-
Notifications
You must be signed in to change notification settings - Fork 6
/
main_dg.py
159 lines (133 loc) · 8.07 KB
/
main_dg.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
# encoding=utf-8
"""
Created on 2018/12/10
@author: Hangwei Qian
"""
import matplotlib
matplotlib.use('Agg')
import network_dg as net
import data_preprocess_dg
import torch
import torch.nn as nn
import tqdm
import argparse
from utils import *
import os
DEVICE = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
result = []
acc_all = []
LOSS_FN_WEIGHT = 1e-5
def train_dg_fixed(model, optimizer, train_loader, test_loader, now_model_name, args):
feature_dim = args.n_feature # for dg dataset
n_batch = len(train_loader.dataset) // args.batch_size
criterion = nn.CrossEntropyLoss()
criterion_ae = nn.MSELoss()
for e in range(args.n_epoch):
if e>0 and e%50 == 0:
plot(result_name)
model.train()
correct, total_loss = 0, 0
total = 0
for index, (sample, target) in enumerate(train_loader):
sample, target = sample.to(DEVICE).float(), target.to(DEVICE).long()
now_len = sample.shape[1]
sample = sample.view(-1, feature_dim, now_len)
sample, target = sample.to(DEVICE).float(), target.to(DEVICE).long()
output, out_decoder = model(sample)
loss_classify = criterion(output, target)
loss_ae = criterion_ae(sample.view(sample.size(0), -1), out_decoder)
loss_mmd = mmd_custorm(sample.view(sample.size(0), -1), out_decoder, [args.sigma])
loss_mmd = loss_mmd.to(DEVICE).float()
loss = loss_classify + LOSS_FN_WEIGHT * loss_ae + args.weight_mmd*loss_mmd
optimizer.zero_grad()
loss.backward()
optimizer.step()
total_loss += loss.item()
_, predicted = torch.max(output.data, 1)
total += target.size(0)
correct += (predicted == target).sum()
if index % 20 == 0:
tqdm.tqdm.write('Epoch: [{}/{}], Batch: [{}/{}], loss_ae:{:.4f}, loss_mmd:{:.4f}, loss_classify:{:.4f}, loss_total:{:.4f}'.format(e + 1, args.n_epoch, index + 1, n_batch,
loss_ae.item(), loss_mmd.item(), loss_classify.item(), loss.item()))
acc_train = float(correct) * 100.0 / (args.batch_size * n_batch)
tqdm.tqdm.write(
'Epoch: [{}/{}], loss: {:.4f}, train acc: {:.2f}%'.format(e + 1, args.n_epoch, total_loss * 1.0 / n_batch, acc_train))
# Testing
model.train(False)
with torch.no_grad():
correct, total = 0, 0
event_acc, event_miF, event_maF, frame_acc, frame_miF, frame_maF = 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
predicted_label_segment, lengths_varying_segment, true_label_segment = torch.LongTensor(), torch.LongTensor(), torch.LongTensor()
for sample, target in test_loader:
sample, target = sample.to(DEVICE).float(), target.to(DEVICE).long()
now_len = sample.shape[1]
# this line would cause error since the batch of last iteration does not have batch_size entries. so use DropLast = True when prep for dataloader
sample = sample.view(-1, feature_dim, now_len)
sample, target = sample.to(DEVICE).float(), target.to(DEVICE).long()
output, out_decoder = model(sample)
_, predicted = torch.max(output.data, 1)
total += target.size(0)
correct += (predicted == target).sum()
lengths_varying = [sample.shape[2]] * sample.shape[0]
lengths_varying = torch.LongTensor(lengths_varying)
predicted_label_segment = torch.LongTensor(torch.cat((predicted_label_segment, predicted.cpu()), dim=0))
lengths_varying_segment = torch.LongTensor(torch.cat((lengths_varying_segment, lengths_varying), dim=0))
true_label_segment = torch.LongTensor(torch.cat((true_label_segment, target.cpu()), dim=0))
# calculate different measurements
# event: accuracy, micro-F1, macro-F1
# frame: accuracy, micro-F1, macro-F1
event_acc, event_miF, event_maF, frame_acc, frame_miF, frame_maF = measure_event_frame(predicted_label_segment, lengths_varying_segment, true_label_segment)
