-
Notifications
You must be signed in to change notification settings - Fork 2
/
decomfl_main.py
362 lines (338 loc) · 13.7 KB
/
decomfl_main.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
import functools
from os import path
import torch
import torch.nn as nn
from peft import LoraConfig, get_peft_model
from tensorboardX import SummaryWriter
from tqdm import tqdm
from transformers import AutoModelForCausalLM, AutoTokenizer
from byzantine import aggregation as byz_agg
from byzantine import attack as byz_attack
from cezo_fl.client import ResetClient
from cezo_fl.fl_helpers import get_client_name
from cezo_fl.models.cnn_fashion import CNN_FMNIST
from cezo_fl.models.cnn_mnist import CNN_MNIST
from cezo_fl.models.lenet import LeNet
from cezo_fl.models.lstm import CharLSTM
from cezo_fl.random_gradient_estimator import RandomGradientEstimator as RGE
from cezo_fl.server import CeZO_Server
from cezo_fl.util import model_helpers
from cezo_fl.util.language_utils import LM_TEMPLATE_MAP, SUPPORTED_LLM, get_lm_loss
from cezo_fl.util.metrics import accuracy
from config import get_args_str, get_params
from preprocess import preprocess
def prepare_settings_underseed(args, device, server_or_client: str = "server"):
torch_dtype = {
"float32": torch.float32,
"float16": torch.float16,
"bfloat16": torch.bfloat16,
}[args.model_dtype]
torch.manual_seed(args.seed)
if args.dataset == "mnist":
model = CNN_MNIST().to(torch_dtype).to(device)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(
model_helpers.get_trainable_model_parameters(model),
lr=args.lr,
weight_decay=1e-5,
momentum=args.momentum,
)
accuracy_func = accuracy
# scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.8)
elif args.dataset == "cifar10":
model = LeNet().to(torch_dtype).to(device)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(
model_helpers.get_trainable_model_parameters(model),
lr=args.lr,
weight_decay=5e-4,
momentum=args.momentum,
)
accuracy_func = accuracy
# scheduler = torch.optim.lr_scheduler.MultiStepLR(
# optimizer, milestones=[200], gamma=0.1
# )
elif args.dataset == "fashion":
model = CNN_FMNIST().to(torch_dtype).to(device)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(
model_helpers.get_trainable_model_parameters(model),
lr=args.lr,
weight_decay=1e-5,
momentum=args.momentum,
)
accuracy_func = accuracy
# scheduler = torch.optim.lr_scheduler.MultiStepLR(
# optimizer, milestones=[200], gamma=0.1
# )
elif args.dataset == "shakespeare":
model = CharLSTM().to(torch_dtype).to(device)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(
model_helpers.get_trainable_model_parameters(model),
lr=args.lr,
momentum=0.9,
weight_decay=5e-4,
)
accuracy_func = accuracy
# scheduler = torch.optim.lr_scheduler.MultiStepLR(
# optimizer, milestones=[200], gamma=0.1
# )
elif args.dataset in LM_TEMPLATE_MAP.keys():
large_model = args.large_model
model_name = SUPPORTED_LLM[large_model]
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch_dtype).to(device)
model.model_name = large_model
tokenizer = AutoTokenizer.from_pretrained(
model_name, padding_side="left", truncate_side="left"
)
template = LM_TEMPLATE_MAP[args.dataset]()
if args.dataset in ["sst2", "cb", "wsc", "wic", "multirc", "rte", "boolq"]:
if args.lora:
# this step initialize lora parameters, which should be under control of seed
lora_config = LoraConfig(
r=args.lora_r,
lora_alpha=args.lora_alpha,
target_modules=["q_proj", "v_proj"],
)
model = get_peft_model(model, lora_config).to(torch_dtype)
verbalizer_id_map = template.get_verbalizer_id(tokenizer)
criterion = get_lm_loss("last_token", verbalizer_id_map=verbalizer_id_map)
optimizer = torch.optim.SGD(
model_helpers.get_trainable_model_parameters(model),
lr=args.lr,
momentum=0,
weight_decay=5e-4,
)
accuracy_func = get_lm_loss("accuracy", verbalizer_id_map=verbalizer_id_map)
