benchmark/metrics/clip_iqa.py

from typing import Dict

import numpy as np
import torch
from PIL import Image
from torchmetrics.multimodal import CLIPImageQualityAssessment


class CLIPIQAMetric:
    def __init__(self):
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.metric = CLIPImageQualityAssessment(
            model_name_or_path="clip_iqa",
            data_range=255.0,
            prompts=("quality",)
        )
        self.metric.to(self.device)

    @property
    def name(self) -> str:
        return "clip_iqa"
    
    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.unsqueeze(0)
        image_tensor = image_tensor.to(self.device)
        scores = self.metric(image_tensor)
        return {"clip_iqa": scores.item()}