-
Notifications
You must be signed in to change notification settings - Fork 0
/
node.py
550 lines (424 loc) · 24.1 KB
/
node.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
547
548
549
550
import copy
import logging
import sys
import evaluate
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
from torch.utils.data import DataLoader
from datasets import concatenate_datasets
from transformers import AutoTokenizer, AutoModelForSequenceClassification, DataCollatorWithPadding, get_scheduler
from tqdm import tqdm
def init_model(name, model_type, num_classes):
# config = AutoConfig.from_pretrained(model_type, num_labels=num_classes, finetuning_task=task_name)
tokenizer = AutoTokenizer.from_pretrained(model_type, use_fast=True)
model = AutoModelForSequenceClassification.from_pretrained(model_type, num_labels=num_classes)
total_params = sum(p.numel() for p in model.parameters())
total_params = total_params / 1000000
logging.info('model parameters of %s_%s: %2.1fM' % (name, model_type, total_params))
return tokenizer, model
def init_optimizer(optimizer_type, model, lr, weight_decay=0., momentum=0.9):
if optimizer_type == 'sgd':
optimizer = torch.optim.SGD(model.parameters(), lr=lr, weight_decay=weight_decay, momentum=momentum)
elif optimizer_type == 'adam':
optimizer = torch.optim.Adam(model.parameters(), lr=lr, weight_decay=weight_decay, betas=(momentum, 0.999), eps=1e-8)
elif optimizer_type == 'adamw':
optimizer = torch.optim.AdamW(model.parameters(), lr=lr, weight_decay=weight_decay, betas=(momentum, 0.999), eps=1e-8)
else:
sys.exit("Not implemented optimizer, code exit, re-run to use correct optimizer")
return optimizer
def init_scheduler(scheduler_type, optimizer, num_warmup_steps=None, num_training_steps=None):
if scheduler_type == 'linear':
scheduler = get_scheduler('linear', optimizer=optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps)
elif scheduler_type == "cosine": # cosine
scheduler = get_scheduler('cosine', optimizer=optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps)
else:
sys.exit("Not implemented learning rate scheduler, code exit, re-run to use correct scheduler")
return scheduler
def preprocessing_raw_datasets(raw_dataset, task_name, tokenizer, max_seq_length, logits=None):
task_to_keys = {
"cola": ("sentence", None),
"mnli": ("premise", "hypothesis"),
"mrpc": ("sentence1", "sentence2"),
"qnli": ("question", "sentence"),
"qqp": ("question1", "question2"),
"rte": ("sentence1", "sentence2"),
"sst2": ("sentence", None),
"stsb": ("sentence1", "sentence2"),
"wnli": ("sentence1", "sentence2"),
"ax": ("premise", "hypothesis"),
}
sentence1_key, sentence2_key = task_to_keys[task_name] # 'sentence1' 'sentence2'
def preprocess_function(examples):
sentences = (examples[sentence1_key],) if sentence2_key is None else (examples[sentence1_key], examples[sentence2_key])
inputs = tokenizer(*sentences, padding=True, max_length=max_seq_length, truncation=True)
return inputs
encoded_dataset = raw_dataset.map(preprocess_function, batched=True)
encoded_dataset = encoded_dataset.remove_columns([sentence1_key, 'idx']) if sentence2_key is None else encoded_dataset.remove_columns([sentence1_key, sentence2_key, 'idx'])
# encoded_dataset = encoded_dataset.rename_column('label', 'labels')
if logits is not None:
if type(logits) == list:
for k in range(len(logits)):
encoded_dataset = encoded_dataset.add_column('logits{}'.format(k), logits[k].tolist())
else:
encoded_dataset = encoded_dataset.add_column('logits', logits.tolist())
return encoded_dataset
class Client:
def __init__(self, args, id, model_type, train_dataset=None):
self.args = args
self.id = id
self.name = 'client' + str(id)
self.model_type = model_type
self.device = args.device
self.optimizer_type = args.optimizer
self.scheduler_type = args.scheduler
self.max_seq_length = args.max_seq_length
self.batch_size = args.batch_size
self.tokenizer, self.model = init_model(self.name, self.model_type, self.args.num_classes)
self.data_collator = DataCollatorWithPadding(tokenizer=self.tokenizer)
self.train_dataset = train_dataset
self.E = args.E
self.lr = args.lr
self.weight_decay = args.weight_decay
self.momentum = args.momentum
self.warmup_steps = args.warmup_steps
def fork(self, w_glob):
self.model.load_state_dict(w_glob)
def local_update(self):
self.model.to(self.device).train()
train_dataset = preprocessing_raw_datasets(self.train_dataset, self.args.task_name, self.tokenizer, self.max_seq_length)
train_dataloader = DataLoader(train_dataset, shuffle=True, batch_size=self.batch_size, collate_fn=self.data_collator)
optimizer = init_optimizer(self.optimizer_type, self.model, self.lr, weight_decay=self.weight_decay, momentum=self.momentum)
scheduler = init_scheduler(self.scheduler_type, optimizer, num_warmup_steps=self.warmup_steps, num_training_steps=len(train_dataloader) * self.E)
for epoch in range(self.E):
train_loss = 0.
