forked from hojunroks/ETRI2022
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_transformer.py
173 lines (146 loc) · 8.78 KB
/
train_transformer.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
from model import LSTM, TransformerBased, generate_square_subsequent_mask
import matplotlib.pyplot as plt
import torch
import torch.nn as nn
import numpy as np
import os
from data import load_data, load_data_lstm, load_data_transformer
from argparse import ArgumentParser
from utils import evaluate, fix_seed
from torch.utils.tensorboard import SummaryWriter
import datetime
import pdb
PERSON_DIRS = [0,6,10,12,20,25,30]
def main():
########## PARSE ARUGMENTS ###########
parser = ArgumentParser()
parser.add_argument('--model_name', type=str, default="transformer")
parser.add_argument('--data_dir', type=str, default="/data/etri_lifelog")
parser.add_argument('--act_flag', type=str, default="act")
parser.add_argument('--person_index', type=int, default=28)
parser.add_argument('--seed', type=int, default=41)
parser.add_argument('--num_epochs', type=int, default=300)
parser.add_argument('--lr', type=float, default=0.00001)
parser.add_argument('--weight_decay', type=float, default=0.01)
parser.add_argument('--d_hid', type=int, default=16)
parser.add_argument('--d_model', type=int, default=16)
parser.add_argument('--nhead', type=int, default=4)
parser.add_argument('--nlayers', type=int, default=3)
parser.add_argument('--bidirectional', type=bool, default=False)
parser.add_argument('--test_every', type=int, default=5)
parser.add_argument('--dropout', type=float, default=0.4)
parser.add_argument('--split_ratio', type=float, default=0.67)
parser.add_argument('--use_timestamp', type=bool, default=False)
parser.add_argument('--sequence_length', type=int, default=10)
parser.add_argument('--sche', action='store_true')
parser.add_argument('--gamma', type=float, default=0.99)
args = parser.parse_args()
############## LOAD DATA ##############
fix_seed(args.seed)
person_dir_index = 0
while person_dir_index < len(PERSON_DIRS) and PERSON_DIRS[person_dir_index] < args.person_index:
person_dir_index += 1
data_path="user{0:02d}-{1:02d}/user{2:02d}".format(PERSON_DIRS[person_dir_index-1]+1, PERSON_DIRS[person_dir_index], args.person_index)
print(data_path)
data_path = os.path.join(args.data_dir, data_path)
train_feat, train_label, train_label_emotion, test_feat, test_label, test_label_emotion, num_classes = load_data_transformer(data_path, split_ratio=args.split_ratio, act_flag=args.act_flag, seq_len=args.sequence_length)
with torch.cuda.device(0):
train_feat = train_feat.cuda()
test_feat = test_feat.cuda()
label_act_train = train_label.cuda()
label_act_test = test_label.cuda()
label_emotion_train = train_label_emotion.cuda()
label_emotion_test = test_label_emotion.cuda()
########### INITIALIZE MODEL ###########
input_size = train_feat.shape[-1]
model = TransformerBased(input_size, args.d_model, args.nhead, args.d_hid, args.nlayers, num_classes, args.dropout).to('cuda:0')
criterion = nn.CrossEntropyLoss()
criterion_emotion = nn.HuberLoss() # L1Loss or HuberLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay = args.weight_decay)
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=args.gamma)
now = datetime.datetime.now()
nowDate = now.strftime('%Y-%m-%d-%H:%M:%S')
if args.sche == True:
exp_name = 'USER_%s_c%s_dh%s_dm%s_he%s_l%s_lr%s_wd%s_dr%s_g%s_s%s_es%s_fin_mep300'%(args.person_index, num_classes, args.d_hid, args.d_model, args.nhead, args.nlayers, args.lr, args.weight_decay, args.dropout, args.gamma, args.sequence_length, args.seed)
else:
exp_name = 'USER_%s_c%s_dh%s_dm%s_he%s_l%s_lr%s_wd%s_dr%s_nosche_s%s_es%s_fin_mep300'%(args.person_index, num_classes, args.d_hid, args.d_model, args.nhead, args.nlayers, args.lr, args.weight_decay, args.dropout, args.sequence_length, args.seed)
writer = SummaryWriter('runs/'+exp_name)
################ TRAIN #################
best_accuracy = 0.
