[JIT] Improve error messaging for using a tensor attribute in ScriptModule

[sidazhang]

🐛 Bug

ScriptModule is not able to have a tensor as an attribute

To Reproduce

Not Constant

import torch
class ConstantTensor(torch.jit.ScriptModule):
    def __init__(self):
        self.cx = torch.ones(100, 100, dtype=torch.float, device='cuda')

    @torch.jit.script_method
    def forward(self, x):
        return x + self.cx


c = ConstantTensor()
print c.graph

attribute 'cx' of type 'Tensor' is not usable in a script method (did you forget to add it __constants__?):

Constant

import torch
class ConstantTensor(torch.jit.ScriptModule):
    __constants__ = ['cx']
    def __init__(self):
        self.cx = torch.ones(100, 100, dtype=torch.float, device='cuda')

    @torch.jit.script_method
    def forward(self, x):
        return x + self.cx


c = ConstantTensor()
print c.graph

TypeError: 'Tensor' object for attribute 'cx' is not a valid constant.
Valid constants are:
  1. a nn.ModuleList
  2. a value of type {bool, float, int, str, NoneType, function, device, layout, dtype}
  3. a list or tuple of (2)

UPDATE:

Seems like the right way to use a tensor attribute is to declare it to be a buffer

self.register_buffer('cx', torch.ones(100, 100, dtype=torch.float, device='cuda'))

Please improve the error message when a user tries to use a tensor as an attribute

[sidazhang]

It seems like that the correct thing to do is

self.register_buffer('cx', torch.ones(100, 100, dtype=torch.float, device='cuda'))

In this case, could you improve the error messaging so it is explicit?