metric = evaluate.load(self.args.dataset, self.args.task_name)
for batch in tqdm(train_dataloader, desc='Iteration'):
batch = {k: v.to(self.device) for k, v in batch.items()}
outputs = self.model(**batch)
loss = outputs.loss
logits = outputs.logits
train_loss += loss.item()
if self.args.num_classes == 1:
prediction = logits.squeeze()
else:
prediction = torch.argmax(logits, dim=-1)
metric.add_batch(predictions=prediction, references=batch['labels'])
loss.backward()
optimizer.step()
scheduler.step()
optimizer.zero_grad()
train_loss = train_loss / len(train_dataloader)
train_results = metric.compute()
logging.info('Epoch {}/{}: train_loss={}, train_results={}'.format(epoch + 1, self.E, train_loss, train_results))
self.model.cpu()
def local_distillation(self, public_dataset, logits_glob):
self.model.to(self.device).train()
public_dataset = preprocessing_raw_datasets(public_dataset, self.args.task_name, self.tokenizer, self.max_seq_length, logits_glob)
public_dataloader = DataLoader(public_dataset, shuffle=True, batch_size=self.batch_size, collate_fn=self.data_collator)
optimizer = init_optimizer(self.optimizer_type, self.model, self.lr, weight_decay=self.weight_decay, momentum=self.momentum)
scheduler = init_scheduler(self.scheduler_type, optimizer, num_warmup_steps=self.warmup_steps, num_training_steps=len(public_dataloader) * self.E)
for epoch in range(self.args.dis_epochs):
train_loss = 0.
metric = evaluate.load(self.args.dataset, self.args.task_name)
for batch in tqdm(public_dataloader, desc='Distilling'):
batch = {k: v.to(self.device) for k, v in batch.items()}
hard_label = batch.pop('labels')
soft_label = batch.pop('logits')
outputs = self.model(**batch)
logits = outputs.logits
loss = F.cross_entropy(logits, torch.argmax(soft_label, dim=-1))
train_loss += loss.item()
if self.args.num_classes == 1:
prediction = logits.squeeze()
else:
prediction = torch.argmax(logits, dim=-1)
metric.add_batch(predictions=prediction, references=hard_label)
loss.backward()
optimizer.step()
scheduler.step()
optimizer.zero_grad()
train_loss = train_loss / len(public_dataloader)
train_results = metric.compute()
logging.info('Epoch {}/{}: train_loss={}, train_results={}'.format(epoch + 1, self.args.dis_epochs, train_loss, train_results))
self.model.cpu()
def compute_logits(self, public_dataset):
self.model.to(self.device).eval()
public_dataset = preprocessing_raw_datasets(public_dataset, self.args.task_name, self.tokenizer, self.max_seq_length)
public_dataloader = DataLoader(public_dataset, shuffle=False, batch_size=self.batch_size, collate_fn=self.data_collator)
logits = None
for batch in tqdm(public_dataloader, desc='Predicting'):
del batch['labels']
batch = {k: v.to(self.device) for k, v in batch.items()}
with torch.no_grad():
outputs = self.model(**batch)
logit = outputs.logits
if logits is None:
logits = logit.detach().cpu()
else:
logits = torch.cat([logits, logit.detach().cpu()], dim=0)
self.model.cpu()
return logits
def compute_logits_batch(self, batch_dataset):
self.model.to(self.device).train()
batch_dataset = preprocessing_raw_datasets(batch_dataset, self.args.task_name, self.tokenizer, self.max_seq_length)
batch_dataloader = DataLoader(batch_dataset, shuffle=False, batch_size=len(batch_dataset), collate_fn=self.data_collator)
logits = None
for batch in batch_dataloader:
del batch['labels']
batch = {k: v.to(self.device) for k, v in batch.items()}
# with torch.no_grad():
outputs = self.model(**batch)
logit = outputs.logits
if logits is None:
logits = logit.detach().cpu()
else:
logits = torch.cat([logits, logit.detach().cpu()], dim=0)
self.model.cpu()
return logits
def query_update(self, dot_product, batch_dataset, batch_logits_local, weight, batch_logits_glob):
self.model.to(self.device).train()
batch_dataset = preprocessing_raw_datasets(batch_dataset, self.args.task_name, self.tokenizer, self.max_seq_length)
batch_dataloader = DataLoader(batch_dataset, shuffle=False, batch_size=len(batch_dataset), collate_fn=self.data_collator)
optimizer = init_optimizer(self.optimizer_type, self.model, self.args.dis_lr, weight_decay=self.args.weight_decay, momentum=self.args.momentum)
scheduler = init_scheduler(self.scheduler_type, optimizer, num_warmup_steps=self.args.warmup_steps, num_training_steps=len(batch_dataloader) * self.args.dis_epochs)
train_loss = 0.
