What does the collate function in pytorch (geometric)?

Question

I am creating a message passing neural network and have some issues with the dataset creation. In pytorch (geometric) it is recommended to create a dataset with the following class. I wonder what is the meaning of the collate function that is called at the end of the process method? In what cases should I use my own collate function? My graphs have mostly different sizes.

import torch
from torch_geometric.data import InMemoryDataset


class MyOwnDataset(InMemoryDataset):
    def __init__(self, root, transform=None, pre_transform=None):
        super(MyOwnDataset, self).__init__(root, transform, pre_transform)
        self.data, self.slices = torch.load(self.processed_paths[0])

    @property
    def raw_file_names(self):
        return ['some_file_1', 'some_file_2', ...]

    @property
    def processed_file_names(self):
        return ['data.pt']

    def download(self):
        # Download to `self.raw_dir`.

    def process(self):
        # Read data into huge `Data` list.
        data_list = [...]

        if self.pre_filter is not None:
            data_list = [data for data in data_list if self.pre_filter(data)]

        if self.pre_transform is not None:
            data_list = [self.pre_transform(data) for data in data_list]

        data, slices = self.collate(data_list)
        torch.save((data, slices), self.processed_paths[0])

The error I get is:

RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 0. Got 4422 and 4032 in dimension 1 at /opt/conda/conda-bld/pytorch_1573049304260/work/aten/src/THC/generic/THCTensorMath.cu:71

Answer 1

What collate does and why:

Because saving a huge python list is really slow, we collate the list into one huge torch_geometric.data.Data object via torch_geometric.data.InMemoryDataset.collate() before saving . The collated data object has concatenated all examples into one big data object and, in addition, returns a slices dictionary to reconstruct single examples from this object. Finally, we need to load these two objects in the constructor into the properties self.data and self.slices.

Source