for epoch in range(args.num_epochs):
model.train()
src_mask = generate_square_subsequent_mask(args.sequence_length).to('cuda:0')
loss_act_total = []
loss_emo_total = []
act_accu_total = []
emo_accu_total = []
for i in range(len(train_feat)):
train_batch = train_feat[i].unsqueeze(dim=1)
outputs = model(train_batch, src_mask)
out_act, out_emotion = outputs
optimizer.zero_grad()
loss_act = criterion(out_act, label_act_train[i].type(torch.LongTensor).to('cuda:0')) #trainY one-hot index
loss_act_total.append(loss_act.item())
loss_emo = criterion_emotion(out_emotion.squeeze(), label_emotion_train[i])
loss_emo_total.append(loss_emo.item())
(loss_act + loss_emo).backward() # TODO: hyperparams?
optimizer.step()
act_accu, emo_accu = evaluate(model, train_feat[i].unsqueeze(dim=1), label_act_train[i], label_emotion_train[i], args.sequence_length, args.model_name)
act_accu_total.append(act_accu)
emo_accu_total.append(emo_accu)
loss_act_total = np.array(loss_act_total)
loss_emo_total = np.array(loss_emo_total)
act_accu_total = np.mean(np.array(act_accu_total), axis=0)
emo_accu_total = np.mean(np.array(emo_accu_total), axis=0)
writer.add_scalar("Loss/train/emo", loss_emo_total.mean(), epoch)
writer.add_scalar("Loss/train/act", loss_act_total.mean(), epoch)
writer.add_scalar("Accuracy/train/act-top-5", act_accu_total[2], epoch)
writer.add_scalar("Accuracy/train/act-top-3", act_accu_total[1], epoch)
writer.add_scalar("Accuracy/train/act-top-1", act_accu_total[0], epoch)
writer.add_scalar("Accuracy/train/emo-top-1", emo_accu_total[0], epoch)
if args.sche == True:
scheduler.step()
if epoch % args.test_every == 0:
with torch.no_grad():
src_mask = generate_square_subsequent_mask(args.sequence_length).to('cuda:0')
model.eval()
loss_act_test = []
loss_emo_test = []
act_accu_test = []
emo_accu_test = []
for i in range(len(test_feat)): # [#samples, #n_classes, #feat_dim]
test_pred = model(test_feat[i].unsqueeze(dim=1), src_mask)
pred_act, pred_emotion = test_pred # pred_act: [#seq_len, #n_class]
loss_act_test.append(criterion(pred_act, label_act_test[i].type(torch.LongTensor).to('cuda:0')).item())
loss_emo_test.append(criterion_emotion(pred_emotion.squeeze(), label_emotion_test[i]).item())
act_accu, emo_accu = evaluate(model, test_feat[i].unsqueeze(dim=1), label_act_test[i], label_emotion_test[i], args.sequence_length, args.model_name)
act_accu_test.append(act_accu)
emo_accu_test.append(emo_accu)
act_accu_test = np.mean(np.array(act_accu_test), axis=0)
emo_accu_test = np.mean(np.array(emo_accu_test), axis=0)
loss_act_test = np.array(loss_act_test)
loss_emo_test = np.array(loss_emo_test)
writer.add_scalar("Loss/test/emo", loss_emo_test.mean(), epoch)
writer.add_scalar("Loss/test/act", loss_act_test.mean(), epoch)
writer.add_scalar("Accuracy/test/act-top-5", act_accu_test[2], epoch)
writer.add_scalar("Accuracy/test/act-top-3", act_accu_test[1], epoch)
writer.add_scalar("Accuracy/test/act-top-1", act_accu_test[0], epoch)
writer.add_scalar("Accuracy/test/emo-top-1", emo_accu_test[0], epoch)
writer.add_scalar("lr", optimizer.param_groups[0]['lr'], epoch)
if best_accuracy <= act_accu_test[0]: # based on act top-1 accuracy
best_accuracy = act_accu_test[0]
best_accus = np.concatenate((act_accu_test, emo_accu_test), axis=-1)
save_dict = {
"epoch": epoch,
"model_state_dict": model.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
"accuracy": best_accuracy,
}
save_path = "./wgt/" + exp_name
if os.path.exists(save_path) == False:
os.mkdir(save_path)
save_model_name = os.path.join(save_path, 'best.ckpt')
torch.save(save_dict, save_model_name)
np.save(os.path.join(save_path, "best_accu.npy"), best_accus)
if __name__=='__main__':
main()