-
Notifications
You must be signed in to change notification settings - Fork 3
/
starter_siqa.py
61 lines (46 loc) · 1.87 KB
/
starter_siqa.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
import trainer
from transformers import BertTokenizer, BertConfig, BertModel
from adapter import AdapterBertModel, AdapterBertForSequenceClassification, ParallelAdapterBertForSequenceClassification
from modeling_adapters import AdapterBertForMultichoiceQA
import helper
import torch
import numpy as np
import logging
import torch
import os
from config import RunConfig
from modeling_adapters import BottleneckAdapterBertConfig
import torch
import sys
logger = logging.getLogger(__name__)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
c = RunConfig()
bertonja = BertModel.from_pretrained(c.pretrained_transformer)
train_data = torch.load(c.train_set)
eval_data = torch.load(c.val_set)
if not os.path.exists(c.output_dir):
os.makedirs(c.output_dir)
# Set seed
trainer.set_seed(c.seed)
# Loading tokenizer, configuration, and model
tokenizer = BertTokenizer.from_pretrained(c.pretrained_transformer)
match = [f for f in os.listdir(c.output_dir) if f.startswith("best")]
if len(match) > 0:
c.model_name_or_path = os.path.join(c.output_dir, 'best')
#config = ParallelAdapterBertConfig.from_pretrained(c.model_name_or_path)
config = BottleneckAdapterBertConfig.from_pretrained(c.model_name_or_path)
if len(match) == 0:
config.layers_to_adapt = c.layers_to_adapt
config.num_labels = 3
config.num_labels = 3
#model = ParallelAdapterBertForSequenceClassification.from_pretrained(c.model_name_or_path, config=config)
model = AdapterBertForMultichoiceQA.from_pretrained(c.model_name_or_path, config=config)
model.to(c.device)
logger.info("Training/evaluation starts...")
params = {"task_type" : "mcqa", "model_params" : {}, "mcqa_config" : c}
_, _, eval_perf = trainer.train(train_data, eval_data, model, params)