Input size for r50x1?

[PenroseTiles]

What is the input shape for the Resnet 50x1 model? Is it 224x224x3? Thanks a lot for answering!

[anish9]

Based on the BiT hyperrule heuristics if the image size is less than 96px then it is resized to 160px and random crops are applied with the size of 128px. If the size is above 96 then images are resized to 448px and random crops of 384px is applied.

[PenroseTiles]

Yes, that is true. But what is the shape of the input layer?

[anish9]

@PenroseTiles the input has been varied based on the task and the dataset

[wyp19930313]

def get_resolution(original_resolution): """Takes (H,W) and returns (precrop, crop).""" area = original_resolution[0] original_resolution[1] return (160, 128) if area < 9696 else (512, 480)

known_dataset_sizes = { 'cifar10': (32, 32), 'cifar100': (32, 32), 'oxford_iiit_pet': (224, 224), 'oxford_flowers102': (224, 224), 'imagenet2012': (224, 224), }

precrop, crop = bit_hyperrule.get_resolution_from_dataset(args.dataset)
train_tx = tv.transforms.Compose([
    tv.transforms.Resize((precrop, precrop)),
    tv.transforms.RandomCrop((crop, crop)),
    tv.transforms.RandomHorizontalFlip(),
    tv.transforms.ToTensor(),
    tv.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])
val_tx = tv.transforms.Compose([
    tv.transforms.Resize((crop, crop)),
    tv.transforms.ToTensor(),
    tv.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])

[lucasb-eyer]

Like @anish9 and @wyp19930313 say, we apply BiT-HyperRule and so the model's input size depends on the original image's resolution.

"the shape of the input layer" is not a valid concept.