PyG dataloder with previously stored data

Hello everyone!

I’m working with a really huge dataset of small graphs. What I’ve done so far is creating a Dataset class that stores the graph objects in a processed folder. What I want is to be able to read those files with a Dataset class without generate them again. How can I do it? It seems pretty silly but I’m wasting a lot of time and I don’t find the way.

The idea is to generate the Data() objects if they haven’t been previously generated and just read it if they already exists. I’m trying to do so creating a Dataset without download and process but it doesn’t work.

Thank you!

I don’t know which format is used to store the generated torch_geometric data but assume you can also load each sample with torch_geometric in a similar way.
If so, I would guess you could write a custom Dataset as described here and load each sample in the __getitem__.


Thank you for your answer! I’ve tried that but it doesn’t work because a standard Pytorch Dataset needs just tensor data, it’s not able to handle format.
I’m now trying to do it with a PyG Dataset, that can obviously handle PyG.Data objects.


In case someone is curious, what I’ve finally done is to load the graph files without any dataset, just like it appears on the PyG tutorial.

from import Data
from torch_geometric.loader import DataLoader

data_list = [Data(…), …, Data(…)]
loader = DataLoader(data_list, batch_size=32)