-
Notifications
You must be signed in to change notification settings - Fork 66
/
tools.py
241 lines (214 loc) · 11.1 KB
/
tools.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
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
import torch
import dgl
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.metrics import f1_score, normalized_mutual_info_score, adjusted_rand_score
from sklearn.cluster import KMeans
from sklearn.svm import LinearSVC
def idx_to_one_hot(idx_arr):
one_hot = np.zeros((idx_arr.shape[0], idx_arr.max() + 1))
one_hot[np.arange(idx_arr.shape[0]), idx_arr] = 1
return one_hot
def kmeans_test(X, y, n_clusters, repeat=10):
nmi_list = []
ari_list = []
for _ in range(repeat):
kmeans = KMeans(n_clusters=n_clusters)
y_pred = kmeans.fit_predict(X)
nmi_score = normalized_mutual_info_score(y, y_pred, average_method='arithmetic')
ari_score = adjusted_rand_score(y, y_pred)
nmi_list.append(nmi_score)
ari_list.append(ari_score)
return np.mean(nmi_list), np.std(nmi_list), np.mean(ari_list), np.std(ari_list)
def svm_test(X, y, test_sizes=(0.2, 0.4, 0.6, 0.8), repeat=10):
random_states = [182318 + i for i in range(repeat)]
result_macro_f1_list = []
result_micro_f1_list = []
for test_size in test_sizes:
macro_f1_list = []
micro_f1_list = []
for i in range(repeat):
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=test_size, shuffle=True, random_state=random_states[i])
svm = LinearSVC(dual=False)
svm.fit(X_train, y_train)
y_pred = svm.predict(X_test)
macro_f1 = f1_score(y_test, y_pred, average='macro')
micro_f1 = f1_score(y_test, y_pred, average='micro')
macro_f1_list.append(macro_f1)
micro_f1_list.append(micro_f1)
result_macro_f1_list.append((np.mean(macro_f1_list), np.std(macro_f1_list)))
result_micro_f1_list.append((np.mean(micro_f1_list), np.std(micro_f1_list)))
return result_macro_f1_list, result_micro_f1_list
def evaluate_results_nc(embeddings, labels, num_classes):
print('SVM test')
svm_macro_f1_list, svm_micro_f1_list = svm_test(embeddings, labels)
print('Macro-F1: ' + ', '.join(['{:.6f}~{:.6f} ({:.1f})'.format(macro_f1_mean, macro_f1_std, train_size) for
(macro_f1_mean, macro_f1_std), train_size in
zip(svm_macro_f1_list, [0.8, 0.6, 0.4, 0.2])]))
print('Micro-F1: ' + ', '.join(['{:.6f}~{:.6f} ({:.1f})'.format(micro_f1_mean, micro_f1_std, train_size) for
(micro_f1_mean, micro_f1_std), train_size in
zip(svm_micro_f1_list, [0.8, 0.6, 0.4, 0.2])]))
print('K-means test')
nmi_mean, nmi_std, ari_mean, ari_std = kmeans_test(embeddings, labels, num_classes)
print('NMI: {:.6f}~{:.6f}'.format(nmi_mean, nmi_std))
print('ARI: {:.6f}~{:.6f}'.format(ari_mean, ari_std))
return svm_macro_f1_list, svm_micro_f1_list, nmi_mean, nmi_std, ari_mean, ari_std
def parse_adjlist(adjlist, edge_metapath_indices, samples=None):
edges = []
nodes = set()
result_indices = []
for row, indices in zip(adjlist, edge_metapath_indices):
row_parsed = list(map(int, row.split(' ')))
nodes.add(row_parsed[0])
if len(row_parsed) > 1:
# sampling neighbors
if samples is None:
neighbors = row_parsed[1:]
result_indices.append(indices)
else:
# undersampling frequent neighbors
unique, counts = np.unique(row_parsed[1:], return_counts=True)
p = []
for count in counts:
p += [(count ** (3 / 4)) / count] * count
p = np.array(p)
p = p / p.sum()
samples = min(samples, len(row_parsed) - 1)
sampled_idx = np.sort(np.random.choice(len(row_parsed) - 1, samples, replace=False, p=p))
neighbors = [row_parsed[i + 1] for i in sampled_idx]
result_indices.append(indices[sampled_idx])
else:
neighbors = []
result_indices.append(indices)
for dst in neighbors:
nodes.add(dst)
edges.append((row_parsed[0], dst))
mapping = {map_from: map_to for map_to, map_from in enumerate(sorted(nodes))}
edges = list(map(lambda tup: (mapping[tup[0]], mapping[tup[1]]), edges))
result_indices = np.vstack(result_indices)
return edges, result_indices, len(nodes), mapping
def parse_minibatch(adjlists, edge_metapath_indices_list, idx_batch, device, samples=None):
g_list = []
result_indices_list = []
idx_batch_mapped_list = []
for adjlist, indices in zip(adjlists, edge_metapath_indices_list):
edges, result_indices, num_nodes, mapping = parse_adjlist(
[adjlist[i] for i in idx_batch], [indices[i] for i in idx_batch], samples)
g = dgl.DGLGraph(multigraph=True)
g.add_nodes(num_nodes)
if len(edges) > 0:
sorted_index = sorted(range(len(edges)), key=lambda i : edges[i])
g.add_edges(*list(zip(*[(edges[i][1], edges[i][0]) for i in sorted_index])))
result_indices = torch.LongTensor(result_indices[sorted_index]).to(device)
else:
result_indices = torch.LongTensor(result_indices).to(device)
#g.add_edges(*list(zip(*[(dst, src) for src, dst in sorted(edges)])))
#result_indices = torch.LongTensor(result_indices).to(device)
