-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
209 lines (155 loc) · 7.41 KB
/
train.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
import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import Dataset, random_split, DataLoader, RandomSampler, SequentialSampler
from torch.nn.utils import clip_grad_norm_
import torch.optim as optim
import torchvision.transforms as transforms
from layer import Sec2SecCoattention
from dataloader import LIRIS_Sec2SecAV
from utils import get_train_val_test_size, set_seed
from datetime import datetime
import argparse
import os
import json
parser = argparse.ArgumentParser()
parser.add_argument('-emotion', type=str, default='arousal')
parser.add_argument('-batch_size', type=int, default=64)
parser.add_argument('-num_of_epochs', type=int, default=250)
parser.add_argument('-device', type=str, default='cuda')
parser.add_argument('-video_length', type=int, default=15)
parser.add_argument('-d_input', type=int, default=1024)
parser.add_argument('-num_of_dense_layers', type=int, default=1)
parser.add_argument('-dense_hidden_dim', type=int, default=1024)
parser.add_argument('-lstm_hidden_dim', type=int, default=1024)
parser.add_argument('-num_of_heads', type=int, default=64)
parser.add_argument('-seed_number', type=int, default=3407)
parser.add_argument('-lr', type=float, default=1e-7)
parser.add_argument('-lstm_layer', type=int, default=1)
parser.add_argument('-norm', type=str, default='layer')
parser.add_argument('-dropout', type=float, default=0.1)
parser.add_argument('-max_grad_norm', type=int, default=1)
parser.add_argument('-rating_path', type=str)
parser.add_argument('-sec2sec_feature_path', type=str)
parser.add_argument('-output_path', type=str)
opt = parser.parse_args()
set_seed(opt.seed_number)
data = LIRIS_Sec2SecAV(opt.emotion, opt.sec2sec_feature_path, opt.rating_path)
train_size, val_size, test_size = get_train_val_test_size(len(data), (7,1,2))
train_dataset, val_dataset, test_dataset = random_split(data, [train_size, val_size, test_size])
train_dataloader = DataLoader(train_dataset, sampler=RandomSampler(train_dataset), batch_size=opt.batch_size)
val_dataloader = DataLoader(val_dataset, sampler=SequentialSampler(val_dataset), batch_size=opt.batch_size)
test_dataloader = DataLoader(test_dataset, sampler=SequentialSampler(test_dataset), batch_size=opt.batch_size)
model = Sec2SecCoattention(opt.d_input, opt.num_of_heads, opt.lstm_hidden_dim, opt.lstm_layer, opt.num_of_dense_layers, \
opt.dense_hidden_dim, opt.device, opt.norm, opt.dropout)
optimizer = torch.optim.Adam(model.parameters(), lr=opt.lr)
loss_fn = nn.BCELoss()
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=20, gamma=0.1)
model.to(opt.device)
loss_fn.to(opt.device)
training_loss_list = []
validation_loss_list = []
best_val_accuracy = 0
best_model = model
prev_time = datetime.now()
early_stop_count = 0
def metrics(preds, targets):
TP, FP, TN, FN = 0, 0, 0, 0
for pred, target in zip(preds, targets):
if (pred == 1) and (target == 1):
TP += 1
elif (pred == 1) and (target == 0):
FP += 1
elif (pred == 0) and (target == 0):
TN += 1
else:
FN += 1
accuracy = (TP+TN)/(TP+TN+FN+FP)
precision = (TP)/(TP+FP)
recall = (TP)/(TP+FN)
f1_score = (2*precision*recall)/(precision+recall)
return {'accuracy': accuracy, 'precision': precision, 'recall' :recall, 'f1_score': f1_score}
for e in range(opt.num_of_epochs):
training_loss_per_epoch = 0
val_loss_per_epoch = 0
model.train()
print('Epoch', e, 'Start training')
for step, batch in enumerate(train_dataloader):
audio = batch[0].to(opt.device)
visual = batch[1].to(opt.device)
rating = batch[2].to(opt.device)
output = model(audio, visual)
loss = loss_fn(output.squeeze(), rating.float())
loss.backward()
clip_grad_norm_(model.parameters(), opt.max_grad_norm) # clipping gradient for avoiding exploding gradients
optimizer.step()
training_loss_per_epoch += float(loss)
scheduler.step()
training_loss_per_epoch = training_loss_per_epoch/train_size
training_loss_list.append(training_loss_per_epoch)
print('Start validating')
model.eval()
val_pred_list = []
val_label_list = []
with torch.no_grad():
for step, batch in enumerate(val_dataloader):
audio = batch[0].to(opt.device)
visual = batch[1].to(opt.device)
rating = batch[2].to(opt.device)
output = model(audio, visual)
loss = loss_fn(output.squeeze(), rating.float())
preds = output.reshape(-1).detach().round().tolist()
labels = rating.reshape(-1).tolist()
val_pred_list += preds
val_label_list += labels
val_loss_per_epoch += float(loss)
val_loss_per_epoch = val_loss_per_epoch/val_size
validation_loss_list.append(val_loss_per_epoch)
val_accuracy = metrics(val_pred_list, val_label_list)['accuracy']
if val_accuracy > best_val_accuracy:
best_model = model
best_val_accuracy = val_accuracy
if (e > 20) and (val_loss_per_epoch > validation_loss_list[-2]):
early_stop_count += 1
if early_stop_count == 5:
break
else:
early_stop_count = 0
current_time = datetime.now()
print('Current time at', current_time, 'Epoch took', current_time - prev_time, \
'Training loss', training_loss_per_epoch, 'Validation loss', val_loss_per_epoch,
'Validation Accuracy', val_accuracy, 'Best val accuracy', best_val_accuracy)
prev_time = current_time
test_pred_list = []
test_label_list = []
model.eval()
with torch.no_grad():
for step, batch in enumerate(test_dataloader):
audio = batch[0].to(opt.device)
visual = batch[1].to(opt.device)
rating = batch[2].to(opt.device)
output = model(audio, visual)
preds = output.reshape(-1).detach().round().tolist()
labels = rating.reshape(-1).tolist()
test_pred_list += preds
test_label_list += labels
report = metrics(test_pred_list, test_label_list)
output_path = opt.output_path + opt.emotion + '/'
if not os.path.exists(output_path):
os.mkdir(output_path)
output_foldername = '_'.join([str(opt.batch_size), 'batchsize', str(opt.num_of_heads), \
'heads', str(opt.lstm_layer), 'lstmlayernum', opt.norm + 'norm', \
str(opt.num_of_dense_layers), 'denselayernum', str(opt.dense_hidden_dim),\
'densehiddim', str(opt.lstm_hidden_dim), 'lstmhiddim', str(opt.seed_number)])
output_path += output_foldername + '/'
if not os.path.exists(output_path):
os.mkdir(output_path)
torch.save(best_model.state_dict(), output_path + 'best_model.pth')
json.dump(training_loss_list, open(output_path + 'training_loss_list.json', 'w'))
json.dump(validation_loss_list, open(output_path + 'val_loss_list.json', 'w'))
json.dump(report, open(output_path + 'testing_report.json', 'w'))
preds_and_labels = {'preds': test_pred_list, 'labels': test_label_list}
json.dump(preds_and_labels, open(output_path + 'preds_labels.json', 'w'))
with open(output_path + 'readme.txt', 'w') as f:
f.write('Stopping epoch number ' + str(e + 1) + '\n')
f.write('Best validation accuracy' + str(best_val_accuracy) + '\n')