benchmark/metrics/clip.py

from typing import Dict

import numpy as np
import torch
from PIL import Image
from torchmetrics.multimodal.clip_score import CLIPScore


class CLIPMetric:
    def __init__(self, model_name_or_path: str = "openai/clip-vit-large-patch14"):
        self.device = torch.device(
            "cuda"
            if torch.cuda.is_available()
            else "mps"
            if torch.backends.mps.is_available()
            else "cpu"
        )
        self.metric = CLIPScore(model_name_or_path="openai/clip-vit-large-patch14")
        self.metric.to(self.device)

    @property
    def name(self) -> str:
        return "clip"

    def compute_score(self, image: Image.Image, prompt: str) -> Dict[str, float]:
        image_tensor = torch.from_numpy(np.array(image)).permute(2, 0, 1).float()
        image_tensor = image_tensor.to(self.device)
        score = self.metric(image_tensor, prompt)
        return {"clip": score.item()}