-
Notifications
You must be signed in to change notification settings - Fork 2
/
utils.py
432 lines (356 loc) · 16.5 KB
/
utils.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
##############################################################
#
# utils.py
# This file contains various functions that are applied in
# the training loops.
# They convert batch data into tensors, feed them to the models,
# compute the loss and propagate it.
#
##############################################################
import math
import torch
import networkx as nx
import numpy as np
from sklearn.preprocessing import LabelEncoder
import re
from sklearn.model_selection import train_test_split
from transformers import BertTokenizer
from transformers import BertForSequenceClassification, AdamW, BertConfig
from torch.utils.data import TensorDataset, DataLoader, RandomSampler, SequentialSampler
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader, random_split
from torch.utils.data.sampler import SubsetRandomSampler
import transformers
# get_linear_schedule_with_warmup
from transformers import BertTokenizer, BertModel, AdamW
from transformers import get_linear_schedule_with_warmup
import time
import timeit
def my_collate1(batches):
# return batches
return [{key: torch.stack(value) for key, value in batch.items()} for batch in batches]
# all_data = []
# for batch in batches:
# data = {}
# for key, value in batch.items():
# import pdb;pdb.set_trace()
# data[key] = torch.stack(value)
# all_data.append(data)
#
# return all_data
def loss_fun(outputs, targets):
loss = nn.CrossEntropyLoss()
return loss(outputs, targets)
# return nn.BCEWithLogitsLoss()(outputs, targets)
def evaluate(target, predicted):
true_label_mask = [1 if (np.argmax(x)-target[i]) ==
0 else 0 for i, x in enumerate(predicted)]
nb_prediction = len(true_label_mask)
true_prediction = sum(true_label_mask)
false_prediction = nb_prediction-true_prediction
accuracy = true_prediction/nb_prediction
return{
"accuracy": accuracy,
"nb exemple": len(target),
"true_prediction": true_prediction,
"false_prediction": false_prediction,
}
def train_loop_fun1_orig(data_loader, model, optimizer, device, scheduler=None):
model.train()
t0 = time.time()
losses = []
#import pdb;pdb.set_trace()
for batch_idx, batch in enumerate(data_loader):
# model.half()
# ids_batch=[data["ids"] for data in batch]
# mask_batch=[data["mask"] for data in batch]
# token_type_ids_batch = [data["token_type_ids"] for data in batch]
# targets_batch = [data["targets"] for data in batch]
# lengt_batch=[data['len'] for data in batch]
ids = [data["ids"] for data in batch] # size of 8
mask = [data["mask"] for data in batch]
token_type_ids = [data["token_type_ids"] for data in batch]
targets = [data["targets"] for data in batch] # length: 8
length = [data['len'] for data in batch] # [tensor([3]), tensor([7]), tensor([2]), tensor([4]), tensor([2]), tensor([4]), tensor([2]), tensor([3])]
ids = torch.cat(ids)
mask = torch.cat(mask)
token_type_ids = torch.cat(token_type_ids)
targets = torch.cat(targets)
length = torch.cat(length)
# for doc in range(len(lengt_batch)):
# ids=ids_batch[doc]
# mask=mask_batch[doc]
# token_type_ids=token_type_ids_batch[doc]
# targets=targets_batch[doc]
# lengt=lengt_batch[doc]
ids = ids.to(device, dtype=torch.long)
mask = mask.to(device, dtype=torch.long)
token_type_ids = token_type_ids.to(device, dtype=torch.long)
targets = targets.to(device, dtype=torch.long)
optimizer.zero_grad()
outputs = model(ids=ids, mask=mask, token_type_ids=token_type_ids)
loss = loss_fun(outputs, targets)
loss.backward()
model.float()
optimizer.step()
if scheduler:
scheduler.step()
losses.append(loss.item())
if batch_idx % 500 == 0:
print(
f"___ batch index = {batch_idx} / {len(data_loader)} ({100*batch_idx / len(data_loader):.2f}%), loss = {np.mean(losses[-10:]):.4f}, time = {time.time()-t0:.2f} secondes ___")
t0 = time.time()
return losses
def get_graph_features(graph):
'more https://networkx.org/documentation/stable/reference/algorithms/approximation.html'
try:
# import pdb;pdb.set_trace()
node_number = nx.number_of_nodes(graph) # int
centrality = nx.degree_centrality(graph) # a dictionary
centrality = sum(centrality.values())/node_number
edge_number = nx.number_of_edges(graph) # int
degrees = dict(graph.degree) # a dictionary
degrees = sum(degrees.values()) /edge_number
density = nx.density(graph) # a float
clustring_coef = nx.average_clustering(graph) # a float Compute the average clustering coefficient for the graph G.
