-
Notifications
You must be signed in to change notification settings - Fork 23
/
AnonymousWalkEmbeddings.py
546 lines (451 loc) · 22.8 KB
/
AnonymousWalkEmbeddings.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
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
'''
Tensorflow implementation of distributed Anonymous Walks Embeddings (AWE).
@author: Sergey Ivanov ([email protected] -- Adaptation for graph2vec via AW
@references:
'''
from __future__ import division, print_function
import os
import math
import random
import argparse
import sys
import time
import re
import shutil
import threading
from collections import Counter
import numpy as np
import tensorflow as tf
from AnonymousWalkKernel import AnonymousWalks, GraphKernel, Evaluation
from sklearn import svm
from sklearn.metrics import accuracy_score
from sklearn.model_selection import train_test_split
SEED = 2018
class AWE(object):
'''
Computes distributed Anonymous Walk Embeddings.
'''
def __init__(self,
dataset='imdb_b',
batch_size=128,
window_size=8,
concat=False,
embedding_size_w=64,
embedding_size_d=64,
loss_type='sampled_softmax',
num_samples=64,
optimize='Adagrad',
learning_rate=1.0,
root = '../',
ext = 'graphml',
steps = 7,
epochs = 1,
batches_per_epoch = 1,
candidate_func = None,
graph_labels = None,
regenerate_corpus = False,
neighborhood_size=1):
'''
Initialize AWE model.
:param dataset: name of the dataset and corresponding name of the folder.
:param batch_size: number of batches per iteration of AWE model.
:param window_size: number of context words.
:param concat: Concatenate context words or not.
:param embedding_size_w: embedding size of word
:param embedding_size_d: embedding size of document
:param loss_type: sampled softmax or nce
:param num_samples: number of (negative) samples for every target word.
:param optimize: SGD or Adagrad
:param learning_rate: learning rate of the model
:param root: root folder of the dataset
:param ext: extension of files with graphs (e.g. graphml)
:param steps: length of anonymous walk
:param epochs: number of epochs for iterations
:param batches_per_epoch: number of batches per epoch for each graph
:param candidate_func: None (loguniform by default) or uniform
:param graph_labels: None, edges, nodes, edges_nodes
'''
# bind params to class
self.batch_size = batch_size
self.window_size = window_size
self.concat = concat
self.embedding_size_w = embedding_size_w
self.embedding_size_d = embedding_size_d
self.loss_type = loss_type
self.num_samples = num_samples
self.optimize = optimize
self.learning_rate = learning_rate
self.candidate_func = candidate_func
self.graph_labels = graph_labels
self.ROOT = root
self.ext = ext
self.steps = steps
self.epochs = epochs
self.dataset = dataset
self.batches_per_epoch = batches_per_epoch
# switch to have batches_per_epoch = N for every graph with N nodes
self.flag2iterations = False
if batches_per_epoch is None:
self.flag2iterations = True
# get all graph filenames (document size)
self.folder = self.ROOT + self.dataset + '/'
folder_graphs = filter(lambda g: g.endswith(max(self.ext, '')), os.listdir(self.folder))
self.sorted_graphs = sorted(folder_graphs, key=lambda g: int(re.findall(r'\d+', g)[0]))
self.document_size = len(self.sorted_graphs)
print('Number of graphs: {}'.format(self.document_size))
print('Generating corpus... ', end='')
self.corpus_fn_name = '{}.corpus'
self.regenerate_corpus = regenerate_corpus
self.neiborhood_size = neighborhood_size
start2gen = time.time()
self.generate_corpus()
print('Finished {}'.format(time.time() - start2gen))
self.vocabulary_size = max(self.walk_ids.values()) + 1
print('Number of words: {}'.format(self.vocabulary_size))
# init all variables in a tensorflow graph
self._init_graph()
# create a session
self.sess = tf.Session(graph=self.graph)
def generate_corpus(self):
# get all AW (vocabulary size)
self.g2v = AnonymousWalks()
if self.graph_labels is None:
self.g2v._all_paths(self.steps, keep_last=True)
