177 loc) · 8.99 KB
/
dataloader.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
import pickle
import sqlite3
from pathlib import Path
# import requests
# import zstandard
from ogb.nodeproppred import NodePropPredDataset
from torch_geometric.data import Data
from torch_geometric.utils import subgraph, to_undirected
from torch_geometric.typing import SparseTensor
import itertools
import torch
from torch.utils.data import DataLoader
import numpy as np
np.random.seed(0)
import torch_cluster # noqa
random_walk = torch.ops.torch_cluster.random_walk
OUT_OF_SAMPLE_SIZE = 100
BATCH_SIZE = 128
def convert_graph_to_data(graph):
return Data(x=torch.from_numpy(graph["node_feat"]), edge_index=graph["edge_index"])
# Gets full dataset graph with all nodes and edges (for final out-of-sample testing)
def get_full_graph():
dataset = NodePropPredDataset(name="ogbn-arxiv", root='dataset/')
return dataset[0]
def get_val_dataloader(graph, model_type):
return get_dataloader(graph, 1, 2, model_type)
def get_dataloader(graph, walk_length, context_size, model_type):
row, col = graph.edge_index
n = graph.num_nodes
edge_adj_matrix = SparseTensor(row=row, col=col, sparse_sizes=(n, n))
if model_type == 'attri2vec':
sample_func = lambda x: sample(x, edge_adj_matrix, graph.x, walk_length, context_size)
elif model_type == 'gnn':
sample_func = lambda x: sample_idxs(x, edge_adj_matrix, graph.x, walk_length, context_size)
else:
raise ValueError('invalid model_type')
return DataLoader(range(edge_adj_matrix.sparse_size(0)),
collate_fn=sample_func, batch_size=BATCH_SIZE)
def pos_walk_sample(batch, edge_adj_mat, walk_length, context_size, walks_per_node=10, p=1, q=1):
batch = batch.repeat(walks_per_node)
rowptr, col, _ = edge_adj_mat.csr()
rw = random_walk(rowptr, col, batch, walk_length, p, q)
if not isinstance(rw, torch.Tensor):
rw = rw[0]
walks = []
num_walks_per_rw = 1 + walk_length + 1 - context_size
for j in range(num_walks_per_rw):
walks.append(rw[:, j:j + context_size])
return torch.cat(walks, dim=0)
def neg_walk_sample(batch, edge_adj_mat, walk_length, context_size, walks_per_node=10, num_negative_samples=1):
batch = batch.repeat(walks_per_node * num_negative_samples)
rw = torch.randint(edge_adj_mat.sparse_size(0),
(batch.size(0), walk_length))
rw = torch.cat([batch.view(-1, 1), rw], dim=-1)
walks = []
num_walks_per_rw = 1 + walk_length + 1 - context_size
for j in range(num_walks_per_rw):
walks.append(rw[:, j:j + context_size])
return torch.cat(walks, dim=0)
def sample(batch, edge_adj_mat, node_feats, walk_length=20, context_size=10):
batch = torch.tensor(batch)
pos_walks_idx = pos_walk_sample(batch, edge_adj_mat, walk_length, context_size)
neg_walks_idx = neg_walk_sample(batch, edge_adj_mat, walk_length, context_size)
d1, d2 = pos_walks_idx.shape
pos_walks_with_emb = torch.index_select(node_feats, 0, pos_walks_idx.flatten()).reshape((d1, d2, -1))
neg_walks_with_emb = torch.index_select(node_feats, 0, neg_walks_idx.flatten()).reshape((d1, d2, -1))
return pos_walks_with_emb, neg_walks_with_emb
def sample_idxs(batch, edge_adj_mat, node_feats, walk_length=20, context_size=10):
batch = torch.tensor(batch)
pos_walks_idx = pos_walk_sample(batch, edge_adj_mat, walk_length, context_size)
neg_walks_idx = neg_walk_sample(batch, edge_adj_mat, walk_length, context_size)
return pos_walks_idx, neg_walks_idx
# Returns train_dataset, val_dataset, (test_graph, full_graph)
# - in-sample data: train and validation Data graph objects for in-sample training
# - test data: all nodes but only train/val edges
def get_datasets():
dataset = NodePropPredDataset(name="ogbn-arxiv", root='dataset/')
graph, label = dataset[0]
node_feat, edge_idx = torch.from_numpy(graph["node_feat"]), torch.from_numpy(graph["edge_index"])
# Node index splits
split_idx = dataset.get_idx_split()
train_idx, val_idx, test_idx = split_idx["train"], split_idx["valid"], split_idx["test"]
# np.random.shuffle(test_idx)
# in_sample_test_idx, out_of_sample_test_idx = torch.from_numpy(test_idx[OUT_OF_SAMPLE_SIZE:]), torch.from_numpy(test_idx[:OUT_OF_SAMPLE_SIZE])
# In-sample training/validation datasets
# Note: confirmed that relabel nodes retains node feature order
train_edges = subgraph(torch.from_numpy(train_idx), edge_idx, relabel_nodes=True)[0]