acc_all.append([event_acc, event_miF, event_maF, frame_acc, frame_miF, frame_maF])
acc_all_T = np.array(acc_all).T.tolist()
best_e_miF = max([row[1] for row in acc_all])
best_iter = acc_all_T[1].index(best_e_miF) + 1
best_e_acc = acc_all[best_iter-1][0]
best_e_maF = acc_all[best_iter-1][2]
best_f_acc = acc_all[best_iter-1][3]
best_f_miF = acc_all[best_iter-1][4]
best_f_maF = acc_all[best_iter-1][5]
if sum(predicted_label_segment) == 0:
print('Note: All predicted labels are 0 in this epoch!\n')
tqdm.tqdm.write(
'Epoch: [{}/{}], e acc:{:.2f}%, e_miF:{:.2f}%, e maF:{:.2f}%, f acc:{:.2f}%, f miF:{:.2f}%, f maF:{:.2f}%, best acc:{:.2f}%, iter:{}'.format(
e + 1, args.n_epoch, event_acc, event_miF, event_maF, frame_acc, frame_miF, frame_maF, best_e_acc,
best_iter))
result.append([acc_train, event_acc, event_miF, event_maF, frame_acc, frame_miF, frame_maF, best_e_acc, best_iter])
result_np = np.array(result, dtype=float)
np.savetxt(result_name, result_np, fmt='%.2f', delimiter=',')
return best_e_acc, best_e_miF, best_e_maF, best_f_acc, best_f_miF, best_f_maF, best_iter
parser = argparse.ArgumentParser(description='argument setting of network')
parser.add_argument('--now_model_name', type=str, default='DDNN', help='the type of model, default DDNN')
parser.add_argument('--n_lstm_layer', type=int, default=1, help='number of lstm layers,default 2')
parser.add_argument('--n_lstm_hidden', type=int, default=64, help= 'number of lstm hidden dim, default 64')
parser.add_argument('--batch_size', type=int, default=64, help='batch size of training')
parser.add_argument('--n_epoch', type=int, default=100, help='number of training epochs')
parser.add_argument('--dataset', type=str, default='dg', help='name of dataset')
parser.add_argument('--n_feature', type=int, default=9, help='name of feature dimension')
parser.add_argument('--len_sw', type=int, default=32, help='length of sliding window')
parser.add_argument('--n_class', type=int, default=2, help='number of class')
parser.add_argument('--d_AE', type=int, default=50, help='dim of AE')
parser.add_argument('--sigma', type=float, default=1, help='parameter of mmd')
parser.add_argument('--weight_mmd', type=float, default=1.0, help='weight of mmd loss')
if __name__ == '__main__':
torch.manual_seed(10)
args = parser.parse_args()
train_loader, val_loader, test_loader = data_preprocess_dg.load_dataset_dg(batch_size=args.batch_size, SLIDING_WINDOW_LEN=32, SLIDING_WINDOW_STEP=16)
model = net.DDNN(args).to(DEVICE)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
result_name = 'results/' + args.dataset + '/' + str(args.n_epoch) + '_' + str(args.batch_size) + '_' + args.now_model_name + '_' + str(args.n_lstm_hidden) + '_' + str(args.n_lstm_layer) + '.csv'
if not os.path.exists('results/' + args.dataset):
os.makedirs('results/' + args.dataset)
if not os.path.isfile(result_name):
with open(result_name, 'w') as my_empty_csv:
pass
best_e_acc, best_e_miF, best_e_maF, best_f_acc, best_f_miF, best_f_maF, best_iter = train_dg_fixed(model, optimizer, train_loader, test_loader, result_name, args)
dataset_name = "{}".format(args.dataset)
with open('results/{}/best_result_{}.txt'.format(dataset_name, dataset_name), 'a') as f:
f.write('now_model_name: ' + args.now_model_name + '\t e_acc: ' + str(best_e_acc) + '\t e_miF: ' + str(
best_e_miF) + '\t e_maF: ' + str(best_e_maF)
+ '\t f_acc: ' + str(best_f_acc) + '\t f_miF: ' + str(best_f_miF) + '\t f_maF: ' + str(
best_f_maF) + '\t best_iter: ' + str(best_iter) + '\t d_AE: ' + str(args.d_AE)
+ '\t n_lstm_hidden: ' + str(args.n_lstm_hidden) + '\t n_lstm_layer: ' + str(args.n_lstm_layer)
+ '\t batch_size: ' + str(args.batch_size) + '\t n_epoch: ' + str(args.n_epoch) + '\n\n')
plot(result_name)