inference.py

import torch
import pickle
from PIL import Image
import torchvision.transforms as T
import os
import sys

# # Ensure project root is on sys.path
# PROJECT_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__)))
# if PROJECT_ROOT not in sys.path:
#     sys.path.insert(0, PROJECT_ROOT)

# # Backward compatibility: vocab.pkl was saved when Vocabulary lived at
# # 'data_processing_vocabulary'. Register an alias so pickle can find it
# # at its new location 'src.data_processing_vocabulary'.
# from src import data_processing_vocabulary
# sys.modules["data_processing_vocabulary"] = data_processing_vocabulary

from image_captioning_model import ImageCaptioningModel


DEVICE = "cpu"


class CaptionGenerator:

    def __init__(
        self,
        model_path="best_model.pth",
        vocab_path="vocab.pkl",
        use_vit=True
    ):
        
        print("Loading vocab...")
        print("Building model...")
        print("Loading state dict...")


        # Load vocab
        with open(vocab_path, "rb") as f:
            self.vocab = pickle.load(f)
        
        print(f"Vocab loaded with {len(self.vocab)} words.")

        # Build model
        self.model = ImageCaptioningModel(
            vocab_size=len(self.vocab),
            pad_id=self.vocab.word2idx["<pad>"],
            use_vit=use_vit
        )
        print("Model built.")

        # Load weights
        state_dict = torch.load(model_path, map_location=DEVICE)
        self.model.load_state_dict(state_dict)
        print(f"Model weights loaded from {model_path}.")

        # Eval mode
        self.model.eval()
        self.model.to(DEVICE)

        # Preprocess
        self.transform = T.Compose([
            T.Resize((224, 224)),
            T.ToTensor(),
            T.Normalize(
                mean=[0.485, 0.456, 0.406],
                std=[0.229, 0.224, 0.225]
            )
        ])


    def preprocess(self, image):

        if not isinstance(image, Image.Image):
            image = Image.open(image).convert("RGB")

        image = self.transform(image)
        return image.unsqueeze(0)


    def generate(self, image, max_len=30, decoding="beam", beam_width=5):
        """
        Generate a caption for the given image.

        Args:
            image: PIL Image or file path.
            max_len: Maximum caption length.
            decoding: 'greedy' or 'beam'.
            beam_width: Number of beams (only used when decoding='beam').

        Returns:
            Generated caption string.
        """
        image = self.preprocess(image).to(DEVICE)

        if decoding == "beam":
            caption = self.model.predict_caption_beam(
                image=image,
                vocab=self.vocab,
                beam_width=beam_width,
                max_len=max_len,
                device=DEVICE
            )
        else:
            caption = self.model.predict_caption(
                image=image,
                vocab=self.vocab,
                max_len=max_len,
                device=DEVICE
            )

        return caption