-
Notifications
You must be signed in to change notification settings - Fork 3
/
finetune_baseline.py
178 lines (160 loc) · 7.63 KB
/
finetune_baseline.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
import torch
import numpy as np
from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score
from tqdm import tqdm
import random
from collections import Counter
from utils import NusaXDataset, NusaTranslationDataset, TatoebaDataset, BUCCDataset, LinceMTDataset, PhincDataset, LinceSADataset, MassiveIntentDataset, Sib200Dataset, NollySentiDataset, MTOPIntentDataset, FIREDataset
from transformers import AutoTokenizer, DataCollatorWithPadding, AutoModelForSequenceClassification, TrainingArguments, Trainer, EarlyStoppingCallback
from datasets import Dataset
import evaluate
import argparse
import os
import json
def set_seed(seed):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
def get_tokenized_dataset(lang_id, tokenizer, max_length):
def tokenize_function(example):
return tokenizer(example['source'], truncation=True, padding='max_length')
train_dataset = Dataset.from_dict(dataset.train_data[lang_id])
val_dataset = Dataset.from_dict(dataset.valid_data[lang_id])
test_dataset = Dataset.from_dict(dataset.test_data[lang_id])
tokenized_train_datasets= train_dataset.map(tokenize_function, batched=True)
tokenized_val_datasets = val_dataset.map(tokenize_function, batched=True)
tokenized_test_datasets = test_dataset.map(tokenize_function, batched=True)
tokenized_train_datasets = tokenized_train_datasets.rename_column("target", "labels")
tokenized_val_datasets = tokenized_val_datasets.rename_column("target", "labels")
tokenized_test_datasets = tokenized_test_datasets.rename_column("target", "labels")
tokenized_train_datasets.set_format(
type="torch", columns=["input_ids", "attention_mask", "labels"]
)
tokenized_val_datasets.set_format(
type="torch", columns=["input_ids", "attention_mask", "labels"]
)
tokenized_test_datasets.set_format(
type="torch", columns=["input_ids", "attention_mask", "labels"]
)
return tokenized_train_datasets, tokenized_val_datasets, tokenized_test_datasets
def compute_metrics(eval_preds):
"""Computes accuracy."""
metric = evaluate.load("accuracy")
logits, labels = eval_preds # raw scores of size (batch_size, num_classes), labels 0-2
predictions = np.argmax(logits, axis=-1) # take max of last dimension (num_classes)
return metric.compute(predictions=predictions, references=labels)
def train_eval_loop(dataset_name, lang_id, output_dir, model_name, label_num, compute_metrics):
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=label_num)
model_output_dir = f"model_out/{dataset_name}_{model_name}_{lang_id}"
BATCH_SIZE = 8
max_length = tokenizer.model_max_length
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
tokenized_train_datasets, tokenized_val_datasets, tokenized_test_datasets = get_tokenized_dataset(lang_id, tokenizer, max_length)
training_args = TrainingArguments(
do_train=True,
do_eval=True,
output_dir=model_output_dir,
num_train_epochs=10,
weight_decay=0.01,
per_device_train_batch_size=BATCH_SIZE,
per_device_eval_batch_size=BATCH_SIZE,
save_strategy="epoch",
save_total_limit=1,
eval_strategy="epoch",
greater_is_better=True,
metric_for_best_model="accuracy",
load_best_model_at_end=True,
logging_dir=f"finetune_logs/{dataset_name}_{model_name}_{lang_id}",
logging_steps=100,
)
early_stopping_callback = EarlyStoppingCallback(
early_stopping_patience=2, # Number of evaluation steps to wait for improvement
early_stopping_threshold=0.0 # Threshold for measuring the new optimum
)
trainer = Trainer(
model=model,
args=training_args,
tokenizer=tokenizer,
data_collator=data_collator,
train_dataset=tokenized_train_datasets,
eval_dataset=tokenized_val_datasets,
compute_metrics=compute_metrics,
callbacks=[early_stopping_callback]
)
print("start training")
trainer.train()
metric = evaluate.load("f1")
model.eval()
for batch in tokenized_test_datasets:
batch = {'input_ids': batch['input_ids'], 'attention_mask': batch['attention_mask'], 'labels': batch['labels']}
with torch.no_grad():
outputs = model(batch['input_ids'].unsqueeze(0).to('cuda'), attention_mask=batch['attention_mask'].unsqueeze(0).to('cuda'))
logits = outputs.logits
predictions = torch.argmax(logits, dim=-1)
metric.add_batch(predictions=[predictions], references=[batch["labels"]])
f1 = metric.compute(average='micro')
if not os.path.exists(f"{output_dir}/{dataset_name}/"):
os.makedirs(f"{output_dir}/{dataset_name}/")
with open(f"{output_dir}/{dataset_name}/{lang_id}.json", "w") as f:
dict_ = {"dataset_name":dataset_name, "lang_id":lang_id, "f1": f1["f1"]}
json.dump(dict_, f)
return f1
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--cross", action="store_true")
parser.add_argument("--model", type=str, default="xlm-roberta-base", help="model name")
parser.add_argument("--dataset", type=str, default="nusax", help="dataset name")
parser.add_argument("--prompt", type=str, default="", help="prompt")
parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
parser.add_argument("--src_lang", type=str, default="x", help="source language")
args = parser.parse_args()
print("###########################")
print("dataset:", args.dataset)
print("model:", args.model)
print("prompt:", args.prompt)
print("seed:", args.seed)
print("cross:", args.cross)
print("src_lang:", args.src_lang)
print("###########################")
set_seed(args.seed)
if args.cross:
output_dir = f"finetune_classification_cross_{args.src_lang}"
else:
output_dir = "finetune_classification"
if args.dataset == "nusax":
dataset = NusaXDataset(prompt=args.prompt, task="classification")
label_num = 3
if args.dataset == "lince_sa":
dataset = LinceSADataset(prompt=args.prompt)
label_num = 3
if args.dataset == "massive_intent":
dataset = MassiveIntentDataset(prompt=args.prompt)
label_num = 60
if args.dataset == "sib200":
dataset = Sib200Dataset(prompt=args.prompt)
label_num = 7
if args.dataset == "nollysenti":
dataset = NollySentiDataset(prompt=args.prompt, task="classification")
label_num = 2
if args.dataset == "mtop_intent":
dataset = MTOPIntentDataset(prompt=args.prompt)
label_num = 113
if args.dataset == 'fire':
dataset = FIREDataset(prompt=args.prompt)
for lang in dataset.LANGS:
print("language: ", lang)
lang_id = lang
if lang == 'malayalam':
label_num = 4
elif lang == 'tamil':
label_num = 5
MODEL_NAME = args.model
train_eval_loop(args.dataset, lang_id, output_dir, MODEL_NAME, label_num, compute_metrics)
else:
for lang in dataset.LANGS:
print("language: ", lang)
lang_id = lang
MODEL_NAME = args.model
train_eval_loop(args.dataset, lang_id, output_dir, MODEL_NAME, label_num, compute_metrics)