Infinite recursion when wrapping torch.nn.Module via `__getattr__`

1

I wrote a simple wrapper to add special methods to a given neural network.
While the below implementation works well for general objects e.g. strings, lists etc. I get a RecursionError when applying it to a torch.nn.Module. It seems that in the latter case the call self.instance inside the __getattr__ method is unsuccessful which hence falls back on __getattr__ again leading to the infinite loop (I also tried self.__dict__['instane'] without luck).

What is going on here? What is the correct way of doing what I want?

import torch

class MyWrapper(torch.nn.Module):
    def __init__(self, instance):
        super().__init__()
        self.instance = instance

    def __getattr__(self, name):
        print("trace", name)
        return getattr(self.instance, name)

# Working example
obj = "test string"
obj_wrapped = MyWrapper(obj)
print(obj_wrapped.split(" ")) # trace split\n ['test', 'string']

# Failing example
net = torch.nn.Linear(12, 12)
net.test_attribute = "hello world"
b = MyWrapper(net)

print(b.test_attribute) # RecursionError: maximum recursion depth exceeded
b.instance # RecursionError: maximum recursion depth exceeded