preprocess.py

# preprocess.py
import os
from PIL import Image
from transformers import BlipProcessor, BlipForConditionalGeneration

def process_dataset(zip_path, output_dir, generate_captions=True):
    os.makedirs(output_dir, exist_ok=True)
    
    # Descompacta dataset
    import zipfile
    with zipfile.ZipFile(zip_path, 'r') as zip_ref:
        zip_ref.extractall(output_dir)
    
    # Carrega BLIP (em inglês — modelo oficial da Salesforce)
    processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
    model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")

    # Processa imagens
    for img_name in os.listdir(output_dir):
        if img_name.lower().endswith(('.png', '.jpg', '.jpeg')):
            img_path = os.path.join(output_dir, img_name)
            image = Image.open(img_path).convert('RGB')
            
            # Redimensiona para evitar erros de memória
            image.thumbnail((512, 512), Image.LANCZOS)
            image.save(img_path)  # Salva imagem redimensionada
            
            if generate_captions:
                inputs = processor(image, return_tensors="pt")
                outputs = model.generate(**inputs, max_new_tokens=50)
                caption = processor.decode(outputs[0], skip_special_tokens=True)
                
                txt_path = os.path.splitext(img_path)[0] + ".txt"
                with open(txt_path, "w", encoding="utf-8") as f:
                    f.write(caption)
    
    return output_dir