elif self.graph_labels == 'nodes':
self.g2v._all_paths_nodes(self.steps, keep_last=True)
elif self.graph_labels == 'edges':
self.g2v._all_paths_edges(self.steps, keep_last=True)
elif self.graph_labels == 'edges_nodes':
self.g2v._all_paths_edges_nodes(self.steps, keep_last=True)
self.walk_ids = dict()
for i, path in enumerate(self.g2v.paths[self.steps]):
self.walk_ids[tuple(path)] = i
self.nodes_per_graphs = dict()
label_suffix = ''
if graph_labels is not None:
label_suffix = '_' + graph_labels
if self.regenerate_corpus == True or not os.path.exists(self.ROOT + self.dataset + '_corpus' + label_suffix):
if not os.path.exists(self.ROOT + self.dataset + '_corpus' + label_suffix):
os.mkdir(self.ROOT + self.dataset + '_corpus' + label_suffix)
for en, graph_fn in enumerate(self.sorted_graphs):
if en > 0 and not en%100:
print(f"Graph {en}")
g2v = AnonymousWalks()
g2v.read_graphml(self.folder + graph_fn)
self.nodes_per_graphs[en] = len(g2v.graph)
g2v.write_corpus(self.neiborhood_size, self.walk_ids, steps, self.graph_labels,
self.ROOT + self.dataset + '_corpus{}/{}'.format(label_suffix, self.corpus_fn_name.format(en)))
def _init_graph(self):
'''
Init a tensorflow Graph containing:
input data, variables, model, loss function, optimizer
'''
self.graph = tf.Graph()
with self.graph.as_default(), tf.device('/cpu:0'):
tf.set_random_seed(SEED)
self.train_dataset = tf.placeholder(tf.int32, shape=[self.batch_size, self.window_size+1])
self.train_labels = tf.placeholder(tf.int32, shape=[self.batch_size, 1])
# embeddings for anonymous walks
self.word_embeddings = tf.Variable(
tf.random_uniform([self.vocabulary_size, self.embedding_size_w], -1.0, 1.0))
# embedding for graphs
self.doc_embeddings = tf.Variable(
tf.random_uniform([self.document_size, self.embedding_size_d], -1.0, 1.0))
if self.concat: # concatenating word vectors and doc vector
combined_embed_vector_length = self.embedding_size_w * self.window_size + self.embedding_size_d
else: # concatenating the average of word vectors and the doc vector
combined_embed_vector_length = self.embedding_size_w + self.embedding_size_d
# softmax weights, W and D vectors should be concatenated before applying softmax
self.weights = tf.Variable(
tf.truncated_normal([self.vocabulary_size, combined_embed_vector_length],
stddev=1.0 / math.sqrt(combined_embed_vector_length)))
# softmax biases
self.biases = tf.Variable(tf.zeros([self.vocabulary_size]))
# shape: (batch_size, embeddings_size)
embed = [] # collect embedding matrices with shape=(batch_size, embedding_size)
if self.concat:
for j in range(self.window_size):
embed_w = tf.nn.embedding_lookup(self.word_embeddings, self.train_dataset[:, j])
embed.append(embed_w)
else:
# averaging word vectors
embed_w = tf.zeros([self.batch_size, self.embedding_size_w])
for j in range(self.window_size):
embed_w += tf.nn.embedding_lookup(self.word_embeddings, self.train_dataset[:, j])
embed.append(embed_w)
embed_d = tf.nn.embedding_lookup(self.doc_embeddings, self.train_dataset[:, self.window_size])
embed.append(embed_d)
# concat word and doc vectors
self.embed = tf.concat(embed, 1)
# choosing negative sampling function
sampled_values = None # log uniform by default
if self.candidate_func == 'uniform': # change to uniform
sampled_values = tf.nn.uniform_candidate_sampler(
true_classes=tf.to_int64(self.train_labels),
num_true=1,
num_sampled=self.num_samples,
unique=True,
range_max=self.vocabulary_size)
# Compute the loss, using a sample of the negative labels each time.
loss = None
if self.loss_type == 'sampled_softmax':
loss = tf.nn.sampled_softmax_loss(self.weights, self.biases, self.train_labels,
self.embed,
self.num_samples,
self.vocabulary_size,
sampled_values = sampled_values)
elif self.loss_type == 'nce':
loss = tf.nn.nce_loss(self.weights, self.biases, self.train_labels,
self.embed, self.num_samples, self.vocabulary_size,
sampled_values=sampled_values)
self.loss = tf.reduce_mean(loss)