closeness_centrality = nx.closeness_centrality(graph) # dict
closeness_centrality = sum(closeness_centrality.values())/len(closeness_centrality)
number_triangles = nx.triangles(graph) # dict
number_triangles = sum(number_triangles.values())/len(number_triangles)
number_clique = nx.graph_clique_number(graph) # a float Returns the number of maximal cliques in the graph.
number_connected_components = nx.number_connected_components(graph) # int Returns the number of connected components.
# avg_shortest_path_len = nx.average_shortest_path_length(graph) # float Return the average shortest path length; The average shortest path length is the sum of path lengths d(u,v) between all pairs of nodes (assuming the length is zero if v is not reachable from v) normalized by n*(n-1) where n is the number of nodes in G.
# diameter = nx.distance_measures.diameter(graph) # int The diameter is the maximum eccentricity.
return {'node_number': node_number, 'edge_number': edge_number, 'centrality': centrality, 'degrees': degrees,
'density': density, 'clustring_coef': clustring_coef, 'closeness_centrality': closeness_centrality,
'number_triangles': number_triangles, 'number_clique': number_clique,
'number_connected_components': number_connected_components,
'avg_shortest_path_len': 0, 'diameter': 0}
except:
return {'node_number': 1, 'edge_number': 1, 'centrality': 0, 'degrees': 0,
'density': 0, 'clustring_coef': 0, 'closeness_centrality': 0,
'number_triangles': 0, 'number_clique': 0,
'number_connected_components': 0,
'avg_shortest_path_len': 0, 'diameter': 0}
def graph_feature_stats(graph_feature_list):
total_number = len(graph_feature_list)
stats = {k:[] for k in graph_feature_list[0].keys()}
for feature_dict in graph_feature_list:
for key in stats.keys():
stats[key].append(feature_dict[key])
'get mean'
stats_mean = {k:sum(v)/len(v) for (k,v) in stats.items()}
return stats_mean
def train_loop_fun1(data_loader, model, optimizer, device, scheduler=None):
'''optimized function for Hi-BERT'''
model.train()
t0 = time.time()
losses = []
#import pdb;pdb.set_trace()
graph_features = []
for batch_idx, batch in enumerate(data_loader):
ids = [data["ids"] for data in batch] # size of 8
mask = [data["mask"] for data in batch]
token_type_ids = [data["token_type_ids"] for data in batch]
targets = [data["targets"] for data in batch] # length: 8
length = [data['len'] for data in batch] # [tensor([3]), tensor([7]), tensor([2]), tensor([4]), tensor([2]), tensor([4]), tensor([2]), tensor([3])]
'cat is not working for hi-bert'
# ids = torch.cat(ids)
# mask = torch.cat(mask)
# token_type_ids = torch.cat(token_type_ids)
# targets = torch.cat(targets)
# length = torch.cat(length)
# ids = ids.to(device, dtype=torch.long)
# mask = mask.to(device, dtype=torch.long)
# token_type_ids = token_type_ids.to(device, dtype=torch.long)
# targets = targets.to(device, dtype=torch.long)
target_labels = torch.stack([x[0] for x in targets]).long().to(device)
optimizer.zero_grad()
# measure time
start = timeit.timeit()
outputs,adj_graph = model(ids=ids, mask=mask, token_type_ids=token_type_ids)
end = timeit.timeit()
model_time = end - start
loss = loss_fun(outputs, target_labels)
loss.backward()
model.float()
optimizer.step()
if scheduler:
scheduler.step()
losses.append(loss.item())
