RecursionError when used on subclass of Tensor

[Varal7]

When using subclasses of Tensor, make_dot crashed with RecursionError.

class DummyTensor(torch.Tensor):
    pass

a = DummyTensor(torch.ones(5, requires_grad=True))
make_dot(a)

Output:

/usr/local/lib/python3.7/dist-packages/torchviz/dot.py in add_base_tensor(var, color)
    152             dot.edge(str(id(var.grad_fn)), str(id(var)))
    153         if var._is_view():
--> 154             add_base_tensor(var._base, color='darkolivegreen3')
    155             dot.edge(str(id(var._base)), str(id(var)), style="dotted")
    156 

RecursionError: maximum recursion depth exceeded in __instancecheck__

The reason is because a is a view, so is a._base, so is a._base._base, so is a._base._base._base, etc.

https://colab.research.google.com/drive/1VlAEQY5lOtp3MsYfgyCkrNmkjzuKoRX1?usp=sharing

[albanD]

I'm confused, why is a a view here?

[Varal7]

So the answer is: a is a view of torch.ones(5).

This actually seems like a Pytorch bug.

a = torch.ones(1)
b = torch.Tensor(a)
print(b._is_view()) # True
print(b._base is a) # True

a = torch.ones(1)
c = DummyTensor(a)
print(c._is_view()) # True
print(c._base is a) # False
print(type(c._base)) # <class '__main__.DummyTensor'>

c._base here creates a new DummyTensor, as per the docs

As of version 1.7.0, methods on torch.Tensor and functions in public torch.* namespaces applied on torch.Tensor subclasses will return subclass instances instead of torch.Tensor instances:

[ezyang]

Yeah, agreed this is a bug.

[ezyang]

As a workaround, do the _base object within the DisableTorchFunction context manager:

with torch._C.DisableTorchFunction():
  base = c._base

[Varal7]

This is indeed fixed by pytorch/pytorch#60464.