finetune_bert.py

from transformers import BertForSequenceClassification, BertTokenizerFast, Trainer, TrainingArguments
from datasets import load_dataset, DatasetDict, Dataset
import torch
import numpy as np
from sklearn.metrics import accuracy_score, precision_recall_fscore_support

model = BertForSequenceClassification.from_pretrained('./bert-base-uncased')
tokenizer = BertTokenizerFast.from_pretrained('./bert-base-uncased')

def tokenize(batch):
    return tokenizer(batch['text'], padding=True, truncation=True)

train_dataset, test_dataset = load_dataset('imdb', split=['train', 'test'])
print(type(train_dataset))
train_dataset = train_dataset.map(tokenize, batched=True, batch_size=len(train_dataset))
test_dataset = test_dataset.map(tokenize, batched=True, batch_size=len(train_dataset))
train_dataset.set_format('torch', columns=['input_ids', 'attention_mask', 'label'])
test_dataset.set_format('torch', columns=['input_ids', 'attention_mask', 'label'])



def compute_metrics(pred):
    labels = pred.label_ids
    preds = pred.predictions.argmax(-1)
    precision, recall, f1, _ = precision_recall_fscore_support(labels, preds, average='binary')
    acc = accuracy_score(labels, preds)
    return {
        'accuracy': acc,
        'f1': f1,
        'precision': precision,
        'recall': recall
    }

training_args = TrainingArguments(
    output_dir='./results',
    num_train_epochs=1,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=64,
    warmup_steps=500,
    weight_decay=0.01,
    logging_dir='./logs',
)

trainer = Trainer(
    model=model,
    args=training_args,
    compute_metrics=compute_metrics,
    train_dataset=train_dataset,
    eval_dataset=test_dataset
)

trainer.train()

trainer.evaluate()