if batch_idx % 500 == 0:
print(
f"___ batch index = {batch_idx} / {len(data_loader)} ({100*batch_idx / len(data_loader):.2f}%), loss = {np.mean(losses[-10:]):.4f}, time = {time.time()-t0:.2f} secondes ___")
t0 = time.time()
graph_features.append(get_graph_features(adj_graph))
stats_mean = graph_feature_stats(graph_features)
import pprint
pprint.pprint(stats_mean)
return losses
def eval_loop_fun1(data_loader, model, device):
model.eval()
fin_targets = []
fin_outputs = []
losses = []
for batch_idx, batch in enumerate(data_loader):
ids = [data["ids"] for data in batch] # size of 8
mask = [data["mask"] for data in batch]
token_type_ids = [data["token_type_ids"] for data in batch]
targets = [data["targets"] for data in batch] # length: 8
with torch.no_grad():
target_labels = torch.stack([x[0] for x in targets]).long().to(device)
outputs, _ = model(ids=ids, mask=mask, token_type_ids=token_type_ids)
loss = loss_fun(outputs, target_labels)
losses.append(loss.item())
fin_targets.append(target_labels.cpu().detach().numpy())
fin_outputs.append(torch.softmax(outputs, dim=1).cpu().detach().numpy())
return np.concatenate(fin_outputs), np.concatenate(fin_targets), losses
# return np.vstack(fin_outputs), np.vstack(fin_targets), losses
def eval_loop_fun1_orgi(data_loader, model, device):
model.eval()
fin_targets = []
fin_outputs = []
losses = []
for batch_idx, batch in enumerate(data_loader):
# model.half()
ids = [data["ids"] for data in batch]
mask = [data["mask"] for data in batch]
token_type_ids = [data["token_type_ids"] for data in batch]
targets = [data["targets"] for data in batch]
lengt = [data['len'] for data in batch]
# ids_batch=[data["ids"] for data in batch]
# mask_batch=[data["mask"] for data in batch]
# token_type_ids_batch = [data["token_type_ids"] for data in batch]
# targets_batch = [data["targets"] for data in batch]
# lengt_batch=[data['len'] for data in batch]
# for doc in range(len(lengt_batch)):
# ids=ids_batch[doc]
# mask=mask_batch[doc]
# token_type_ids=token_type_ids_batch[doc]
# targets=targets_batch[doc]
# lengt=lengt_batch[doc]
ids = torch.cat(ids)
mask = torch.cat(mask)
token_type_ids = torch.cat(token_type_ids)
targets = torch.cat(targets)
lengt = torch.cat(lengt)
ids = ids.to(device, dtype=torch.long)
mask = mask.to(device, dtype=torch.long)
token_type_ids = token_type_ids.to(device, dtype=torch.long)
targets = targets.to(device, dtype=torch.long)
with torch.no_grad():
outputs = model(ids=ids, mask=mask, token_type_ids=token_type_ids)
loss = loss_fun(outputs, targets)
losses.append(loss.item())
fin_targets.append(targets.cpu().detach().numpy())
fin_outputs.append(torch.softmax(
outputs, dim=1).cpu().detach().numpy())
return np.concatenate(fin_outputs), np.concatenate(fin_targets), losses
# return np.vstack(fin_outputs), np.vstack(fin_targets), losses
def rnn_train_loop_fun1(data_loader, model, optimizer, device, scheduler=None):
model.train()
t0 = time.time()
losses = []
for batch_idx, batch in enumerate(data_loader):
# model.half()
# ids_batch=[data["ids"] for data in batch]
# mask_batch=[data["mask"] for data in batch]
# token_type_ids_batch = [data["token_type_ids"] for data in batch]
# targets_batch = [data["targets"] for data in batch]
# lengt_batch=[data['len'] for data in batch]
ids = [data["ids"] for data in batch]