# Optimizer.
if self.optimize == 'Adagrad':
self.optimizer = tf.train.AdagradOptimizer(self.learning_rate).minimize(loss)
elif self.optimize == 'SGD':
self.optimizer = tf.train.GradientDescentOptimizer(self.learning_rate).minimize(loss)
# Normalize embeddings
norm_w = tf.sqrt(tf.reduce_sum(tf.square(self.word_embeddings), 1, keep_dims=True))
self.normalized_word_embeddings = self.word_embeddings/norm_w
norm_d = tf.sqrt(tf.reduce_sum(tf.square(self.doc_embeddings), 1, keep_dims=True))
self.normalized_doc_embeddings = self.doc_embeddings/norm_d
self.init_op = tf.global_variables_initializer()
self.saver = tf.train.Saver()
def _train_thread_body(self):
'''Train model on random anonymous walk batches.'''
label_suffix = ''
if self.graph_labels is not None:
label_suffix = '_' + graph_labels
while True:
batch_data, batch_labels = self.g2v.generate_file_batch(batch_size, window_size, self.doc_id,
self.ROOT + self.dataset + '_corpus{}/{}'.format(label_suffix, self.corpus_fn_name.format(self.doc_id)),
self.nodes_per_graphs[self.doc_id])
# batch_data, batch_labels = self.g2v.generate_random_batch(batch_size=self.batch_size,
# window_size=self.window_size,
# steps=self.steps, walk_ids=self.walk_ids,
# doc_id=self.doc_id,
# graph_labels = self.graph_labels)
feed_dict = {self.train_dataset: batch_data, self.train_labels: batch_labels}
op, l = self.sess.run([self.optimizer, self.loss], feed_dict=feed_dict)
self.sample += 1
self.global_step += 1
self.average_loss += l
# The average loss is an estimate of the loss over the last 100 batches.
# if self.global_step % 100 == 0:
# print('Average loss at step %d: %f' % (self.global_step, self.average_loss))
# self.average_loss = 0
if self.sample >= self.batches_per_epoch:
break
def train(self):
'''Train the model.'''
session = self.sess
session.run(self.init_op)
self.average_loss = 0
self.global_step = 0
print('Initialized')
random_order = list(range(len(self.sorted_graphs)))
random.shuffle(random_order)
for ep in range(self.epochs):
print('Epoch: {}'.format(ep))
time2epoch = time.time()
for rank_id, doc_id in enumerate(random_order):
# for doc_id, graph_fn in enumerate(self.sorted_graphs):
# graph_fn = self.sorted_graphs[doc_id]
time2graph = time.time()
self.sample = 0
self.doc_id = doc_id
# self.g2v.read_graphml(self.folder + graph_fn)
# self.g2v.create_random_walk_graph()
# print('{}-{}. Graph-{}: {} nodes'.format(ep, rank_id, doc_id, len(self.g2v.rw_graph)))
# if self.flag2iterations == True: # take sample of N words per each graph with N nodes
# self.batches_per_epoch = len(self.g2v.rw_graph)
self._train_thread_body()
if rank_id > 0 and not rank_id%100:
print('Graph {}-{}: {:.2f}'.format(ep, rank_id, time.time() - time2graph))
print('Time for epoch {}: {:.2f}'.format(ep, time.time() - time2epoch))
# save temporary embeddings
if not ep%10:
self.graph_embeddings = session.run(self.normalized_doc_embeddings)
np.savez_compressed(RESULTS_FOLDER + '/' + dataset + '/tmp/embeddings_{}.txt'.format(ep), E=self.graph_embeddings)
self.graph_embeddings = session.run(self.normalized_doc_embeddings)
return self
if __name__ == '__main__':
# Set random seeds
SEED = 2018
random.seed(SEED)
np.random.seed(SEED)
dataset = 'mutag'
batch_size = 100
window_size = 16
embedding_size_w = 128
embedding_size_d = 128
num_samples = 10
concat = False
loss_type = 'sampled_softmax'
optimize = 'Adagrad'
learning_rate = 0.1
root = '../Datasets/'
ext = 'graphml'
steps = 10
epochs = 100
batches_per_epoch = 100
candidate_func = None
graph_labels = None
KERNEL = 'rbf'
RESULTS_FOLDER = 'doc2vec_results/'
TRIALS = 10 # number of cross-validation
