Cannot obtain the accuracy stated in the doc for inception_v3 pretrained on Imagenet

Hi.
I'm trying to evaluate inception_v3 pretrained model from the hub on Imagenet (ILSVRC 2012) test set. I use the following evaluation code:

def compute_accuracy(self):
        num_correct = 0
        num_images = 0

        _IMAGE_MEAN_VALUE = [0.485, 0.456, 0.406]
        _IMAGE_STD_VALUE = [0.229, 0.224, 0.225]
        imgnet_loader = torch.utils.data.DataLoader(
            ImageFolder('/home/amin/dataset/ILSVRC/val', 
            transforms.Compose([
                transforms.Resize(299),
                transforms.CenterCrop(299),
                transforms.ToTensor(),
                transforms.Normalize(mean=_IMAGE_MEAN_VALUE, std=_IMAGE_STD_VALUE),
            ])),
            batch_size=16, shuffle=True,
            num_workers=8, pin_memory=True)
        self.model = torch.hub.load('pytorch/vision:v0.10.0', 'inception_v3', pretrained=True).cuda()
        self.model.eval()

        for i, (images, targets) in \
            enumerate(tqdm(imgnet_loader,  desc="Compute Accuracy", total=len(imgnet_loader))):
            images = images.cuda() if torch.cuda.is_available() else images.cpu()
            targets = targets.cuda() if torch.cuda.is_available() else targets.cpu()
            output_dict = self.model(images)
            pred=output_dict.argmax(dim=1)
            
            num_correct += (pred == targets).sum().item()
            num_images += images.size(0)

        classification_acc = num_correct / float(num_images) * 100
        return classification_acc

However, the accuracy I get is 77.216 while it should be 77.45 according to this page. I figured that the model has a transform_input as preprocess in itself. So if we are doing normalization beforehand (as suggested in the example code), we should set transform_input to false. So if I add self.model.transform_input = False, I get 77.472, Which is closer to the expected value but not exactly the same.

Assuming that the issue isn't from my code, I also found the thread on std, mean values (#1439). So I tested some of the values suggested there as well, and got these results:

I appreciate any input on this.
Thanks.

Start by reproducing the reported evaluation with torch.hub.load('pytorch/vision:v0.10.0', 'inception_v3', pretrained=True), the shown ImageFolder transforms, and both transform_input settings. Compare the measured accuracy with the hub documentation and determine which preprocessing and expected result should be considered correct; done means the discrepancy is explained and the behavior or documentation is aligned.