# metric = evaluate.load(self.args.dataset, self.args.task_name)
for batch in batch_dataloader:
batch = {k: v.to(self.device) for k, v in batch.items()}
hard_labels = batch.pop('labels')
outputs = self.model(**batch)
logits = outputs.logits
loss1 = weight * dot_product * F.cross_entropy(logits, torch.argmax(batch_logits_local, dim=-1).to(self.device))
loss2 = F.cross_entropy(logits, torch.argmax(batch_logits_glob, dim=-1).to(self.device))
loss = loss1 + loss2
train_loss += loss.item()
# if self.args.num_classes == 1:
# prediction = logits.squeeze()
# else:
# prediction = torch.argmax(logits, dim=-1)
# metric.add_batch(predictions=prediction, references=hard_labels)
loss.backward()
optimizer.step()
scheduler.step()
optimizer.zero_grad()
train_loss = train_loss / len(batch_dataloader)
# train_results = metric.compute()
# logging.info('Epoch {}/{}: train_loss={}, train_results={}'.format(epoch + 1, self.args.dis_epochs, train_loss, train_results))
# logging.info('train_loss={}, train_results={}'.format(train_loss, train_results))
self.model.cpu()
class Server:
def __init__(self, args, id, model_type, public_dataset=None):
self.args = args
self.id = id
self.name = 'server'
self.model_type = model_type
self.device = args.device
self.optimizer_type = args.optimizer
self.scheduler_type = args.scheduler
self.max_seq_length = args.max_seq_length
self.batch_size = args.batch_size
# self.loss_fn = nn.CrossEntropyLoss() if args.task_name != 'stsb' else nn.MSELoss()
self.tokenizer, self.model = init_model(self.name, self.model_type, self.args.num_classes) # 全局模型
self.data_collator = DataCollatorWithPadding(tokenizer=self.tokenizer)
self.public_dataset = public_dataset
def merge(self, w_locals, weights):
w_avg = copy.deepcopy(w_locals[0])
for n in w_avg.keys():
w_avg[n] = 0.
for k in range(len(w_locals)):
w_avg[n] += weights[k] * w_locals[k][n]
return w_avg
def centralized_training(self, train_dataset):
self.model.to(self.device).train()
train_dataset = preprocessing_raw_datasets(train_dataset, self.args.task_name, self.tokenizer, self.max_seq_length)
train_dataloader = DataLoader(train_dataset, shuffle=True, batch_size=self.batch_size, collate_fn=self.data_collator)
optimizer = init_optimizer(self.optimizer_type, self.model, self.args.lr, weight_decay=self.args.weight_decay, momentum=self.args.momentum)
scheduler = init_scheduler(self.scheduler_type, optimizer, num_warmup_steps=self.args.warmup_steps, num_training_steps=len(train_dataloader) * self.args.E)
for epoch in range(self.args.E):
train_loss = 0.