elif args.dataset in ["squad", "drop", "xsum"]:
if server_or_client == "server":
criterion = get_lm_loss("f1", tokenizer=tokenizer)
optimizer = torch.optim.SGD(
model_helpers.get_trainable_model_parameters(model),
lr=args.lr,
momentum=0,
weight_decay=0,
)
accuracy_func = get_lm_loss("f1", tokenizer=tokenizer)
elif server_or_client == "client":
criterion = get_lm_loss("full_sentence", verbalizer_id_map={})
optimizer = torch.optim.SGD(
model_helpers.get_trainable_model_parameters(model),
lr=args.lr,
momentum=0,
weight_decay=0,
)
accuracy_func = get_lm_loss("full_sentence", verbalizer_id_map={})
else:
raise ValueError(
"server_or_client must be either 'server' or 'client'. "
f"But get {server_or_client}"
)
else:
raise ValueError(f"Dataset {args.dataset} is not supported")
else:
raise Exception(f"Dataset {args.dataset} is not supported")
if args.grad_estimate_method in ["rge-central", "rge-forward"]:
method = args.grad_estimate_method[4:]
print(f"Using RGE {method}")
if args.dataset in ["squad", "drop"] and server_or_client == "server":
generation_mode = True
# TODO move this setting partially to the args
generation_mode_kwargs = {
"do_sample": True,
"temperature": 1.0,
"num_beams": 2,
"top_p": 0.3,
"top_k": None,
"num_return_sequences": 1,
"max_new_tokens": 5, # will be adjusted dynamically later
"max_length": 2048,
"length_penalty": 2,
"early_stopping": True,
"eos_token_id": [
tokenizer.encode("\n", add_special_tokens=False)[-1],
tokenizer.eos_token_id,
],
}
elif args.dataset in ["xsum"] and server_or_client == "server":
generation_mode = True
# TODO move this setting partially to the args
generation_mode_kwargs = {
"do_sample": True,
"temperature": 1.0,
"num_beams": 2,
"top_p": 0.95,
"top_k": None,
"num_return_sequences": 1,
"max_new_tokens": 500, # will be adjusted dynamically later
"max_length": 2048,
"early_stopping": True,
"eos_token_id": [
tokenizer.encode("\n", add_special_tokens=False)[-1],
tokenizer.eos_token_id,
],
}
else:
generation_mode = False
generation_mode_kwargs = None
grad_estimator = RGE(
model,
parameters=model_helpers.get_trainable_model_parameters(model),
mu=args.mu,
num_pert=args.num_pert,
grad_estimate_method=method,
device=device,
torch_dtype=torch_dtype,
# To save memory consumption, we have to use parameter-wise perturb + no_optim together.
sgd_only_no_optim=args.no_optim,
paramwise_perturb=args.no_optim,
# For generation mode, the forward style is different
generation_mode=generation_mode,
generation_mode_kwargs=generation_mode_kwargs,
)
else:
raise Exception(f"Grad estimate method {args.grad_estimate_method} not supported")
return model, criterion, optimizer, grad_estimator, accuracy_func
def setup_server_and_clients(
args, device_map: dict[str, torch.device], train_loaders
) -> CeZO_Server:
clients = []
for i in range(args.num_clients):
client_name = get_client_name(i)
client_device = device_map[client_name]
(
client_model,
client_criterion,
client_optimizer,
client_grad_estimator,
client_accuracy_func,
) = prepare_settings_underseed(args, client_device, "client")
client_model.to(client_device)
client = ResetClient(
client_model,
train_loaders[i],
client_grad_estimator,
client_optimizer,
client_criterion,
client_accuracy_func,
client_device,
)
clients.append(client)
server_device = device_map["server"]
server = CeZO_Server(
clients,
server_device,
num_sample_clients=args.num_sample_clients,
local_update_steps=args.local_update_steps,
)
# set server tools
(
server_model,
server_criterion,
server_optimizer,
server_grad_estimator,
server_accuracy_func,
) = prepare_settings_underseed(args, server_device, "server")
server_model.to(server_device)
server.set_server_model_and_criterion(
server_model,
server_criterion,
server_accuracy_func,
server_optimizer,
server_grad_estimator,
)
