Why Tensor shape is not a Tensor?

vision
1

Why is tensor.shape/tensor.size() not a tenosr? what is the need to also have a tensor.Size?
When tracing tensor.shape[0] is a tensor but when not tracing tensor.shape[0] is an int. So I have to do check with isinstance like this:

@torch.jit.script
def center_slice_helper(x, h_offset, w_offset, h_end, w_end):
    return x[:, :, h_offset:h_end, w_offset:w_end]


class CenterCrop(nn.Module):
    def __init__(self, crop_size):
        """Crop from the center of a 4d tensor
        Input shape can be dynamic
        :param crop_size: the center crop size
        """
        super(CenterCrop, self).__init__()
        self.crop_size = crop_size

    def extra_repr(self):
        """Extra information
        """
        return 'crop_size={}'.format(
            self.crop_size
        )

    def forward(self, x):
        h_offset = (x.shape[2] - self.crop_size) / 2
        w_offset = (x.shape[3] - self.crop_size) / 2
        if not isinstance(h_offset, torch.Tensor):
            h_offset, w_offset = torch.tensor(h_offset), torch.tensor(w_offset)
        h_end = h_offset + self.crop_size
        w_end = w_offset + self.crop_size
        return center_slice_helper(x, h_offset, w_offset, h_end, w_end)

If tensor.shape was a tensor we could use it like other tensors in the network

1 Like