RuntimeError: dictionary changed size during iteration?

[animitta]

pytorch ddp mode, dataloader wrapper batch data by EasyDict.
line 120, in init for k in self.class.dict.keys():

        # Class attributes
        for k in self.__class__.__dict__.keys():
            if not (k.startswith('__') and k.endswith('__')) and not k in ('update', 'pop'):
                setattr(self, k, getattr(self, k))
```python

[boray-tw]

Please provide a minimal reproduceable code with PyTorch and easydict package versions, or people cannot understand your problem, nor do answer.

I suspect this issue relates to the Python 3 implementation of dict.keys().

In Python 3, dict.keys() returns a view object, while in Python 2 it returns a copy of keys. A view object is similar to a real-time dynamic view of the original object (dictionary in this case). It is an undefined behavior to iterate over a view object (iterator) modified in the flight, and thus causes an RuntimeError.

To enforce the Python 2 behavior, you need to either convert the view object to a list list(dict.keys()) or modify a copy of the original dictionary like this (see the last code block).

Suppose that your attached code is the EasyDict initializer. Could you modify the line 143 from for k in self.__class__.__dict__.keys(): to for k in list(self.__class__.__dict__.keys()):, and see if the error persists?

If no error, could you open a pull request concerning this improvement, as indicated in this PyTorch EasyDict support issue?

[animitta]

OK, I will testing.