How to batch graphs with other normal data?

[595693085]

Hi,
my data looks like:

data_list=[
[graph_1a, graph_1b, seq_1a, seq_1b, seq_1c, y1], 
[graph_2a, graph_2b, seq_2a, seq_2b, seq_2c, y2],
....,
[graph_na, graph_nb, seq_na, seq_nb, seq_nc, yn]
],

which erery data consists of five parts (two graphs and three sequences) and the corresponding label.
The two graphs were input two GNNs, and the three sequences were input three CNNs, then the five hidden vectors were concatenated to predict the label.

The following codes are my implementations.

from torch_geometric.data import Data
from torch_geometric.nn import GCNConv, global_max_pool
from torch_geometric.loader import DataLoader
import torch.nn as nn

# dataset
class PairData(Data):
    def __inc__(self, key, value, *args, **kwargs):
        if key == 'edge_index_s':  # graph1
            return self.x_s.size(0)
        if key == 'edge_index_t':  # graph2
            return self.x_t.size(0)
        return super().__inc__(key, value, *args, **kwargs)
#the PairData implementation is followed the official documentation in [link](https://pytorch-geometric.readthedocs.io/en/latest/notes/batching.html)

class MyDataset(Dataset):
     def __init__(self, data, ...)
         self.process(...)
     def process(self, data,...)
          data_list=[]
          for i in range(len(data)):
                ....#some processings
                pari_data = PairData(x_s=x_s, # graph1
                                     edge_index_s=edge_index_s, # graph1
                                     x_t=x_t, # graph2
                                     edge_index_t=edge_index_t, # graph2
                                     y=labels)
                pari_data.seq1=seq1
                pari_data.seq2=seq2
                pari_data.seq3=seq3
                data_list.append(pari_data)
          self.data_list=data_list
     def __getitem__(self, index):
           return self.data_list[index]

#model 
class MyModel(nn.Module)
     def __init__(self, ...)
          self.gnn1=GCNConv(...)
          self.gnn2=GCNConv(...)
          self.cnn1=nn.Conv1d(...)
          self.cnn2=nn.Conv1d(...)
          self.cnn3=nn.Conv1d(...)
          ...
     def forward(self, data)
          x_s, edge_index_s, x_s_batch = data.x_s, data.edge_index_s, data.x_s_batch
          x_t, edge_index_t, x_t_batch = data.x_t, data.edge_index_t, data.x_t_batch
          seq1= data.seq1
          seq2= data.seq2
          seq3= data.seq3
         x1=self.gnn1(x_s,edge_index_s)
         x1=global_max_pool(x1,x_s_batch )
         x2=self.gnn2(x_t,edge_index_t)
         x2=global_max_pool(x_t,x_t_batch)
         x3=self.cnn1(seq1)
         x4=self.cnn2(seq2)
         x5=self.cnn3(seq3)
         x=torch.cat(x1,x2,x3,x4,x5 ....)
         ....
         prediction= ...
         return prediction

#for training
....
dataset=MyDataset(...,data)
dataloader=DataLoader(dataset, batch_size=16, shuffle=True, num_workers=4, follow_batch=["x_s", "x_t"])
model=MyModel(...)

for idx,batch_data in enumerate(dataloader):
    output=model(batch_data)
    ....

But when I run the code, there is an error:

File "/home/ubuntu/.conda/envs/***/lib/python3.11/site-packages/...", line 974, in forward
    batch_size, seq_length = input_shape
ValueError: not enough values to unpack (expected 2, got 1)

The error menas that the input sequences are not in a batch_size shape (batch_size,d), the shape of them are:

seq1.size()===>(d*batch_size)
seq2.size()===>(d*batch_size)
seq3.size()===>(d*batch_size)

So the sequences could not be fit into the CNNs.

So how to batch the data, which every data consisits of multiple graphs and mutiple normal matrix.

[wsad1]

Try updating the __cat_dim__ function in your Data class. As described here.