-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathmulticls_novalid_vtuning.py
191 lines (151 loc) · 8.63 KB
/
multicls_novalid_vtuning.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
import numpy as np
import torch
import tqdm
import time
import os
from src.multicls_trainer import PromptBoostingTrainer
from src.ptuning import BaseModel, OPTVTuningClassification, RoBERTaVTuningClassification
from src.saver import PredictionSaver, TestPredictionSaver
from src.template import SentenceTemplate, TemplateManager
from src.utils import ROOT_DIR, BATCH_SIZE, create_logger, MODEL_CACHE_DIR
from src.data_util import get_class_num, get_weak_cls_num, load_dataset, get_task_type, get_template_list
import wandb
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--adaboost_lr", type = float, default = 1.0)
parser.add_argument("--dataset", type = str, default = 'sst')
parser.add_argument("--model", type = str, default = 'roberta')
parser.add_argument("--template_name", type = str, default = 't5_template3')
parser.add_argument("--stop_criterior", type = str, default = 'best')
parser.add_argument("--label_set_size", type = int, default = 5)
parser.add_argument("--max_template_num", type = int, default = 0)
parser.add_argument("--pred_cache_dir", type = str, default = '')
parser.add_argument("--use_logits", action = 'store_true')
parser.add_argument("--use_wandb", action = 'store_true')
parser.add_argument("--change_template", action = 'store_true')
parser.add_argument("--manual", action = 'store_true')
parser.add_argument("--use_part_templates", action = 'store_true')
parser.add_argument("--start_idx", type = int, default = 0)
parser.add_argument("--end_idx", type = int, default = 10)
parser.add_argument("--second_best", action = 'store_true')
parser.add_argument("--sort_dataset", action = 'store_true')
parser.add_argument("--fewshot", action = 'store_true')
parser.add_argument("--fewshot_k", type = int, default = 0)
parser.add_argument("--fewshot_seed", type = int, default = 100, choices = [100, 13, 21, 42, 87])
args = parser.parse_args()
if __name__ == '__main__':
device = torch.device('cuda')
adaboost_lr = args.adaboost_lr
template_name = args.template_name
dataset = args.dataset
sentence_pair = get_task_type(dataset)
num_classes = get_class_num(dataset)
adaboost_weak_cls = get_weak_cls_num(dataset)
model = args.model
pred_cache_dir = args.pred_cache_dir
sort_dataset = args.sort_dataset
stop_criterior = args.stop_criterior
use_logits = args.use_logits
use_wandb = args.use_wandb
label_set_size = args.label_set_size
max_template_num = args.max_template_num
adaboost_maximum_epoch = 20000
fewshot = args.fewshot
fewshot_k = args.fewshot_k
fewshot_seed = args.fewshot_seed
filter_templates = args.filter_templates
suffix = ""
if args.use_part_templates:
suffix = f"range({args.start_idx}-{args.end_idx})"
if filter_templates:
suffix += f"filtered"
logger, log_dir = create_logger(logger_name='ensemble_novalid',filename = f'{model}-{dataset}-{suffix}-novalid')
wandb_name = f"{model}-{dataset}-{suffix}-novalid"
if use_wandb:
if fewshot:
wandb_name += f"-{fewshot_k}shot-seed{fewshot_seed}"
wandb.init(project = f'vtuning-{dataset}', name = f'{wandb_name}')
train_dataset, _, test_dataset = load_dataset(dataset_name = dataset, sort_dataset = sort_dataset, fewshot = fewshot, k = fewshot_k, rand_seed = fewshot_seed,
use_valid_for_train = True)
# label_token_set_path = ROOT_DIR + 'label_maps/sst_roberta_template1.json'
num_training = len(train_dataset[0])
num_test = len(test_dataset[0])
train_labels = torch.LongTensor(train_dataset[1]).to(device)
test_labels = torch.LongTensor(test_dataset[1]).to(device)
weight_tensor = torch.ones(num_training, dtype = torch.float32).to(device) / num_training
if model == 'roberta':
vtuning_model = RoBERTaVTuningClassification(model_type = 'roberta-large', cache_dir = os.path.join(MODEL_CACHE_DIR, 'roberta_model/roberta-large/'),
