forked from RS2002/CSI-BERT
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathfinetune.py
172 lines (155 loc) · 7.35 KB
/
finetune.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
from model import CSIBERT,Sequence_Classifier,Classification
from transformers import BertConfig
import argparse
import tqdm
import torch
from dataset import load_all,load_zero_people
from torch.utils.data import DataLoader
import torch.nn as nn
import copy
import numpy as np
from sklearn.model_selection import train_test_split
pad=np.array([-1000]*52)
def get_args():
parser = argparse.ArgumentParser(description='')
parser.add_argument('--batch_size', type=int, default=64)
parser.add_argument('--mask_percent', type=float, default=0.15)
parser.add_argument('--normal', action="store_true", default=False)
parser.add_argument('--hs', type=int, default=64)
parser.add_argument('--layers', type=int, default=4)
parser.add_argument('--max_len', type=int, default=100)
parser.add_argument('--heads', type=int, default=4)
parser.add_argument('--position_embedding_type', type=str, default="absolute")
parser.add_argument('--time_embedding', action="store_true", default=False) # whether to use time embedding
parser.add_argument("--cpu", action="store_true",default=False)
parser.add_argument("--cuda", type=str, default='0')
parser.add_argument("--carrier_dim", type=int, default=52)
parser.add_argument("--carrier_attn", action="store_true",default=False)
parser.add_argument('--lr', type=float, default=0.0005)
# parser.add_argument("--test_people", type=int, nargs='+', default=[0,1])
parser.add_argument('--epoch', type=int, default=50)
parser.add_argument('--class_num', type=int, default=6) #action:6, people:8
parser.add_argument('--task', type=str, default="action") # "action" or "people"
parser.add_argument("--path", type=str, default='./csibert_pretrain.pth')
args = parser.parse_args()
return args
def main():
args=get_args()
device_name = "cuda:"+args.cuda
device = torch.device(device_name if torch.cuda.is_available() and not args.cpu else 'cpu')
bertconfig=BertConfig(max_position_embeddings=args.max_len, hidden_size=args.hs, position_embedding_type=args.position_embedding_type,num_hidden_layers=args.layers,num_attention_heads=args.heads)
csibert=CSIBERT(bertconfig,args.carrier_dim,args.carrier_attn, args.time_embedding)
csibert.load_state_dict(torch.load(args.path))
csi_dim=args.carrier_dim
# model=Sequence_Classifier(csibert,args.class_num)
model = Classification(csibert, args.class_num)
model=model.to(device)
total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print('total parameters:', total_params)
optim = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=0.01)
dataset=load_all()
train_data,test_data=train_test_split(dataset, test_size=0.1)
train_loader = DataLoader(train_data, batch_size=args.batch_size, shuffle=True)
test_loader = DataLoader(test_data, batch_size=args.batch_size, shuffle=False)
loss_func = nn.CrossEntropyLoss()
best_loss=None
for j in range(args.epoch):
model.train()
torch.set_grad_enabled(True)
loss_list=[]
acc_list=[]
pbar = tqdm.tqdm(train_loader, disable=False)
for x,_,action,people,timestamp in pbar:
x=x.to(device)
timestamp=timestamp.to(device)
if args.task=="action":
label=action.to(device)
elif args.task=="people":
label=people.to(device)
else:
print("ERROR")
exit(-1)
input = copy.deepcopy(x)
max_values, _ = torch.max(input, dim=-2, keepdim=True)
input[input == pad[0]] = -pad[0]
min_values, _ = torch.min(input, dim=-2, keepdim=True)
input[input == -pad[0]] = pad[0]
if args.normal: # 在时间维度归一化
non_pad = (input != pad[0]).float()
avg = copy.deepcopy(input)
avg[input == pad[0]] = 0
avg = torch.sum(avg, dim=-2, keepdim=True) / torch.sum(non_pad, dim=-2, keepdim=True)
std = (input - avg) ** 2
std[input == pad[0]] = 0
std = torch.sum(std, dim=-2, keepdim=True) / torch.sum(non_pad, dim=-2, keepdim=True)
std = torch.sqrt(std)
input = (input - avg) / std
batch_size,seq_len,carrier_num=input.shape
attn_mask = (x[:, :, 0] != pad[0]).float().to(device) # (batch, seq_len)
if args.position_embedding_type=="absolute":
y = model(input, attn_mask)
else:
y = model(input, attn_mask, timestamp)
loss = loss_func(y,label)
output = torch.argmax(y, dim=-1)
acc=torch.sum(output==label)/batch_size
model.zero_grad()
loss.backward()
# nn.utils.clip_grad_norm_(model.parameters(), 3.0) # 用于裁剪梯度，防止梯度爆炸
optim.step()
loss_list.append(loss.item())
acc_list.append(acc.item())
log="Epoch {} | Train Loss {:06f}, Train Acc {:06f}, ".format(j+1,np.mean(loss_list),np.mean(acc_list))
print(log)
with open("Finetune.txt", 'a') as file:
file.write(log)
model.eval()
torch.set_grad_enabled(False)
loss_list=[]
acc_list=[]
pbar = tqdm.tqdm(test_loader, disable=False)
for x,_,action,people,timestamp in pbar:
x=x.to(device)
timestamp=timestamp.to(device)
if args.task=="action":
label=action.to(device)
elif args.task=="people":
label=people.to(device)
else:
print("ERROR")
exit(-1)
input = copy.deepcopy(x)
max_values, _ = torch.max(input, dim=-2, keepdim=True)
input[input == pad[0]] = -pad[0]
min_values, _ = torch.min(input, dim=-2, keepdim=True)
input[input == -pad[0]] = pad[0]
if args.normal: # 在时间维度归一化
non_pad = (input != pad[0]).float()
avg = copy.deepcopy(input)
avg[input == pad[0]] = 0
avg = torch.sum(avg, dim=-2, keepdim=True) / torch.sum(non_pad, dim=-2, keepdim=True)
std = (input - avg) ** 2
std[input == pad[0]] = 0
std = torch.sum(std, dim=-2, keepdim=True) / torch.sum(non_pad, dim=-2, keepdim=True)
std = torch.sqrt(std)
input = (input - avg) / std
batch_size,seq_len,carrier_num=input.shape
attn_mask = (x[:, :, 0] != pad[0]).float().to(device) # (batch, seq_len)
if args.position_embedding_type=="absolute":
y = model(input, attn_mask)
else:
y = model(input, attn_mask, timestamp)
loss = loss_func(y,label)
output = torch.argmax(y, dim=-1)
acc=torch.sum(output==label)/batch_size
loss_list.append(loss.item())
acc_list.append(acc.item())
log="Test Loss {:06f}, Test Acc {:06f}, ".format(np.mean(loss_list),np.mean(acc_list))
print(log)
with open("Finetune.txt", 'a') as file:
file.write(log+"\n")
if best_loss is None or np.mean(loss_list)<best_loss:
best_loss=np.mean(loss_list)
torch.save(model.state_dict(), "finetune.pth")
if __name__ == '__main__':
main()