# TODO(lizhe) move this into a seperate main file.
# Prepare the Byzantine attack
if args.byz_type == "no_byz":
server.register_attack_func(byz_attack.no_byz)
elif args.byz_type == "gaussian":
server.register_attack_func(
functools.partial(byz_attack.gaussian_attack, num_attack=args.num_byz)
)
elif args.byz_type == "sign":
server.register_attack_func(
functools.partial(byz_attack.sign_attack, num_attack=args.num_byz)
)
elif args.byz_type == "trim":
server.register_attack_func(
functools.partial(byz_attack.trim_attack, num_attack=args.num_byz)
)
elif args.byz_type == "krum":
server.register_attack_func(
functools.partial(byz_attack.krum_attack, num_attack=args.num_byz, lr=args.lr)
)
else:
raise Exception(
"byz_type should be one of no_byz, gaussian, sign, trim, krum."
+ f"But get {args.byz_type}"
)
if args.aggregation == "mean":
server.register_aggregation_func(byz_agg.mean)
elif args.aggregation == "median":
server.register_aggregation_func(byz_agg.median)
elif args.aggregation == "trim":
server.register_aggregation_func(byz_agg.trim)
elif args.aggregation == "krum":
server.register_aggregation_func(byz_agg.krum)
else:
raise Exception(
"aggregation type should be one of mean, median, trim, krum. "
+ f"But get {args.aggregation}"
)
return server
# get_warmup_lr is not used for now.
def get_warmup_lr(args, current_epoch: int, current_iter: int, iters_per_epoch: int) -> float:
overall_iterations = args.warmup_epochs * iters_per_epoch + 1
current_iterations = current_epoch * iters_per_epoch + current_iter + 1
return args.lr * current_iterations / overall_iterations
def get_size_of_model(model):
return sum(p.numel() * p.element_size() for p in model.parameters())
if __name__ == "__main__":
args = get_params().parse_args()
if args.dataset == "shakespeare":
args.num_clients = 139
print(args)
device_map, train_loaders, test_loader = preprocess(args)
server = setup_server_and_clients(args, device_map, train_loaders)
if args.log_to_tensorboard:
tensorboard_sub_folder = "-".join(
[
get_args_str(args),
server.server_model.model_name,
model_helpers.get_current_datetime_str(),
]
)
writer = SummaryWriter(
path.join(
"tensorboards",
"cezo_fl",
args.dataset,
args.log_to_tensorboard,
tensorboard_sub_folder,
)
)
with tqdm(total=args.iterations, desc="Training:") as t, torch.no_grad():
for ite in range(args.iterations):
step_loss, step_accuracy = server.train_one_step(ite)
t.set_postfix({"Loss": step_loss, "Accuracy": step_accuracy})
t.update(1)
if args.adjust_perturb:
if ite == 500:
server.set_learning_rate(args.lr * 0.8)
server.set_perturbation(args.num_pert * 2)
elif ite == 1000:
server.set_learning_rate(args.lr * 0.5)
server.set_perturbation(args.num_pert * 4)
elif ite == 2000:
server.set_learning_rate(args.lr * 0.3)
server.set_perturbation(args.num_pert * 8)
if args.log_to_tensorboard:
writer.add_scalar("Loss/train", step_loss, ite)
writer.add_scalar("Accuracy/train", step_accuracy, ite)
# eval
if args.eval_iterations != 0 and (ite + 1) % args.eval_iterations == 0:
eval_loss, eval_accuracy = server.eval_model(test_loader)
if args.log_to_tensorboard:
writer.add_scalar("Loss/test", eval_loss, ite)
writer.add_scalar("Accuracy/test", eval_accuracy, ite)