train_graph = Data(x=node_feat[train_idx], edge_index=to_undirected(train_edges))
# train_and_val_idx = np.sort(np.concatenate((train_idx, valid_idx)))
val_edges = subgraph(torch.from_numpy(val_idx), edge_idx, relabel_nodes=True)[0]
val_graph = Data(x=node_feat[val_idx], edge_index=to_undirected(val_edges))
# Creates test graph with all train/val edges and separate test nodes
train_and_val_idx = torch.from_numpy(np.sort(np.concatenate((train_idx, val_idx))))
train_and_val_edges = clean_edges(edge_idx, train_and_val_idx, True)
# train_val_test_idx = np.sort(np.concatenate((train_and_val_idx, test_idx)))
# test_edges = subgraph(torch.arange(graph["num_nodes"]), train_and_val_edges, relabel_nodes=False)[0]
test_graph = Data(x=node_feat, edge_index=to_undirected(train_and_val_edges))
return train_graph, val_graph, graph #(test_graph, graph, test_idx)
def get_title(id) -> str:
url = "https://zenodo.org/record/6511057/files/Papers.txt.zst?download=1"
data_filepath = Path("dataset") / "papers.txt"
db_filepath = Path("dataset") / "papers.db"
if not db_filepath.exists():
raise FileNotFoundError("Papers database not found")
# if not data_filepath.exists():
# with requests.get(url, stream=True) as r:
# r.raise_for_status()
# with open(data_filepath, "wb") as f:
# dctx = zstandard.ZstdDecompressor()
# dctx.copy_stream(r.raw, f)
con = sqlite3.connect(db_filepath)
cur = con.cursor()
res = cur.execute("SELECT PaperTitle from papers WHERE PaperId = ? LIMIT 1", (id,))
return res.fetchone()[0]
# Returns all edges where either one (not in_sample) or both (in_sample) nodes are within the split idx
def clean_edges(full_edge_idx, split_idxs, in_sample):
edge_elem_mask = torch.isin(full_edge_idx, split_idxs)
edge_mask = torch.sum(edge_elem_mask, axis=0) > in_sample
split_edge_idxs = full_edge_idx[:, edge_mask]
return split_edge_idxs
# Returns a tuple of edges and edge labels where 1 is existing and 0 is not in graph
# TODO: do we want to shuffle?
def get_pos_neg_edges(graph, in_sample):
dataset = NodePropPredDataset(name="ogbn-arxiv", root='dataset/')
split_idx = dataset.get_idx_split()
train_idx, val_idx, test_idx = split_idx["train"], split_idx["valid"], split_idx["test"]
if in_sample:
idxs = torch.from_numpy(np.concatenate((train_idx, val_idx)))
all_edges = torch.combinations(idxs)
else:
idxs = torch.from_numpy(test_idx)
all_edges = torch.combinations(torch.arange(graph.num_nodes))
pos_out_sample_edges = clean_edges(graph.edge_index, idxs, in_sample)
# TODO: check dimensions
labels = torch.zeros(all_edges.shape[0])
labels[all_edges == pos_out_sample_edges] = 1
return all_edges, labels
def test_relabel_node():
dataset = NodePropPredDataset(name="ogbn-arxiv", root='dataset/')
graph, label = dataset[0]
node_feat, edge_idx = torch.from_numpy(graph["node_feat"]), torch.from_numpy(graph["edge_index"])
split_idx = dataset.get_idx_split()
train_idx = torch.from_numpy(split_idx["train"])
train_edges = subgraph(train_idx, edge_idx, relabel_nodes=True)[0]
train_graph = Data(x=node_feat[train_idx], edge_index=train_edges)
random_shuffle_idx = train_idx.clone().numpy()
np.random.shuffle(random_shuffle_idx)
random_shuffle_idx = torch.from_numpy(random_shuffle_idx)
train_shuffle_edges = subgraph(random_shuffle_idx, edge_idx, relabel_nodes=True)[0]
# Node features need to be in order. Subgraph node indices can be shuffled.
train_shuffle_graph = Data(x=node_feat[train_idx], edge_index=train_shuffle_edges)
assert(torch.all(train_graph.edge_index == train_shuffle_graph.edge_index).item())
assert(torch.all(train_graph.x == train_shuffle_graph.x).item())
def test_randomwalk(graph, walk_length, context_size):
edges = []
for i in range(499):
edges.append([i, i+1])
for i in range(0, 498, 2):
edges.append([i, i + 2])
# graph = Data(x=torch.arange(500), edge_index=torch.tensor(edges).T)
dataloader = get_dataloader(graph, walk_length, context_size)
sample_data = []
for _ in range(10):
data = next(iter(dataloader))
sample_data.append(data)
print(sample_data[0][0].shape)
print(sample_data[0][1].shape)
if __name__ == "__main__":
train, val, graph = get_datasets()
print("train", train)
print("train", train["x"])
print("train", train.edge_index)
test_relabel_node()
print("train")
test_randomwalk(train, 20, 10)
print("val")
test_randomwalk(val, 1, 2)