import torch
from PIL import Image
from torchvision import transforms
from torchvision.transforms.functional import InterpolationMode
from BLIP.models.blip import blip_decoder

class BLIPDecoderService:
    _instance = None
    _model = None

    def __new__(cls):
        if cls._instance is None:
            cls._instance = super(BLIPDecoderService, cls).__new__(cls)
            cls.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
            cls.image_size = 384
            
            # Load Model
            model_url = 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model_base_capfilt_large.pth'
            cls._model = blip_decoder(pretrained=model_url, image_size=cls.image_size, vit='base')
            cls._model.eval()
            cls._model = cls._model.to(cls.device)
            
            # Preprocess
            cls.transform = transforms.Compose([
                transforms.Resize((cls.image_size, cls.image_size), interpolation=InterpolationMode.BICUBIC),
                transforms.ToTensor(),
                transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
            ])
        return cls._instance

    def generate_caption(self, pil_image):
        image = self.transform(pil_image).unsqueeze(0).to(self.device)
        with torch.no_grad():
            # Beam search as per your requirement
            caption = self._model.generate(image, sample=False, num_beams=3, max_length=20, min_length=5)
            return caption[0]