app.py

#허깅페이스에서 돌아갈 수 있도록 바꾸어 보았음

import torch
from transformers import BertTokenizerFast, BertForQuestionAnswering, Trainer, TrainingArguments
from datasets import load_dataset
from collections import defaultdict

# 데이터 불러오기
dataset_load = load_dataset('Multimodal-Fatima/OK-VQA_train')
dataset = dataset_load['train'].select(range(300))

# 불필요한 특성 선택
selected_features = ['image', 'answers', 'question']
selected_dataset = dataset.map(lambda ex: {feature: ex[feature] for feature in selected_features})

# 소프트 인코딩
answers_to_id = defaultdict(lambda: len(answers_to_id))
selected_dataset = selected_dataset.map(lambda ex: {
    'answers': [answers_to_id[ans] for ans in ex['answers']],
    'question': ex['question'],
    'image': ex['image']
})

id_to_answers = {v: k for k, v in answers_to_id.items()}
id_to_labels = {k: ex['answers'] for k, ex in enumerate(selected_dataset)}

selected_dataset = selected_dataset.map(lambda ex: {'answers': id_to_labels.get(ex['answers'][0]),
                                                   'question': ex['question'],
                                                   'image': ex['image']})

flattened_features = []

for ex in selected_dataset:
    flattened_example = {
        'answers': ex['answers'],
        'question': ex['question'],
        'image': ex['image'],
    }
    flattened_features.append(flattened_example)

# 모델 가져오기
from transformers import AutoModelForSequenceClassification, TrainingArguments, Trainer

model_name = 'microsoft/git-base-vqav2'
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Trainer를 사용하여 모델 학습
tokenizer = BertTokenizerFast.from_pretrained('bert-base-multilingual-cased')

def preprocess_function(examples):
    tokenized_inputs = tokenizer(examples['question'], truncation=True, padding=True)
    return {
        'input_ids': tokenized_inputs['input_ids'],
        'attention_mask': tokenized_inputs['attention_mask'],
        'pixel_values': [(4, 3, 244, 244)] * len(tokenized_inputs['input_ids']),
        'pixel_mask': [1] * len(tokenized_inputs['input_ids']),
        'labels': [[label] for label in examples['answers']]
    }

dataset = load_dataset("Multimodal-Fatima/OK-VQA_train")['train'].select(range(300))
ok_vqa_dataset = dataset.map(preprocess_function, batched=True)
ok_vqa_dataset.set_format(type='torch', columns=['input_ids', 'attention_mask', 'pixel_values', 'pixel_mask', 'labels'])

training_args = TrainingArguments(
    output_dir='./results',
    num_train_epochs=20,
    per_device_train_batch_size=4,
    logging_steps=500,
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=ok_vqa_dataset
)

# 모델 학습
trainer.train()