metric = evaluate.load(self.args.dataset, self.args.task_name)
for batch in tqdm(train_dataloader, desc='Iteration'):
batch = {k: v.to(self.device) for k, v in batch.items()}
outputs = self.model(**batch)
loss = outputs.loss
logits = outputs.logits
train_loss += loss.item()
if self.args.num_classes == 1:
prediction = logits.squeeze()
else:
prediction = torch.argmax(logits, dim=-1)
metric.add_batch(predictions=prediction, references=batch['labels'])
loss.backward()
optimizer.step()
scheduler.step()
optimizer.zero_grad()
train_loss = train_loss / len(train_dataloader)
train_results = metric.compute()
logging.info('Epoch {}/{}: train_loss={}, train_results={}'.format(epoch + 1, self.args.E, train_loss, train_results))
self.model.cpu()
def logit_ensemble(self, logits_locals, weights):
logits_glob = torch.zeros_like(logits_locals[0])
for k in range(len(logits_locals)):
logits_glob += weights[k] * logits_locals[k]
# logits_glob /= len(logits_locals)
return logits_glob
def logit_ensemble_with_ERA(self, logits_locals, weights):
logits_glob = torch.zeros_like(logits_locals[0])
for k in range(len(logits_locals)):
logits_glob += weights[k] * logits_locals[k]
# logits_glob /= len(logits_locals)
T = 0.1
logits_glob = torch.softmax(logits_glob / T, dim=-1)
return logits_glob
def batch_logit_ensemble(self, batch_logits_locals, weights):
batch_logits_glob = torch.zeros_like(batch_logits_locals[0])
for k in range(len(batch_logits_locals)):
batch_logits_glob += weights[k] * batch_logits_locals[k]
# batch_logits_glob /= len(batch_logits_locals)
return batch_logits_glob
def ensemble_distillation(self, public_dataset, logits_glob):
self.model.to(self.device).train()
public_dataset = preprocessing_raw_datasets(public_dataset, self.args.task_name, self.tokenizer, self.max_seq_length, logits_glob)
public_dataloader = DataLoader(public_dataset, shuffle=True, batch_size=self.batch_size, collate_fn=self.data_collator)
optimizer = init_optimizer(self.optimizer_type, self.model, self.args.dis_lr, weight_decay=self.args.weight_decay, momentum=self.args.momentum)
scheduler = init_scheduler(self.scheduler_type, optimizer, num_warmup_steps=self.args.warmup_steps, num_training_steps=len(public_dataloader) * self.args.dis_epochs)
for epoch in range(self.args.dis_epochs):
train_loss = 0.
metric = evaluate.load(self.args.dataset, self.args.task_name)
for batch in tqdm(public_dataloader, desc='Distilling'):
batch = {k: v.to(self.device) for k, v in batch.items()}
hard_label = batch.pop('labels')
soft_label = batch.pop('logits')
outputs = self.model(**batch)
logits = outputs.logits
if self.args.algorithm in ['fed_df', 'fed_ed']:
T = 1
loss = F.kl_div(F.log_softmax(logits / T, dim=-1), F.softmax(soft_label / T, dim=-1), reduction='batchmean') * (T ** 2)
elif self.args.algorithm == 'fed_kd':
loss = F.mse_loss(logits, soft_label)
elif self.args.algorithm == 'ds_fl':
loss = F.cross_entropy(logits, torch.argmax(soft_label, dim=-1))
else:
sys.exit("Not implemented algorithm, code exit, re-run to use correct algorithm")
train_loss += loss.item()
if self.args.num_classes == 1:
prediction = logits.squeeze()
else:
prediction = torch.argmax(logits, dim=-1)
metric.add_batch(predictions=prediction, references=hard_label)
loss.backward()
optimizer.step()
scheduler.step()
optimizer.zero_grad()
train_loss = train_loss / len(public_dataloader)
train_results = metric.compute()
logging.info('Epoch {}/{}: train_loss={}, train_results={}'.format(epoch + 1, self.args.dis_epochs, train_loss, train_results))
self.model.cpu()
def mhat_distillation(self, public_dataset, logits_locals, weights):
self.model.to(self.device).train()
public_dataset = preprocessing_raw_datasets(public_dataset, self.args.task_name, self.tokenizer, self.max_seq_length, logits_locals)
public_dataloader = DataLoader(public_dataset, shuffle=True, batch_size=self.batch_size, collate_fn=self.data_collator)
optimizer = init_optimizer(self.optimizer_type, self.model, self.args.dis_lr, weight_decay=self.args.weight_decay, momentum=self.args.momentum)
scheduler = init_scheduler(self.scheduler_type, optimizer, num_warmup_steps=self.args.warmup_steps, num_training_steps=len(public_dataloader) * self.args.dis_epochs)
for epoch in range(self.args.dis_epochs):
train_loss = 0.
metric = evaluate.load(self.args.dataset, self.args.task_name)
for batch in tqdm(public_dataloader, desc='Distilling'):
batch = {k: v.to(self.device) for k, v in batch.items()}
hard_label = batch.pop('labels')
soft_label = []
for k in range(len(logits_locals)):
soft_label.append(batch.pop('logits{}'.format(k)))
outputs = self.model(**batch)
logits = outputs.logits
loss = 0.