mask = [data["mask"] for data in batch]
token_type_ids = [data["token_type_ids"] for data in batch]
targets = [data["targets"][0] for data in batch]
lengt = [data['len'] for data in batch]
ids = torch.cat(ids)
mask = torch.cat(mask)
token_type_ids = torch.cat(token_type_ids)
targets = torch.stack(targets)
lengt = torch.cat(lengt)
lengt = [x.item() for x in lengt]
# for doc in range(len(lengt_batch)):
# ids=ids_batch[doc]
# mask=mask_batch[doc]
# token_type_ids=token_type_ids_batch[doc]
# targets=targets_batch[doc]
# lengt=lengt_batch[doc]
ids = ids.to(device, dtype=torch.long)
mask = mask.to(device, dtype=torch.long)
token_type_ids = token_type_ids.to(device, dtype=torch.long)
targets = targets.to(device, dtype=torch.long)
optimizer.zero_grad()
outputs = model(ids=ids, mask=mask,
token_type_ids=token_type_ids, lengt=lengt)
loss = loss_fun(outputs, targets)
loss.backward()
model.float()
optimizer.step()
if scheduler:
scheduler.step()
losses.append(loss.item())
if batch_idx % 640 == 0:
print(
f"___ batch index = {batch_idx} / {len(data_loader)} ({100*batch_idx / len(data_loader):.2f}%), loss = {np.mean(losses[-10:]):.4f}, time = {time.time()-t0:.2f} secondes ___")
t0 = time.time()
return losses
def rnn_eval_loop_fun1(data_loader, model, device):
model.eval()
fin_targets = []
fin_outputs = []
losses = []
for batch_idx, batch in enumerate(data_loader):
# model.half()
ids = [data["ids"] for data in batch]
mask = [data["mask"] for data in batch]
token_type_ids = [data["token_type_ids"] for data in batch]
targets = [data["targets"][0] for data in batch]
lengt = [data['len'] for data in batch]
# ids_batch=[data["ids"] for data in batch]
# mask_batch=[data["mask"] for data in batch]
# token_type_ids_batch = [data["token_type_ids"] for data in batch]
# targets_batch = [data["targets"] for data in batch]
# lengt_batch=[data['len'] for data in batch]
# for doc in range(len(lengt_batch)):
# ids=ids_batch[doc]
# mask=mask_batch[doc]
# token_type_ids=token_type_ids_batch[doc]
# targets=targets_batch[doc]
# lengt=lengt_batch[doc]
ids = torch.cat(ids)
mask = torch.cat(mask)
token_type_ids = torch.cat(token_type_ids)
targets = torch.stack(targets)
lengt = torch.cat(lengt)
lengt = [x.item() for x in lengt]
ids = ids.to(device, dtype=torch.long)
mask = mask.to(device, dtype=torch.long)
token_type_ids = token_type_ids.to(device, dtype=torch.long)
targets = targets.to(device, dtype=torch.long)
with torch.no_grad():
outputs = model(ids=ids, mask=mask,
token_type_ids=token_type_ids, lengt=lengt)
loss = loss_fun(outputs, targets)
losses.append(loss.item())
fin_targets.append(targets.cpu().detach().numpy())
fin_outputs.append(torch.softmax(
outputs, dim=1).cpu().detach().numpy())
return np.concatenate(fin_outputs), np.concatenate(fin_targets), losses
# return np.vstack(fin_outputs), np.vstack(fin_targets), losses
def kronecker_generator(node_number):
"""Return the Kron graph
"""
n = math.ceil(math.sqrt(node_number))
# binomial edge
nb, pb = 10, .5
A = np.random.binomial(nb, pb, n*n).reshape(n,n) / 10
B = np.random.binomial(nb, pb, n*n).reshape(n,n) / 10
A[A < 0.5] = 0
B[B < 0.5] = 0
prod = np.kron(A,B)
prod[prod > 0] = 1
# return truncated version
prod = prod[:node_number,:node_number]
G = nx.from_numpy_matrix(prod)
# add noise
return G