regenerate_corpus = True
neighborhood_size = window_size + 1
parser = argparse.ArgumentParser(description='Getting classification accuracy for Graph Kernel Methods')
parser.add_argument('--dataset', default=dataset, help='Dataset with graphs to classify')
parser.add_argument('--batch_size', default=batch_size, help='Number of target words in a batch', type=int)
parser.add_argument('--window_size', default=window_size, help='Number of context words for target', type=int)
parser.add_argument('--embedding_size_w', default=embedding_size_w, help='Dimension of word embeddings', type=int)
parser.add_argument('--embedding_size_d', default=embedding_size_d, help='Dimension of document embeddings', type=int)
parser.add_argument('--num_samples', default=num_samples, help='Number of (negative) samples for objective functions', type=int)
parser.add_argument('--concat', default=concat, help='Concatenate or Average context words', type=bool)
parser.add_argument('--loss_type', default=loss_type, help='sampled_softmax or nce')
parser.add_argument('--optimize', default=optimize, help='Adagrad or SGD')
parser.add_argument('--learning_rate', default=learning_rate, help='Learning rate of optimizer')
parser.add_argument('--root', default=root, help='Root folder of dataset')
parser.add_argument('--ext', default=ext, help='Extension of graph filenames')
parser.add_argument('--results_folder', default=RESULTS_FOLDER, help='Folder to store results')
parser.add_argument('--steps', default=steps, help='Number of steps for AW', type=int)
parser.add_argument('--epochs', default=epochs, help='Number of epochs to train', type=int)
parser.add_argument('--batches_per_epoch', default=batches_per_epoch, help='Number of iterations per epoch for each graph', type=int)
parser.add_argument('--candidate_func', default=candidate_func, help='Sampling function for negatives: uniform or loguniform (None, by default)')
parser.add_argument('--graph_labels', default=graph_labels,
help='Graph labels to use (none, nodes, edges, edges_nodes)')
parser.add_argument('--regenerate_corpus', default=regenerate_corpus, type=bool,
help='If regenerate corpus for training. ')
parser.add_argument('--neighborhood_size', default=neighborhood_size, type=int,
help='Number of context words per line.')
args = parser.parse_args()
dataset = args.dataset
batch_size = args.batch_size
window_size = args.window_size
embedding_size_w = args.embedding_size_w
embedding_size_d = args.embedding_size_d
num_samples = args.num_samples
concat = args.concat
loss_type = args.loss_type
optimize = args.optimize
learning_rate = args.learning_rate
root = args.root
ext = args.ext
steps = args.steps
epochs = args.epochs
batches_per_epoch = args.batches_per_epoch
RESULTS_FOLDER = args.results_folder
candidate_func = args.candidate_func
graph_labels = args.graph_labels
regenerate_corpus = args.regenerate_corpus
neighborhood_size = args.neighborhood_size
if not os.path.exists(RESULTS_FOLDER):
os.makedirs(RESULTS_FOLDER)
if not os.path.exists(RESULTS_FOLDER + '/' + dataset):
os.makedirs(RESULTS_FOLDER + '/' + dataset)
if not os.path.exists(RESULTS_FOLDER + '/' + dataset + '/tmp/'):
os.makedirs(RESULTS_FOLDER + '/' + dataset + '/tmp/')
print('DATASET: {}'.format(dataset))
print('BATCHES PER EPOCH: {}'.format(batches_per_epoch))
print('BATCH SIZE: {}'.format(batch_size))
print('WINDOW SIZE: {}'.format(window_size))
print('NEGATIVES: {}'.format(num_samples))
print('EPOCHS: {}'.format(epochs))
print('')
print('EMBEDDING WORD SIZE: {}'.format(embedding_size_w))
print('EMBEDDING GRAPH SIZE: {}'.format(embedding_size_w))
print('LENGTH: {}'.format(steps))
