/
run_end_model_ag.py
executable file
·177 lines (145 loc) · 5.72 KB
/
run_end_model_ag.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
import random
import numpy as np
import sys
import torch
from allennlp.data.data_loaders import SimpleDataLoader
from allennlp.data import Vocabulary, Instance, Token
from allennlp.data.token_indexers import ELMoTokenCharactersIndexer
from copy import deepcopy as dcopy
from allennlp.data.fields import LabelField, ArrayField, TextField
from allennlp.training.optimizers import AdamWOptimizer
from allennlp.training.learning_rate_schedulers import CosineWithRestarts
from allennlp.training.trainer import Trainer, GradientDescentTrainer
from allennlp.training.util import evaluate
import joblib
from end.backbones.text_classifiers import BertClassifier
from allennlp.data.tokenizers.pretrained_transformer_tokenizer \
import PretrainedTransformerTokenizer
from allennlp.data.token_indexers.pretrained_transformer_indexer \
import PretrainedTransformerIndexer
from allennlp.modules.token_embedders.pretrained_transformer_embedder \
import PretrainedTransformerEmbedder
from allennlp.modules.text_field_embedders.basic_text_field_embedder \
import BasicTextFieldEmbedder
raw_dataset = 'data/agnews/{:s}.jkl'
def load_jkl(file):
data = joblib.load(file)
return data
def proc_inst(instance):
return instance['string']
def extract_inst_label_pair(file, pack=False, ignore_label=False):
inst = []
labels = []
for item in file:
sentence = proc_inst(item)
inst.append(sentence)
if not ignore_label:
label = item['label']
labels.append(label)
else:
pack = False
if pack:
return list(zip(inst, labels))
return inst, labels
def to_instance(tuples, tokenizer, token_indexers, soft_labeled=False):
onehot = np.eye(4)
data = []
for item, label in tuples:
if soft_labeled:
label_field = ArrayField(label)
else:
label_field = ArrayField(onehot[label])
fields = {'text': TextField(tokenizer.tokenize(item),
token_indexers=token_indexers), 'label': label_field}
data.append(Instance(fields))
return data
def build_allentrainer(model, train_dl, val_dl, save_loc='output/'):
parameters = [
[n, p]
for n, p in model.named_parameters() if p.requires_grad
]
optimizer = AdamWOptimizer(parameters, lr=5e-7)
default_epoch = 20
trainer = GradientDescentTrainer(
model=model,
data_loader=train_dl,
validation_data_loader=val_dl,
num_epochs=default_epoch,
optimizer=optimizer,
patience=5,
serialization_dir=save_loc,
cuda_device=0,
use_amp=False,
grad_clipping=2.0,
validation_metric='+average_F1'
)
return trainer
def process_voc(inst, save=False):
vocab = Vocabulary.from_instances(inst)
if save:
vocab.save_to_files('vocabulary/')
vocab.print_statistics()
return vocab
def main(soft_label_loc=None, rand=-2,):
if rand >= 0:
seeds = [500, 600, 700, 800, 900]
random_seed = pt_seed = np_seed = seeds[rand]
random.seed(random_seed)
torch.manual_seed(pt_seed)
np.random.seed(np_seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(pt_seed)
else:
rand = 1e3
model_string = "bert-base-uncased"
tokenizer = PretrainedTransformerTokenizer(model_string)
token_indexer = {'tokens': PretrainedTransformerIndexer(
model_string)}
train = load_jkl(dcopy(raw_dataset).format('train'))
val = load_jkl(dcopy(raw_dataset).format('dev'))
print('Soft Label :', soft_labels_loc)
sla = np.load(soft_label_loc, allow_pickle=True).item()
training_soft_labels = sla['sl']
train_inst, _ = extract_inst_label_pair(train, pack=False)
idx = sla['idx']
if idx is not None:
train_inst = [inst for i, inst in enumerate(train_inst) if i in idx]
train_inst = [inst for i, inst in enumerate(train_inst) if i in idx]
print(len(train_inst), len(idx), training_soft_labels.shape)
train_tuple = list(zip(train_inst, training_soft_labels))
soft_labeled = True
save_loc = 'scratch/agnews/'
val_tuple = extract_inst_label_pair(val, pack=True)
train_insts = to_instance(train_tuple, tokenizer=tokenizer, token_indexers=token_indexer, soft_labeled=soft_labeled)
val_insts = to_instance(val_tuple, tokenizer=tokenizer, token_indexers=token_indexer)
test = load_jkl(dcopy(raw_dataset).format('test'))
test_tuple = extract_inst_label_pair(test, pack=True)
test_insts = to_instance(test_tuple, tokenizer=tokenizer,
token_indexers=token_indexer)
vocabulary = process_voc(train_insts + val_insts + test_insts, save=True)
num_class = 4
bert_token_embedder = PretrainedTransformerEmbedder(model_string)
bert_textfield_embedder = BasicTextFieldEmbedder(
{"tokens": bert_token_embedder})
model = BertClassifier(
vocabulary, bert_textfield_embedder, freeze_encoder=False, out_features=4)
model = model.to('cuda:0')
train_dl = SimpleDataLoader(train_insts, batch_size=16, shuffle=True, vocab=vocabulary)
val_dl = SimpleDataLoader(val_insts, batch_size=16, shuffle=False, vocab=vocabulary)
trainer = build_allentrainer(model, train_dl, val_dl, save_loc)
print("Starting training")
trainer.train()
print("Finished training")
test_dl = SimpleDataLoader(test_insts, batch_size=16, shuffle=False, vocab=vocabulary)
results = evaluate(model, test_dl, cuda_device=0)
print(results)
if __name__ == '__main__':
try:
soft_labels_loc = sys.argv[1]
except IndexError:
soft_labels_loc = None
try:
seed = int(sys.argv[2])
except:
seed = 1
main(soft_label_loc=soft_labels_loc, rand=seed)