g_list.append(g)
result_indices_list.append(result_indices)
idx_batch_mapped_list.append(np.array([mapping[idx] for idx in idx_batch]))
return g_list, result_indices_list, idx_batch_mapped_list
def parse_adjlist_LastFM(adjlist, edge_metapath_indices, samples=None, exclude=None, offset=None, mode=None):
edges = []
nodes = set()
result_indices = []
for row, indices in zip(adjlist, edge_metapath_indices):
row_parsed = list(map(int, row.split(' ')))
nodes.add(row_parsed[0])
if len(row_parsed) > 1:
# sampling neighbors
if samples is None:
if exclude is not None:
if mode == 0:
mask = [False if [u1, a1 - offset] in exclude or [u2, a2 - offset] in exclude else True for u1, a1, u2, a2 in indices[:, [0, 1, -1, -2]]]
else:
mask = [False if [u1, a1 - offset] in exclude or [u2, a2 - offset] in exclude else True for a1, u1, a2, u2 in indices[:, [0, 1, -1, -2]]]
neighbors = np.array(row_parsed[1:])[mask]
result_indices.append(indices[mask])
else:
neighbors = row_parsed[1:]
result_indices.append(indices)
else:
# undersampling frequent neighbors
unique, counts = np.unique(row_parsed[1:], return_counts=True)
p = []
for count in counts:
p += [(count ** (3 / 4)) / count] * count
p = np.array(p)
p = p / p.sum()
samples = min(samples, len(row_parsed) - 1)
sampled_idx = np.sort(np.random.choice(len(row_parsed) - 1, samples, replace=False, p=p))
if exclude is not None:
if mode == 0:
mask = [False if [u1, a1 - offset] in exclude or [u2, a2 - offset] in exclude else True for u1, a1, u2, a2 in indices[sampled_idx][:, [0, 1, -1, -2]]]
else:
mask = [False if [u1, a1 - offset] in exclude or [u2, a2 - offset] in exclude else True for a1, u1, a2, u2 in indices[sampled_idx][:, [0, 1, -1, -2]]]
neighbors = np.array([row_parsed[i + 1] for i in sampled_idx])[mask]
result_indices.append(indices[sampled_idx][mask])
else:
neighbors = [row_parsed[i + 1] for i in sampled_idx]
result_indices.append(indices[sampled_idx])
else:
neighbors = [row_parsed[0]]
indices = np.array([[row_parsed[0]] * indices.shape[1]])
if mode == 1:
indices += offset
result_indices.append(indices)
for dst in neighbors:
nodes.add(dst)
edges.append((row_parsed[0], dst))
mapping = {map_from: map_to for map_to, map_from in enumerate(sorted(nodes))}
edges = list(map(lambda tup: (mapping[tup[0]], mapping[tup[1]]), edges))
result_indices = np.vstack(result_indices)
return edges, result_indices, len(nodes), mapping
def parse_minibatch_LastFM(adjlists_ua, edge_metapath_indices_list_ua, user_artist_batch, device, samples=None, use_masks=None, offset=None):
g_lists = [[], []]
result_indices_lists = [[], []]
idx_batch_mapped_lists = [[], []]
for mode, (adjlists, edge_metapath_indices_list) in enumerate(zip(adjlists_ua, edge_metapath_indices_list_ua)):
for adjlist, indices, use_mask in zip(adjlists, edge_metapath_indices_list, use_masks[mode]):
if use_mask:
edges, result_indices, num_nodes, mapping = parse_adjlist_LastFM(
[adjlist[row[mode]] for row in user_artist_batch], [indices[row[mode]] for row in user_artist_batch], samples, user_artist_batch, offset, mode)
else:
edges, result_indices, num_nodes, mapping = parse_adjlist_LastFM(
[adjlist[row[mode]] for row in user_artist_batch], [indices[row[mode]] for row in user_artist_batch], samples, offset=offset, mode=mode)
g = dgl.DGLGraph(multigraph=True)
g.add_nodes(num_nodes)
if len(edges) > 0:
sorted_index = sorted(range(len(edges)), key=lambda i : edges[i])
g.add_edges(*list(zip(*[(edges[i][1], edges[i][0]) for i in sorted_index])))
result_indices = torch.LongTensor(result_indices[sorted_index]).to(device)
else:
result_indices = torch.LongTensor(result_indices).to(device)
g_lists[mode].append(g)
result_indices_lists[mode].append(result_indices)
idx_batch_mapped_lists[mode].append(np.array([mapping[row[mode]] for row in user_artist_batch]))
return g_lists, result_indices_lists, idx_batch_mapped_lists
class index_generator:
def __init__(self, batch_size, num_data=None, indices=None, shuffle=True):
if num_data is not None:
self.num_data = num_data
self.indices = np.arange(num_data)
if indices is not None:
self.num_data = len(indices)
self.indices = np.copy(indices)
self.batch_size = batch_size
self.iter_counter = 0
self.shuffle = shuffle
if shuffle:
np.random.shuffle(self.indices)
def next(self):
if self.num_iterations_left() <= 0:
self.reset()
self.iter_counter += 1
return np.copy(self.indices[(self.iter_counter - 1) * self.batch_size:self.iter_counter * self.batch_size])
def num_iterations(self):
return int(np.ceil(self.num_data / self.batch_size))
def num_iterations_left(self):
return self.num_iterations() - self.iter_counter
def reset(self):
if self.shuffle:
np.random.shuffle(self.indices)
self.iter_counter = 0