/
train_generative.py
189 lines (144 loc) · 6.33 KB
/
train_generative.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
import os
import torch
import torch.nn as nn
import torch.optim as optim
import numpy as np
import math
import random
import datetime
import pickle
from torch.utils.data import DataLoader, ConcatDataset, TensorDataset
from sklearn.metrics import matthews_corrcoef, f1_score
from tqdm import tqdm
from datasets import load_dataset
from data_loader import P2CDataset
from model import load_backbone, Classifier, Classifier_pref_ensemble
from common import parse_args
from utils import Logger, set_seed, set_model_path, save_model, AverageMeter, cut_input, ECE
from src.train import train_base, train_preference
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
CKPT_PATH = './checkpoint'
def main():
args = parse_args()
# Set seed
set_seed(args)
if args.pref_type == 'none':
log_name = f"{args.dataset}_{args.train_type}_{args.base}_S{args.seed}"
else:
log_name = f"{args.dataset}_{args.train_type}_{args.pref_type}_cons{args.lambda_cons}_div{args.lambda_div}_S{args.seed}"
logger = Logger(log_name)
log_dir = logger.logdir
logger.log(args)
logger.log(log_name)
logger.log('Loading pre-trained backbone network... ({})'.format(args.backbone))
backbone, tokenizer = load_backbone(args.backbone)
logger.log('Initializing model and optimizer...')
if 'dynasent' in args.dataset:
args.n_class = 3
elif 'emo' in args.dataset:
args.n_class = 6
else:
args.n_class = 2
if args.pref_type == 'none':
model = Classifier(args, args.backbone, backbone, args.n_class, args.train_type).to(device)
else:
model = Classifier_pref_ensemble(args, args.backbone, backbone, args.n_class, args.train_type).to(device)
if args.pre_ckpt is not None:
logger.log('Loading from pre-trained model')
model.load_state_dict(torch.load(args.pre_ckpt))
# Set optimizer (1) fixed learning rate and (2) no weight decay
optimizer = optim.Adam(model.parameters(), lr=args.model_lr, weight_decay=0)
logger.log('Initializing dataset...')
dataset = P2CDataset(args.dataset, tokenizer, args.backbone)
# Added for preference
train_loader = DataLoader(dataset.train_dataset, shuffle=True, drop_last=True, batch_size=args.batch_size, num_workers=4)
val_loader = DataLoader(dataset.val_dataset, shuffle=False, batch_size=args.batch_size, num_workers=4)
test_loader = DataLoader(dataset.test_dataset, shuffle=False, batch_size=args.batch_size, num_workers=4)
logger.log('==========> Start training ({})'.format(args.train_type))
best_acc, final_acc, final_ece = 0, 0, 0
train_labels = dataset.train_dataset[:][1]
args.n_samples = len(train_labels)
pref_train = None
prob_train = None
for epoch in range(1, 1+args.epochs):
# Set Dataloader
if args.pref_type == 'none':
train_base(args, train_loader, model, optimizer, epoch, logger)
else:
pref_train, prob_train = train_preference(args, train_loader, model, optimizer, epoch, logger)
best_acc, final_acc, final_ece = eval_func(args, model, val_loader, test_loader, logger, log_dir, epoch,
best_acc, final_acc, final_ece)
logger.log('===========>>>>> Final ECE: {}'.format(final_ece))
logger.log('===========>>>>> Final Test Accuracy: {}'.format(final_acc))
def eval_func(args, model, val_loader, test_loader, logger, log_dir, epoch, best_acc, final_acc, final_ece):
acc, ece_temp = test_acc(args, val_loader, model, logger)
if acc > best_acc:
t_acc, ece = test_acc(args, test_loader, model, logger, ece_temp)
# Update test accuracy based on validation performance
best_acc = acc
final_acc = t_acc
final_ece = ece
logger.log('========== Val ACC ==========')
logger.log('Val acc: {:.3f}'.format(best_acc))
logger.log('========== Test ACC ==========')
logger.log('Test acc: {:.3f}'.format(final_acc))
logger.log('========== Test ECE ==========')
logger.log('Test ece: {:.3f}'.format(final_ece))
# Save model
if args.save_ckpt:
logger.log('Save model...')
save_model(args, model, log_dir, epoch)
return best_acc, final_acc, final_ece
def test_acc(args, loader, model, logger=None, temp_opt=None):
if logger is not None:
logger.log('Compute test accuracy...')
model.eval()
all_preds = []
all_labels = []
for i, (tokens, labels, indices) in enumerate(loader):
tokens = tokens.long().to(device)
labels = labels.to(device)
with torch.no_grad():
if args.base == 'multi':
all_outputs = model(tokens)
outputs = 0
for i in range(len(all_outputs)):
outputs += (all_outputs[i].softmax(dim=-1))
outputs /= len(all_outputs)
else:
outputs = model(tokens)
all_preds.append(outputs)
all_labels.append(labels)
all_preds = torch.cat(all_preds, dim=0)
all_labels = torch.cat(all_labels, dim=0)
if temp_opt is None:
ece = ECE(all_preds, all_labels)
else:
ece = 100.0 * ECE(all_preds, all_labels, temp_opt=temp_opt)
all_preds = all_preds.cpu().max(1)[1]
all_labels = all_labels.cpu()
acc = 100.0 * (all_preds == all_labels).float().sum() / len(all_preds)
mcc = 100.0 * matthews_corrcoef(all_preds, all_labels)
if args.dataset == 'cola':
metric = mcc
elif args.dataset == 'spam' or args.dataset == 'hate':
bacc, worst_acc = [], 0
for i in range(2):
bacc.append(100.0 * (all_preds[all_labels == i] == i).float().sum() / (all_labels == i).float().sum())
worst_acc = min(bacc)
bacc = sum(bacc) / len(bacc)
logger.log("Acc: {}, bAcc: {}, wAcc: {}".format(acc, bacc, worst_acc))
metric = bacc
elif args.dataset == 'emo':
bacc, worst_acc = [], 0
for i in range(6):
bacc.append(100.0 * (all_preds[all_labels == i] == i).float().sum() / (all_labels == i).float().sum())
worst_acc = min(bacc)
bacc = sum(bacc) / len(bacc)
logger.log("Acc: {}, bAcc: {}, wAcc: {}".format(acc, bacc, worst_acc))
metric = bacc
else:
metric = acc
return metric, ece
if __name__ == "__main__":
main()