batch size 1 throws index error

[alex-de-zwart]

I'd like to benchmark different models also in terms of latency, but when I use a batch size of 1 it torch geometric message passing throws the following error:

Dimension out of range (expected to be in range of [-1, 0], but got -2)
Edge_index is created with the following code

class GraphDataset(pyg_data.InMemoryDataset):
def init(self, root, file_name, transform=None, pre_transform=None):
self.file_name = file_name
super().init(root, transform, pre_transform)
self.data, self.slices = torch.load(self.processed_paths[0])

@property
def raw_file_names(self):
    return [f'{self.file_name}.txt']

@property
def processed_file_names(self):
    return [f'{self.file_name}.pt']

def download(self):
    pass

def process(self):
    raw_data_file = f'{self.raw_dir}/{self.raw_file_names[0]}'
    print(raw_data_file)
    with open(raw_data_file, 'rb') as f:
        sessions = pickle.load(f)
    data_list = []

    for session in range(len(sessions[0])):
      session, y = sessions[0][session], sessions[1][session]
      codes, uniques = pd.factorize(session)
      senders, receivers = codes[:-1], codes[1:]

      # Build Data instance
      edge_index = torch.tensor([senders, receivers], dtype=torch.long)
      x = torch.tensor(uniques, dtype=torch.long).unsqueeze(1)
      y = torch.tensor([y[0]], dtype=torch.long)
      data_list.append(pyg_data.Data(x=x, edge_index=edge_index, y=y))
    
    data, slices = self.collate(data_list)
    torch.save((data, slices), self.processed_paths[0])

[rusty1s]

This likely indicates that x is only one-dimensional (although I see that you explicitly unsqueeze(1) already). Can you still confirm that this is not the case?

[alex-de-zwart]

Hi @rusty1s thanks for the prompt reply much appreciated!

Before x is passed to a gru cell it's squeezed again, similar to what happens in this notebook:

https://colab.research.google.com/drive/1XwQ0njqSZL8vbHJMnRRHlH4ar0kYYFVz?usp=sharing#scrollTo=TNsAHq8PCN3k

[rusty1s]

Ok, what does that mean? Whenever you pass x to a MessagePassing module, it should be two-dimensional.