-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathMetaPrompting.py
161 lines (133 loc) · 7.92 KB
/
MetaPrompting.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
import os
import argparse
import random
import torch
import numpy as np
# import datetime
import dataloader as loader
from model import MetaTransformerModelWrapper
from utils import tprint
def parse_args():
parser = argparse.ArgumentParser(
description="MetaPrompting")
# data configuration
parser.add_argument("--data_path", type=str, default="data/",
help="path to dataset")
parser.add_argument("--dataset", type=str, default="huffpost",
help="name of the dataset. "
"Options: [20newsgroup, amazon, huffpost, reuters]")
parser.add_argument("--stored_episodes_dir", default="./episode_data/",
help="path to stored episode, "
"if given, will directly load the episode from the path, "
"otherwise new episodes will be built from training data.")
parser.add_argument("--output_dir", default="./output_dir/",
type=str, required=False,
help="The output directory where the model predictions and checkpoints will be written")
# model configuration
parser.add_argument("--pretrained_model", default="bert",
help="use PLM embedding (only available for sent-level datasets: huffpost, fewrel")
parser.add_argument("--model_type", default="bert-base-uncased")
parser.add_argument("--pretrained_cache_dir", default="./pretrained_models/",
type=str, help="path to the cache_dir of transformers")
parser.add_argument("--embed_size", default=768, type=int,
help="Prompt tokens embedding size")
# task configuration
parser.add_argument("--n_way", type=int, default=5, help="the number of classes for each task")
parser.add_argument("--k_shot", type=int, default=5,
help="the number of support examples for each class for each task")
parser.add_argument("--l_query", type=int, default=15,
help="the number of query examples for each class for each task")
# train/test configuration
parser.add_argument("--train_epochs", type=int, default=300,
help="max num of training epochs")
parser.add_argument("--val_episodes", type=int, default=100,
help="episodes sampled during each validation")
parser.add_argument("--test_episodes", type=int, default=1000,
help="episodes sampled during each test")
# training options
parser.add_argument("--eval_every_step", type=int, default=100, help="eval_every_step")
parser.add_argument("--seed", type=int, default=1999, help="seed")
parser.add_argument("--patience", type=int, default=5, help="patience")
parser.add_argument("--clip_grad", type=float, default=1., help="gradient clipping")
parser.add_argument("--prompt_template", type=int, default=0)
parser.add_argument("--n_adapt_epochs", type=int, default=15,
help="the number of adaption epochs on evaluation few-shot episodes")
parser.add_argument("--inner_steps", type=int, default=15, help="the number of adaption steps")
parser.add_argument("--batch_size", type=int, default=16)
parser.add_argument("--lm_learning_rate", type=float, default=1e-5)
parser.add_argument("--prompt_learning_rate", type=float, default=5e-5)
parser.add_argument("--no_train", type=int, default=0)
parser.add_argument("--no_eval", type=int, default=0)
parser.add_argument("--no_load", type=int, default=0)
return parser.parse_args()
def set_seed(seed):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
def main():
args = parse_args()
args.data_path = args.data_path + args.dataset + ".json"
config_str = str(args.n_way) + 'way' + str(args.k_shot) + 'shot_' + str(args.inner_steps) + 'ada' + \
str(args.seed) + 'seed' + str(args.prompt_template) + 'template'
tprint(config_str)
args.output_dir = os.path.join(args.output_dir, args.dataset + '_' + config_str + '')
set_seed(args.seed)
stored_episodes_dir = os.path.join(args.stored_episodes_dir, args.dataset,
str(args.n_way) + 'way' + str(args.k_shot) + 'shot')
if not (os.path.exists(os.path.join(stored_episodes_dir, "train.json")) and
os.path.exists(os.path.join(stored_episodes_dir, "val.json")) and
os.path.exists(os.path.join(stored_episodes_dir, "test.json"))):
train_data, val_data, test_data, train_classes, val_classes, test_classes = loader.load_dataset(args)
else:
train_data, val_data, test_data = [], [], []
train_episodes = loader.build_episodes(train_data, n_way=args.n_way, k_shot=args.k_shot, l_query=args.l_query,
n_episodes=6000, stored_episodes=os.path.join(stored_episodes_dir,
"train.json"))
val_episodes = loader.build_episodes(val_data, n_way=args.n_way, k_shot=args.k_shot, l_query=args.l_query,
n_episodes=1000, stored_episodes=os.path.join(stored_episodes_dir,
"val.json"))
test_episodes = loader.build_episodes(test_data, n_way=args.n_way, k_shot=args.k_shot, l_query=args.l_query,
n_episodes=3000, stored_episodes=os.path.join(stored_episodes_dir,
"test.json"))
n_gpu = torch.cuda.device_count()
per_gpu_train_batch_size = args.batch_size // n_gpu
print("\n=======================================================")
print("==================Meta training stage==================")
print("=======================================================\n")
wrapper = MetaTransformerModelWrapper(args)
if args.no_train == 0:
wrapper.train(train_data=train_episodes,
eval_data=val_episodes,
task_output_dir=args.output_dir,
per_gpu_train_batch_size=per_gpu_train_batch_size,
n_gpu=n_gpu,
eval_every_step=args.eval_every_step,
n_adapt_epochs=args.n_adapt_epochs,
n_train_epochs=args.train_epochs,
weight_decay=0.1,
n_inner_steps=3,
lm_learning_rate=args.lm_learning_rate,
prompt_learning_rate=args.prompt_learning_rate,
max_grad_norm=args.clip_grad
)
if args.no_eval == 0:
print("\n=======================================================")
print("==================Meta testing stage===================")
print("=======================================================\n")
print("Loading best model")
pretrained_model_path = os.path.join(args.output_dir, 'best') if args.no_load == 0 else None
average_scores = wrapper.eval(eval_data=test_episodes,
# classes=test_classes,
pretrained_model_path=pretrained_model_path,
n_eval_episodes=args.test_episodes,
per_gpu_eval_batch_size=per_gpu_train_batch_size,
n_gpu=n_gpu,
n_adapt_epochs=args.n_adapt_epochs,
lm_learning_rate=args.lm_learning_rate,
prompt_learning_rate=args.prompt_learning_rate,
)
print(average_scores)
print("\n=======================================================")
if __name__ == "__main__":
main()