Neighborloader with torch_geometric.data.Dataset

[MATTEMAINA]

🚀 Feature

Implement Neighborloader that accept not only one Data object at time, but a Dataset object too. The goal is to manage entire dataset with only one call of neighborloader instead manage each datum of dataset every time

Motivation

I working with temporal graphs, and i use so much The Neighborloader to separate my graph in training and validation, but it can process only one Data object at time instead of a Dataset (since i have snapshots of a evolving graph)

[rusty1s]

I'm happy to learn more about your use-case. The current strategy for this is to convert your dataset into a single data object and pass it to NeighborLoader, i.e.

from torch_geometric.data import Batch
from torch_geometric.loader import NeighborLoader

data = Batch.from_data_list(dataset)
loader = NeighborLoader(data, ...)

[MATTEMAINA]

i tried this approach, but i could notice that i cannot more use the 'input_nodes' parameter, i use it to define the set of nodes to use in training and validation. how have i to modify the parameter 'input_nodes' for having the same result?

[rusty1s]

That depends. If training, validation nodes are described for each data objects via masks, then these masks can be used on the Batch object as well:

data = Batch.from_data_list(dataset)
train_loader = NeighborLoader(data, input_nodes=data.train_mask, ...)

If you have designated train_dataset and val_dataset, you do not need to make use of input_nodes at all:

train_loader = NeighborLoader(Batch.from_data_list(train_dataset), ...)
val_loader = NeighborLoader(Batch.from_data_list(val_dataset), ...)

[MATTEMAINA]

Perfect if i understand correctly, Thank this approach the NeighborLoader return a batch of single objact Data as if i use Dataloader(dataset) ?

[rusty1s]

It will return a subset of your graph data. As such, the returned data object is of type Data rather than of type Batch. The only real difference between the two is the missing batch vector.