-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathmain_lamb.py
174 lines (150 loc) · 8.17 KB
/
main_lamb.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
import os
import copy
import torch
import socket
import datetime
from torch import nn
from torch.utils.tensorboard import SummaryWriter
from torch.optim import SGD
from torch.optim.lr_scheduler import MultiStepLR
from datasets import load_dataset
from networks import load_model
from workers.worker_vision import *
from utils.scheduler import Warmup_MultiStepLR
from utils.utils import *
from lamb.pytorch_lamb import Lamb, log_lamb_rs
os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
dir_path = os.path.dirname(__file__)
nfs_dataset_path1 = '/mnt/nfs4-p1/ckx/datasets/'
nfs_dataset_path2 = '/nfs4-p1/ckx/datasets/'
# torch.set_num_threads(4)
def main(args):
set_seed(args)
# check nfs dataset path
if os.path.exists(nfs_dataset_path1):
args.dataset_path = nfs_dataset_path1
elif os.path.exists(nfs_dataset_path2):
args.dataset_path = nfs_dataset_path2
log_id = datetime.datetime.now().strftime('%b%d_%H:%M:%S') + '_' + socket.gethostname() + '_' + args.identity
writer = SummaryWriter(log_dir=os.path.join(args.runs_data_dir, log_id))
probe_train_loader, probe_valid_loader, _, classes = load_dataset(root=args.dataset_path, name=args.dataset_name, image_size=args.image_size,
train_batch_size=256, valid_batch_size=64)
worker_list = []
split = [1.0 / args.size for _ in range(args.size)]
for rank in range(args.size):
train_loader, _, _, classes = load_dataset(root=args.dataset_path, name=args.dataset_name, image_size=args.image_size,
train_batch_size=args.batch_size,
distribute=True, rank=rank, split=split, seed=args.seed)
model = load_model(args.model, classes, pretrained=args.pretrained).to(args.device)
optimizer = Lamb(model.parameters(), lr=args.lr, weight_decay=args.wd, betas=(.9, .999), adam=('lamb' == 'adam'))
scheduler = MultiStepLR(optimizer, milestones=[max(args.epoch, args.early_stop), max(args.epoch, args.early_stop)], gamma=args.gamma)
if args.amp:
worker = Worker_Vision_AMP(model, rank, optimizer, scheduler, train_loader, args.device)
else:
worker = Worker_Vision(model, rank, optimizer, scheduler, train_loader, args.device)
worker_list.append(worker)
# 定义 中心模型 center_model
center_model = copy.deepcopy(worker_list[0].model)
# center_model = copy.deepcopy(worker_list[0].model)
for name, param in center_model.named_parameters():
for worker in worker_list[1:]:
param.data += worker.model.state_dict()[name].data
param.data /= args.size
P = generate_P(args.mode, args.size)
iteration = 0
for epoch in range(args.epoch):
for worker in worker_list:
worker.update_iter()
for _ in range(train_loader.__len__()):
if args.mode == 'csgd':
for worker in worker_list:
worker.model.load_state_dict(center_model.state_dict())
worker.step()
worker.update_grad()
else: # dsgd
# 每个iteration，传播矩阵P中的worker做random shuffle（自己的邻居在下一个iteration时改变）
if args.shuffle == "random":
P_perturbed = np.matmul(np.matmul(PermutationMatrix(args.size).T,P),PermutationMatrix(args.size))
elif args.shuffle == "fixed":
P_perturbed = P
model_dict_list = []
for worker in worker_list:
model_dict_list.append(worker.model.state_dict())
for worker in worker_list:
worker.step()
for name, param in worker.model.named_parameters():
param.data = torch.zeros_like(param.data)
for i in range(args.size):
p = P_perturbed[worker.rank][i]
param.data += model_dict_list[i][name].data * p
# worker.step() # 效果会变差
worker.update_grad()
center_model = copy.deepcopy(worker_list[0].model)
for name, param in center_model.named_parameters():
for worker in worker_list[1:]:
param.data += worker.model.state_dict()[name].data
param.data /= args.size
if iteration % 50 == 0:
start_time = datetime.datetime.now()
eval_iteration = iteration
if args.amp:
train_acc, train_loss, valid_acc, valid_loss = eval_vision_amp(center_model, probe_train_loader, probe_valid_loader,
None, iteration, writer, args.device)
else:
train_acc, train_loss, valid_acc, valid_loss = eval_vision(center_model, probe_train_loader, probe_valid_loader,
None, iteration, writer, args.device)
print(f"\n|\033[0;31m Iteration:{iteration}|{args.early_stop}, epoch: {epoch}|{args.epoch},\033[0m",
f'train loss:{train_loss:.4}, acc:{train_acc:.4%}, '
f'valid loss:{valid_loss:.4}, acc:{valid_acc:.4%}.',
flush=True, end="\n")
else:
end_time = datetime.datetime.now()
print(f"\r|\033[0;31m Iteration:{eval_iteration}-{iteration}, time: {(end_time - start_time).seconds}s\033[0m", flush=True, end="")
iteration += 1
if iteration == args.early_stop: break
if iteration == args.early_stop: break
state = {
'acc': train_acc,
'epoch': epoch,
'state_dict': center_model.state_dict()
}
if not os.path.exists(args.perf_dict_dir):
os.mkdir(args.perf_dict_dir)
torch.save(state, os.path.join(args.perf_dict_dir, log_id + '.t7'))
writer.close()
print('ending')
if __name__=='__main__':
parser = argparse.ArgumentParser()
## dataset
parser.add_argument("--dataset_path", type=str, default='datasets')
parser.add_argument("--dataset_name", type=str, default='CIFAR10',
choices=['CIFAR10','CIFAR100','TinyImageNet'])
parser.add_argument("--image_size", type=int, default=56, help='input image size')
parser.add_argument("--batch_size", type=int, default=512)
parser.add_argument('--n_swap', type=int, default=None)
# mode parameter
parser.add_argument('--mode', type=str, default='ring', choices=['csgd', 'ring', 'meshgrid', 'exponential'])
parser.add_argument('--shuffle', type=str, default="fixed", choices=['fixed', 'random'])
parser.add_argument('--size', type=int, default=16)
parser.add_argument('--port', type=int, default=29500)
parser.add_argument('--backend', type=str, default="gloo")
# deep model parameter
parser.add_argument('--model', type=str, default='ResNet18',
choices=['ResNet18', 'AlexNet', 'DenseNet121', 'AlexNet_M','ResNet18_M', 'ResNet34_M', 'DenseNet121_M'])
parser.add_argument("--pretrained", type=int, default=1)
# optimization parameter
parser.add_argument('--lr', type=float, default=0.1, help='learning rate')
parser.add_argument('--wd', type=float, default=0.0, help='weight decay')
parser.add_argument('--gamma', type=float, default=0.1)
parser.add_argument('--momentum', type=float, default=0.0)
parser.add_argument('--warmup_step', type=int, default=0)
parser.add_argument('--epoch', type=int, default=6000)
parser.add_argument('--early_stop', type=int, default=6000, help='w.r.t., iterations')
parser.add_argument('--milestones', type=int, nargs='+', default=[2400, 4800])
parser.add_argument('--seed', type=int, default=666)
parser.add_argument("--device", type=int, default=0)
parser.add_argument("--amp", action='store_true', help='automatic mixed precision')
args = parser.parse_args()
args = add_identity(args, dir_path)
# print(args)
main(args)