-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_fedprox.py
262 lines (203 loc) · 11.8 KB
/
train_fedprox.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
#!/usr/bin/env python
# -*-coding:utf-8 -*-
'''
@File : train_fedprox.py
@Time : 2022/09/16 14:15:42
@Author : Bo
'''
import torch
import numpy as np
import math
import os
import torch.nn as nn
from tqdm import tqdm
import fed_model.cls_model as cls_model
import opt.opt as opt
import copy
import data.create_dataset as create_dataset
import data.prepare_cifar as prepare_cifar
import configs.conf as const
import utils.utils as utils
import shutil
import sys
device=torch.device("cuda")
class Train(object):
def __init__(self, conf, model_dir, data_group, ckpt_dir, local_epoch_update):
self.conf = conf
self.model_dir = model_dir
self.data_group = data_group
self.local_epoch_update = local_epoch_update
self.model_use = cls_model.get_model(conf, device=torch.device("cuda"))
if ckpt_dir:
ckpt_use = torch.load(ckpt_dir, map_location=device)
self.model_use.load_state_dict(ckpt_use)
self.original_model_params = self._get_weight()
self.server_model = copy.deepcopy(self.model_use)
if ckpt_dir:
for k in self.original_model_params.keys():
assert (self.original_model_params[k].data - ckpt_use[k].data).max() == 0
self.optimizer = opt.define_optimizer(conf, self.model_use, "sgd", lr=conf.lr)
self.loss_fn = nn.CrossEntropyLoss(reduction='sum')
self.tr_data_loader, self.val_data_loader, self.tt_data_loader = data_group
print("Training data size", len(self.tr_data_loader))
print("Validating data size", len(self.val_data_loader))
print("Testing data size", len(self.tt_data_loader))
parameter_list = [p for p in self.model_use.parameters() if p.requires_grad == False]
assert len(parameter_list) == 0
self.saved_model_iters = {}
def _get_weight(self):
weight_init = {}
for name, p in self.model_use.named_parameters():
if p.requires_grad:
weight_init[name] = p.data.detach().clone()
return weight_init
def _update_lr(self, global_epoch):
for i in range(len(self.optimizer.param_groups)):
self.optimizer.param_groups[i]["lr"] = self.conf.lr
def _update_batch_tr(self, _image, _label, global_step):
_data_batch = create_dataset.load_data_batch(_image, _label, is_on_cuda=True)
self.optimizer.zero_grad()
_pred = self.model_use(_data_batch["input"])
_loss = self.loss_fn(_pred, _data_batch["target"]) / len(_image)
_loss.backward()
for name, p in self.model_use.named_parameters():
if p.requires_grad:
p.grad.data += (p.data - self.original_model_params[name].data) * self.conf.beta
self.optimizer.step()
accu = (_pred.argmax(axis=-1) == _data_batch["target"]).sum().div(len(_image))
print("Training loss: {:.4f} and Training accuracy {:.2f} and learning rate {:.4f}".format(
_loss.item(), accu.item(), self.optimizer.param_groups[0]["lr"]))
def _eval(self, global_step, data_use, str_use):
self.model_use.eval()
val_loss, val_accu = 0.0, 0.0
for i, (_image, _label) in enumerate(data_use):
_data = create_dataset.load_data_batch(_image, _label, is_on_cuda=True)
_pred = self.model_use(_data["input"])
_loss = self.loss_fn(_pred, _data["target"])
_accu = (_pred.argmax(axis=-1) == _data["target"]).sum()
val_loss += _loss.detach().cpu().numpy()
val_accu += _accu.detach().cpu().numpy()
print("{} loss: {:.4f} and {} accuracy {:.2f}".format(str_use, val_loss / len(data_use) / len(_image),
str_use, val_accu / len(data_use) / len(_image)))
return val_loss / len(data_use) / len(_image), val_accu / len(data_use) / len(_image)
def _select_top_model(self, global_step, val_loss):
num_save=1
if len(self.saved_model_iters.keys()) < num_save:
self._save_model(global_step, val_loss)
self.saved_model_iters["%d" % global_step] = val_loss
self.best_model = self._get_weight()
self.epoch_use = global_step
else:
key_g = list(self.saved_model_iters.keys())
val_accu_g = [self.saved_model_iters[k] for k in key_g]
if val_loss <= np.max(val_accu_g):
self.best_model = self._get_weight()
_ind = np.argmax(val_accu_g)
del self.saved_model_iters[key_g[_ind]]
self.saved_model_iters["%d" % global_step] = val_loss
self._save_model(global_step, val_loss)
self.epoch_use = global_step
ckpt_remove = [v for v in os.listdir(self.model_dir) if "model-{:05d}-{:.4f}.pt".format(int(key_g[_ind]),
np.min(val_accu_g)) in v][0]
if os.path.isfile(self.model_dir + "/" + ckpt_remove):
os.remove(self.model_dir + "/" + ckpt_remove)
print("The best model epoch :", global_step)
def _save_model(self, epoch, val_accu):
torch.save(self.model_use.state_dict(), self.model_dir + "/model-{:05d}-{:.4f}.pt".format(epoch, val_accu))
def run(self):
global_step = 0
for _, epoch in enumerate(tqdm(range(self.local_epoch_update+1)[1:])):
