lazy_instance could not be applied for torch.nn.Module

[quancs]

code:

>>> import torch.nn as nn
>>> from jsonargparse import lazy_instance
>>> lazy_instance(nn.Module)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/quancs/miniconda3/lib/python3.8/site-packages/jsonargparse/typehints.py", line 561, in lazy_instance
    return LazyInitClass(class_type, kwargs)
  File "/home/quancs/miniconda3/lib/python3.8/site-packages/jsonargparse/typehints.py", line 520, in __init__
    for name, method in inspect.getmembers(self, predicate=inspect.ismethod):
  File "/home/quancs/miniconda3/lib/python3.8/inspect.py", line 334, in getmembers
    names = dir(object)
  File "/home/quancs/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1557, in __dir__
    parameters = list(self._parameters.keys())
  File "/home/quancs/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 947, in __getattr__
    raise AttributeError("'{}' object has no attribute '{}'".format(
AttributeError: 'LazyInitClass' object has no attribute '_parameters'

[quancs]

change the __dir__ of a subclass of Module to

    def __dir__(self):
        try:
            return super().__dir__()
        except:
            return ['__init__', 'forward']

can ignore the exception

[mauvilsa]

Thanks for reporting. Indeed it is not working and after trying a few things I still don't understand what is the underlying issue. Implementing __dir__ is not a solution since you would need to manually give all possible methods that could be called with an instance. Will keep looking at how to fix it.

[mauvilsa]

I have pushed a fix to master. The lazy_instance idea isn't as simple as I initially thought. There might be classes for which it still does not work. Specific issues can be created if this is the case. For now closing this one. The fix will be included in version 4.0.0 scheduled for next week.