Method compute

utils/torchmetric_kid.py:247–273

Calculate KID score based on accumulated extracted features from the two distributions. Returns a tuple of mean and standard deviation of KID scores calculated on subsets of extracted features. Implementation inspired by `Fid Score`_

(self)

            self.fake_features.append(features)

    def compute(self) -> Tuple[Tensor, Tensor]:
        """Calculate KID score based on accumulated extracted features from the two distributions. Returns a tuple
        of mean and standard deviation of KID scores calculated on subsets of extracted features.

        Implementation inspired by `Fid Score`_
        """
        real_features = dim_zero_cat(self.real_features)
        fake_features = dim_zero_cat(self.fake_features)

        n_samples_real = real_features.shape[0]
        if n_samples_real < self.subset_size:
            raise ValueError("Argument `subset_size` should be smaller than the number of samples")
        n_samples_fake = fake_features.shape[0]
        if n_samples_fake < self.subset_size:
            raise ValueError("Argument `subset_size` should be smaller than the number of samples")

        kid_scores_ = []
        for _ in range(self.subsets):
            perm = torch.randperm(n_samples_real)
            f_real = real_features[perm[: self.subset_size]]
            perm = torch.randperm(n_samples_fake)
            f_fake = fake_features[perm[: self.subset_size]]

            o = poly_mmd(f_real, f_fake, self.degree, self.gamma, self.coef)
            kid_scores_.append(o)
        kid_scores = torch.stack(kid_scores_)
        return kid_scores.mean(), kid_scores.std(unbiased=False)

    def reset(self) -> None:
        if not self.reset_real_features: