[BUG] Defining a __setattr__ funcion inside a @tensorclass creates weird problems

[alex-bene]

Describe the bug

Defining a __setattr__ function inside a @tensorclass seems to make the class unable to traverse its fields in some cases (the fields are not shown when printing an instance).
(Probably) related to this is that when defining __setattr__, auto_batch_size_ and auto_device_ do not work (I suppose it uses the same logic to traverse the fields as print)

To Reproduce

import torch

@tensorclass()
class PoC:
    boring_variable: Tensor | None = None


@tensorclass()
class PoC_bug:
    boring_variable: Tensor | None = None

    def __setattr__(self, name, value):
        super().__setattr__(name, value)


poc = PoC()
poc_bug = PoC_bug()
poc.boring_variable = torch.rand(2, 3)
poc_bug.boring_variable = torch.rand(2, 3)

print(poc)
"""
PoC(
    boring_variable=Tensor(shape=torch.Size([2, 3]), device=cpu, dtype=torch.float32, is_shared=False),
    batch_size=torch.Size([]),
    device=None,
    is_shared=False)
"""
print(poc_bug)
"""
PoC_bug(
,
    batch_size=torch.Size([]),
    device=None,
    is_shared=False)
"""

assert poc.batch_size == torch.Size([])
assert poc_bug.batch_size == torch.Size([])
assert poc.device is None
assert poc_bug.device is None

poc.auto_batch_size_(1)
poc_bug.auto_batch_size_(1)
assert poc.batch_size == torch.Size([2])
assert poc_bug.batch_size == torch.Size([])

poc.auto_device_()
poc_bug.auto_device_()
assert poc.device == torch.device("cpu")
assert poc_bug.device is None

Expected behavior

PoC and PoC_bug classes should behave the same.

Reason and Possible fixes

If you know or suspect the reason for this bug, paste the code lines and suggest modifications.

Checklist

• I have checked that there is no similar issue in the repo (required)
• I have read the documentation (required)
• I have provided a minimal working example to reproduce the bug (required)

[vmoens]

Likely because super will call the setattr of the parent class which is object (tensorclass does not inherit). That should work with TensorClass though!