When 'RuntimeError: The 'data' object was created by an older version of PyG.' evokes?

[songsong0425]

Dear PyG team,
Greetings, I have a question about the RuntimeError.
While running the code from paper, I counted 'RuntimeError: The 'data' object was created by an older version of PyG' error.
I confirmed the link https://github.com/pyg-team/pytorch_geometric/blob/master/torch_geometric/data/data.py that looks related to this error, but I could not understand when this error evoke.
As described in the comment, I checked the dataset's root folder to remove the 'processed/' directory, but it didn't exist in my folder anywhere.
My PyG version is 2.0.3 and running with jupyter notebook, python 3.9.7, and pytorch 1.10.0.
Can anyone deal with this error?
Sincerely,
Songyeon

[rusty1s]

The data format changed slightly in torch-geometric==2.0. Thus, we cannot longer load in the data that were created by previous PyG versions. I am really sorry for the inconveniences! Once thing you can do is to re-process your to fit into the PyG 2.0 format. That is:

# Install PyG 1.7.2
data = torch.load('data.pt').to_dict() 
torch.save(data, 'data_dict.pt')

# Install PyG 2.0
data_dict = torch.load('data_dict.pt')
data = Data.from_dict(data_dict)
torch.save(data, 'data.pt')

[rusty1s]

Try

data, slices = torch.load('data.pt')
data_dict = data.to_dict()

[eason1021]

Dear rusty1s,

It occurs new attribute error: AttributeError: 'Data' object has no attribute 'to_dict'. Did you mean: 'from_dict'?

I think this is weird, I also print the data, here is the information:
Data(edge_index=[2, 184764384], test_mask=[26394912], train_mask=[26394912], val_mask=[26394912], x=[26394912, 21], y=[26394912])

[rusty1s]

Data.to_dict() should exist in 1.7.2, see https://github.com/pyg-team/pytorch_geometric/blob/1.7.2/torch_geometric/data/data.py#L113.

However, you can also do data_dict = {key: data[key] for key in data}.

[eason1021]

Dear rusty1s,

Thanks, the code execute successfully in PyG 1.7.2 with
data, slices = torch.load('data.pt')
data_dict = data.to_dict()

However, when I do the following code in PyG 2.0:
data_dict = torch.load('data_dict.pt')
data = Data.from_dict(data_dict)
torch.save(data, 'data.pt')

The code can execute with no error message, however, I found I could not print the dataset information with new pt file and PyG Inmemory Dataset function:

If I print(len(dataset)), the error message would appear: AttributeError: 'tuple' object has no attribute 'items'

[rusty1s]

Wouldn't you want to save the tuple instead of the data object?

torch.save((data, slices), 'data.pt')

[tonyriverms]

A possible simpler solution from my experience might be directly reading from __dict__. All data seem actually saved there. In this way the Data object seems now compatible with the old version at least for reading in my use case.

For example,

from torch_geometric.data.data import Data
from typing import Mapping

def is_new_data_object(data: Data):
    return '_store' in data.__dict__

data: Data = pickle.loads(self.env.begin().get(f"{idx}".encode()))
if not is_new_data_object(data):
    data: Mapping = data.__dict__
return dict(
            pos=torch.as_tensor(data["pos"]).float(),
            pos_relaxed=torch.as_tensor(data["pos_relaxed"]).float(),
            cell=torch.as_tensor(data["cell"]).float().view(3, 3),
            atoms=torch.as_tensor(data["atomic_numbers"]).long(),
            tags=torch.as_tensor(data["tags"]).long(),
            relaxed_energy=data["y_relaxed"],  # python float
   )