DataLoader doesn't support batches composed of graph with different edge_index

Hello,

l would like to build a DataLoader from my custom dataset. Each sample has its own adjacency matrix. The latter varies in terms of the number of nodes and connectivity from one sample to another. As a consequence, each sample has its own edge_index and edge_weight.

When l use dataLoader with batch_size=1 it works. However when batch_size > 1 l get the following error :

from torch_geometric.data import DataLoader
loader = DataLoader(train_folder, batch_size=2, shuffle=True)
batch=next(iter(loader))

***** RuntimeError: stack expects each tensor to be equal size, but got [2, 107050] at entry 0 and [2, 106190] at entry 1**

Is DataLoader not supporting varying edge_index ?

[rusty1s]

1. No, it should only be applied to indices.
2. Yes, that’s what DataLoader is doing out of the box (your third point results in wrong behaviour)

[pinkfloyd06]

Thank you for the links and details.

1/ cumsum is applied also on edge_weight ?

2/

train_folder=[ ]

for data_sample in data_list:

                           features,coordinates,edge_index,edge_weight,targets=data_sample

                           train_folder.append(
                            torch_geometric.data.Data(
                            x=features,
                            pos=coordinates, 
                            edge_index=edge_index, 
                            edge_attr=edge_weight,
                            y=targets, 
                            )
                            )

train_loader = torch_geometric.data.DataLoader(train_folder, batch_size=batch_size_train, shuffle=True)

Do you think that my edge_index are accumulated correctly in 2/ or l need to consider cumsum as follow in 3/ ?

3/

train_folder=[ ]

cumsum = 0

for data_sample in data_list:

                           features,coordinates,edge_index,edge_weight,targets=data_sample

                           num_nodes=coordinates.shape[0]

                           cumsum += num_nodes

                           edge_index = edge_index + cumsum 

                           train_folder.append(
                            torch_geometric.data.Data(
                            x=features,
                            pos=coordinates, 
                            edge_index=edge_index, 
                            edge_attr=edge_weight,
                            y=targets, 
                            )
                            )

train_loader = torch_geometric.data.DataLoader(train_folder, batch_size=batch_size_train, shuffle=True)

4/ It seems to me that cumsum plays a role of mask. It control the number of nodes per samples involved in matrix multiplications, in order not to confuse/mix the nodes between the different samples.

Than you