请问图像重建，计算出来的psnr值是负数怎么办？

训练验证和测试用的是同一个计算函数，训练验证时psnr值是正常的，到验证时psnr值就变成负数了，社区里类似的问题说是图像值的范围不对，但是我验证集跟测试集用的是同一个提取代码啊，请问是我哪里出了什么问题吗？

    def __getitem__(self, index):

        if self.split == 'train':
            LR_img = cv2.imread(os.path.join(self.LR_dir, self.img_names[index][:-4] + 'x' + str(self.SR_rate) + '.png')) / 255.
            HR_img = cv2.imread(os.path.join(self.HR_dir, self.img_names[index])) / 255.
            if self.augment:
                # random crop
                LR_img, HR_img = random_crop(LR_img, HR_img, self.crop_size, self.SR_rate)

                # geometric transformations
                if random.random() < 0.5:  # hflip
                    LR_img, HR_img = LR_img[:, ::-1, :], HR_img[:, ::-1, :]
                if random.random() < 0.5:  # vflip
                    LR_img, HR_img = LR_img[::-1, :, :], HR_img[::-1, :, :]
                if random.random() < 0.5:  # rot90
                    LR_img, HR_img = LR_img.transpose(1, 0, 2), HR_img.transpose(1, 0, 2)

                # intensity scale
                intensity_scale = random.choice(self.intensity_list)
                LR_img *= intensity_scale
                HR_img *= intensity_scale

       #这里else包括了验证和测试
        else:
            LR_img = cv2.imread(os.path.join(self.LR_dir, self.img_names[index][:-4] + 'x' + str(self.SR_rate) + '.png')) / 255.
            HR_img = cv2.imread(os.path.join(self.HR_dir, self.img_names[index])) / 255.


        # Convert
        LR_img = np.ascontiguousarray(LR_img.transpose(2, 0, 1))  # HWC => CHW
        HR_img = np.ascontiguousarray(HR_img.transpose(2, 0, 1))

        return torch.from_numpy(LR_img), torch.from_numpy(HR_img), self.img_names[index]

        for LR_img, HR_img, img_name in dataloader:
            LR_img, HR_img = LR_img.cuda().float(), HR_img.cuda().float()
            if device != 'cpu':
                torch.cuda.synchronize()
            t0 = time.perf_counter()
            HR_pred = model(LR_img)
            if device != 'cpu':
                torch.cuda.synchronize()
            t1 = time.perf_counter()
            psnr = cal_psnr2(HR_pred,HR_img).item()
            inference_time = t1 - t0

def cal_psnr2(x,y):
    mse = torch.mean((x - y)**2,dim=[1,2,3])
    if mse < 1.0e-10:
        return 100
    score = 20 * torch.log10(1.0 / torch.sqrt(mse))
    return score