device = device, verbalizer_dict = None, sentence_pair = sentence_pair)
elif model == 'opt1.3b':
vtuning_model = OPTVTuningClassification(model_type = 'facebook/opt-1.3b', cache_dir = os.path.join(MODEL_CACHE_DIR, 'opt_model/opt-1.3b/'),
device = device, verbalizer_dict = None, sentence_pair = sentence_pair)
template_dir_list = get_template_list(dataset)
template_manager = TemplateManager(template_dir_list = template_dir_list, output_token = vtuning_model.tokenizer.mask_token, max_template_num = max_template_num,
use_part_templates = args.use_part_templates, start_idx = args.start_idx, end_idx = args.end_idx)
dir_list = "\n\t".join(template_manager.template_dir_list)
print(f"using templates from: {dir_list}",)
trainer = PromptBoostingTrainer(adaboost_lr = adaboost_lr, num_classes = num_classes, adaboost_maximum_epoch = adaboost_maximum_epoch)
if pred_cache_dir != '':
prediction_saver = PredictionSaver(save_dir = os.path.join(ROOT_DIR, pred_cache_dir, 'novalid/'), model_name = model,
fewshot = fewshot, fewshot_k = fewshot_k, fewshot_seed = fewshot_seed,
)
else:
prediction_saver = PredictionSaver(model_name = model,
fewshot = fewshot, fewshot_k = fewshot_k, fewshot_seed = fewshot_seed,
)
test_pred_saver = TestPredictionSaver(save_dir = os.path(ROOT_DIR, f'cached_test_preds/{dataset}/'), model_name = model)
train_probs, valid_probs = [],[]
word2idx = vtuning_model.tokenizer.get_vocab()
for model_id in tqdm.tqdm(range(adaboost_weak_cls)):
if args.change_template:
del train_probs
del valid_probs
template = template_manager.change_template()
template.visualize()
cached_preds, flag = prediction_saver.load_preds(template)
if not flag:
train_probs = trainer.pre_compute_logits(vtuning_model, template, train_dataset,)
valid_probs = []
prediction_saver.save_preds(template, train_probs, valid_probs)
else:
train_probs, valid_probs = cached_preds
trainer.record_dataset_weights(weight_tensor)
verbalizer, train_error,train_acc, wrong_flags,train_preds= trainer.train(train_dataset, vtuning_model, train_probs, train_labels,
weight_tensor = weight_tensor,label_set_size = label_set_size,
)
print(verbalizer)
if train_error < 1 - (1 / (num_classes)):
print(f"\tmodel {model_id + 1} finished")
print(f"\ttrain error {train_error}, train_acc {train_acc}")
logger.info(f"\tmodel {model_id + 1} finished")
logger.info(f"\ttrain error {train_error}, train_acc {train_acc}")
succ_flag = True
else:
print(f"error {train_error}; train_acc {train_acc}\n Ensemble is worse than random, ensemble can not be fit.")
logger.info(f"error {train_error}; train_acc {train_acc}\n Ensemble is worse than random, ensemble can not be fit.")
continue
alpha, weight_tensor = trainer.adaboost_step(train_error, wrong_flags, weight_tensor)
print(f"\talpha {alpha}")
logger.info(f"\talpha {alpha}")
trainer.save_prediction(train_preds, split = 'train')
train_ensemble_acc = trainer.ensemble_result(train_labels, split = 'train')
trainer.best_epoch = len(trainer.model_weight_tensor)
trainer.save_weak_learner(verbalizer, template.template_name)
tolog = {
'train_error': train_error,
'alpha': alpha,
'train_acc': train_acc,
'ensemble_train_acc': train_ensemble_acc,
}
if use_wandb:
wandb.log(tolog)
print(f"finish training with {len(trainer.model_weight_tensor)} weak classifier")
print(f"best ensemble classfier: 0 - {trainer.best_epoch}")
logger.info(f"finish training with {len(trainer.model_weight_tensor)} weak classifier")
logger.info(f"best ensemble classfier: 0 - {trainer.best_epoch}")
all_template_used = template_manager.get_all_template()
test_ensemble_acc = trainer.final_eval(test_dataset, vtuning_model, all_template_used, test_pred_saver)
print(f"best test acc {test_ensemble_acc}")
logger.info(f"best test acc {test_ensemble_acc}")
to_log = {"best_test":test_ensemble_acc}
wandb.log(to_log)