print('')
# initialize model
awe = AWE(dataset = dataset, batch_size = batch_size, window_size = window_size,
embedding_size_w = embedding_size_w, embedding_size_d = embedding_size_d,
num_samples = num_samples, concat = concat, loss_type = loss_type,
optimize = optimize, learning_rate = learning_rate, root = root,
ext = ext, steps = steps, epochs = epochs, batches_per_epoch = batches_per_epoch,
candidate_func = candidate_func, graph_labels=graph_labels, regenerate_corpus=regenerate_corpus,
neighborhood_size=neighborhood_size)
print()
start2emb = time.time()
awe.train() # get embeddings
finish2emb = time.time()
print()
print('Time to compute embeddings: {:.2f} sec'.format(finish2emb - start2emb))
# E = np.load('imdb_b.embeddings.txt')['E']
gk = GraphKernel()
gk.embeddings = awe.graph_embeddings
gk.write_embeddings(RESULTS_FOLDER + '/' + dataset + '/embeddings.txt')
# gk.embeddings = E
# read classes for each graph
y = []
with open(root + dataset + '/labels.txt') as f:
for line in f:
y.extend(list(map(int, line.strip().split())))
y = np.array(y)
### testing on embeddings
accuracies = []
for _ in range(10):
E = gk.embeddings
idx_train, idx_test = train_test_split(list(range(E.shape[0])), test_size=0.2)
E_train = E[idx_train, :]
y_train = y[idx_train]
E_test = E[idx_test, :]
y_test = y[idx_test]
model = svm.SVC(kernel='rbf', C=1, gamma = 1)
model.fit(E_train, y_train)
y_predicted = model.predict(E_test)
accuracies.append(accuracy_score(y_test, y_predicted))
print('On Embeddings:', accuracy_score(y_test, y_predicted))
print(np.max(accuracies), np.mean(accuracies), np.std(accuracies))
################## Estimate results: Classification Accuracy ########################
print()
for KERNEL in ['rbf', 'linear', 'poly']:
if KERNEL == 'rbf':
sigma_grid = [0.00001, 0.0001, 0.001, 0.1, 1, 10]
else:
sigma_grid = [1]
# try:
for s_ix in range(len(sigma_grid)):
print('Setup: ',dataset, KERNEL, sigma_grid[s_ix])
sys.stdout.flush()
print('Computing Kernel Matrix...')
start2kernelmatrix = time.time()
gk.kernel_matrix(kernel_method=KERNEL, build_embeddings=False, sigma=sigma_grid[s_ix])
finish2kernelmatrix = time.time()
print('Time to compute Kernel Matrix: ', finish2kernelmatrix - start2kernelmatrix)
sys.stdout.flush()
# write kernel matrix and embeddings
# gk.write_kernel_matrix('{}/{}/kernel_{}_{}.txt'.format(RESULTS_FOLDER, dataset, KERNEL, sigma_grid[s_ix]))
# dump = np.load('{}/{}/kernel_{}_{}.txt.npz'.format(RESULTS_FOLDER, dataset, KERNEL, sigma_grid[s_ix]))
# gk.K = dump['K']
N, M = gk.K.shape
print('Kernel matrix shape: {}x{}'.format(N, M))
sys.stdout.flush()
# run k-fold SVM with cross-validation on C
print('Evaluating Kernel Matrix on SVM...')
ev = Evaluation(gk.K, y, verbose=False)
optimal_test_scores = []
for _ in range(TRIALS):
print(TRIALS - _, end=' ')
sys.stdout.flush()
accs = ev.evaluate()
optimal_test_scores.extend(accs)
print()
print('Average Performance on Test: {:.2f}% +-{:.2f}%'.format(100*np.mean(optimal_test_scores),
100*np.std(optimal_test_scores)))
sys.stdout.flush()
# append results of dataset to the file
with open('{}/{}/performance_{}_{}_{}.txt'.format(RESULTS_FOLDER, dataset, dataset, KERNEL, steps), 'a') as f:
f.write('{} {} {} {} {} {} {}\n'.format(dataset, KERNEL, steps, sigma_grid[s_ix],
np.mean(optimal_test_scores), np.std(optimal_test_scores),
finish2kernelmatrix - start2kernelmatrix))
print()
# except Exception as e:
# print('ERROR FOR', dataset, KERNEL, steps)
# raise e
label_suffix = ''
if graph_labels is not None:
label_suffix = '_' + graph_labels
shutil.rmtree(root + dataset + '_corpus{}'.format(label_suffix))
console = []