self.model_use.train()
self._update_lr(epoch)
for i, (_image, _label) in enumerate(self.tr_data_loader):
self._update_batch_tr(_image, _label, global_step)
global_step += 1
_val_loss, _val_accu = self._eval(global_step, self.val_data_loader, "validation")
self._select_top_model(epoch, _val_loss)
self._eval(global_step, self.tt_data_loader, "testing")
if np.isnan(_val_loss) == False:
self.epoch_use = self.local_epoch_update
self._save_model(global_step, _val_loss)
print("The final best model is from epoch %d" % epoch)
self.best_model = self._get_weight() # This is the actually best model
all_ckpts = [v for v in os.listdir(self.model_dir) if ".pt" in v and "model" in v]
steps = [int(v.split("-")[1]) for v in all_ckpts]
os.remove(self.model_dir + "/" + all_ckpts[np.argmin(steps)]) # this is for removing the previous saved model
def run_train(conf):
if conf.dataset == "cifar10" or conf.dataset == "cifar100":
train_dataset = prepare_cifar.get_dataset(conf, conf.dataset, conf.image_path, split="train")
test_dataset = prepare_cifar.get_dataset(conf, conf.dataset, conf.image_path, split="test")
print("==========================================")
print("The shape of the training data", len(train_dataset))
print("The shape of the testing data", len(test_dataset))
if conf.dataset == "cifar10" or conf.dataset == "cifar100":
num_im_per_client = 50000 / conf.n_clients
num_im_update = num_im_per_client + conf.num_synthetic_images
local_epoch_update = int(num_im_per_client * conf.local_n_epochs / num_im_update)
conf.updated_local_epoch = local_epoch_update
train_dataset_update, \
val_dataset_update, \
test_dataset = create_dataset.define_val_dataset(conf, train_dataset, test_dataset)
data_group = {"train": train_dataset_update, "val": val_dataset_update, "test": test_dataset}
_, data_partitioner_use = create_dataset.define_data_loader(conf, dataset=data_group["train"],
localdata_id=0,
is_train=True, data_partitioner=None) # this is only for getting a data_partitioner
tr_data_loader, _ = create_dataset.define_data_loader(conf, dataset=data_group["train"],
localdata_id=conf.use_local_id,
is_train=True, data_partitioner=data_partitioner_use)
val_data_loader, _ = create_dataset.define_data_loader(conf, dataset=data_group["val"],
localdata_id=0,
is_train=False)
tt_data_loader, _ = create_dataset.define_data_loader(conf, dataset=data_group["test"],
localdata_id=0,
is_train=False)
print("The length of the tr_loader ", len(tr_data_loader), " val loader ", len(val_data_loader), " test loader ", len(tt_data_loader),
)
data_loader_group = [tr_data_loader, val_data_loader, tt_data_loader]
if conf.communication_round > 0:
ckpt_dir = conf.model_dir.split("client_id")[0].replace("communication_round_%03d" % conf.communication_round,
"communication_round_%03d" % (conf.communication_round - 1)) + "aggregated_model.pt"
if conf.communication_round > 2:
m_dir_delete = conf.model_dir.split("client_id")[0].replace("communication_round_%03d" % (conf.communication_round),
"communication_round_%03d" % (conf.communication_round - 2))
if conf.free_up_space == True and conf.sample_ratio == 1.0:
if conf.use_local_id == 0:
c_group = [v for v in os.listdir(m_dir_delete) if "client_id" in v]
for v in c_group:
shutil.rmtree(m_dir_delete + "/" + v + "/")
else:
ckpt_dir = None
Train(conf, conf.model_dir, data_loader_group, ckpt_dir, local_epoch_update).run()
def train_with_conf(conf):
path_init = utils.get_path_init(conf.loc)
conf = utils.get_dir_name(conf)
model_dir = path_init + "/%s/%s/communication_round_%03d/" % (conf.folder_name,
conf.dir_name,
conf.communication_round)
selected_clients = utils.select_random_clients(conf.n_clients, conf.sample_ratio, model_dir)
conf.selected_clients = selected_clients
print("The selected clients", selected_clients)
print("Local id", conf.use_local_id)
model_dir += "client_id_%02d/" % selected_clients[conf.use_local_id]
conf.use_local_id = selected_clients[conf.use_local_id]
# utils.create_dir(model_dir)
utils.create_dir(model_dir.split("/communication_round")[0].replace(path_init, "../exp_data/"))
conf.model_dir = model_dir
stdoutOrigin = sys.stdout
sys.stdout = open(conf.model_dir + "training_statistics.txt", 'w')
if conf.communication_round == 0:
conf.random_state = np.random.RandomState(conf.seed_use)
utils.seed_everything(conf.seed_use)
else:
conf.seed_use = np.random.randint(1, 20000, [1])[0]
conf.random_state = np.random.RandomState(conf.seed_use)
utils.seed_everything(conf.seed_use)
for arg in vars(conf):
print(arg, getattr(conf, arg))
run_train(conf)
sys.stdout.close()
sys.stdout = stdoutOrigin
if __name__ == "__main__":
a = torch.zeros([1]).to(device)
conf = const.get_args()
train_with_conf(conf)