for k in range(len(logits_locals)):
tmp_kd_loss = weights[k] * F.cross_entropy(logits, torch.argmax(soft_label[k], dim=-1), reduction='mean')
loss += tmp_kd_loss
train_loss += loss.item()
if self.args.num_classes == 1:
prediction = logits.squeeze()
else:
prediction = torch.argmax(logits, dim=-1)
metric.add_batch(predictions=prediction, references=hard_label)
loss.backward()
optimizer.step()
scheduler.step()
optimizer.zero_grad()
train_loss = train_loss / len(public_dataloader)
train_results = metric.compute()
logging.info('Epoch {}/{}: train_loss={}, train_results={}'.format(epoch + 1, self.args.dis_epochs, train_loss, train_results))
self.model.cpu()
def compute_logits(self, public_dataset):
self.model.to(self.device).eval()
public_dataset = preprocessing_raw_datasets(public_dataset, self.args.task_name, self.tokenizer, self.max_seq_length)
public_dataloader = DataLoader(public_dataset, shuffle=False, batch_size=self.batch_size, collate_fn=self.data_collator)
logits = None
for batch in tqdm(public_dataloader, desc='Predicting'):
del batch['labels']
batch = {k: v.to(self.device) for k, v in batch.items()}
with torch.no_grad():
outputs = self.model(**batch)
logit = outputs.logits
if logits is None:
logits = logit.detach().cpu()
else:
logits = torch.cat([logits, logit.detach().cpu()], dim=0)
self.model.cpu()
return logits
def get_query_dataloader(self, query_dataset):
query_dataset = preprocessing_raw_datasets(query_dataset, self.args.task_name, self.tokenizer, self.max_seq_length)
query_dataloader = DataLoader(query_dataset, shuffle=True, batch_size=self.batch_size, collate_fn=self.data_collator)
return query_dataloader
def compute_query_loss(self, query_batch):
self.model.to(self.device).train()
batch = {k: v.to(self.device) for k, v in query_batch.items()}
hard_label = batch.pop('labels')
outputs = self.model(**batch)
logits = outputs.logits
loss = F.cross_entropy(logits.detach(), hard_label, reduction='mean')
self.model.cpu()
return loss
def batch_distillation(self, batch_dataset, batch_logits_locals, weights):
self.model.to(self.device).train()
batch_dataset = preprocessing_raw_datasets(batch_dataset, self.args.task_name, self.tokenizer, self.max_seq_length, batch_logits_locals)
batch_dataloader = DataLoader(batch_dataset, shuffle=True, batch_size=len(batch_dataset), collate_fn=self.data_collator)
optimizer = init_optimizer(self.optimizer_type, self.model, self.args.dis_lr, weight_decay=self.args.weight_decay, momentum=self.args.momentum)
scheduler = init_scheduler(self.scheduler_type, optimizer, num_warmup_steps=self.args.warmup_steps, num_training_steps=len(batch_dataloader) * self.args.dis_epochs)
# for epoch in range(self.args.dis_epochs):
train_loss = 0.
# metric = evaluate.load(self.args.dataset, self.args.task_name)
for batch in batch_dataloader:
batch = {k: v.to(self.device) for k, v in batch.items()}
hard_label = batch.pop('labels')
soft_label = []
for k in range(len(batch_logits_locals)):
soft_label.append(batch.pop('logits{}'.format(k)))
outputs = self.model(**batch)
logits = outputs.logits
loss = 0.
for k in range(len(batch_logits_locals)):
tmp_loss = weights[k] * F.cross_entropy(logits, torch.argmax(soft_label[k], dim=-1), reduction='mean')
# tmp_loss = F.cross_entropy(logits, torch.argmax(soft_labels[k], dim=-1), reduction='mean')
loss += tmp_loss
train_loss += loss.item()
# if self.args.num_classes == 1:
# prediction = logits.squeeze()
# else:
# prediction = torch.argmax(logits, dim=-1)
# metric.add_batch(predictions=prediction, references=hard_label)
loss.backward()
optimizer.step()
scheduler.step()
optimizer.zero_grad()
train_loss = train_loss / len(batch_dataloader)
# train_results = metric.compute()
# logging.info('Epoch {}/{}: train_loss={}, train_results={}'.format(epoch + 1, self.args.dis_epochs, train_loss, train_results))
self.model.cpu()
# return train_loss
if __name__ == '__main__':
class args:
num_classes = 2
model_type = 'bert-base-uncased'