How to convert MNIST dataset into PyG dataset?

I want to convert MNIST dataset into PYG dataset. The idea behind this to replicate Defferrard et al’s result (https://arxiv.org/pdf/1606.09375.pdf) [Table 1. Proposed graph CNN GC32-P4-GC64-P4-FC512]. I have an Adjacency matrix which is 784 X 784. The adjacency matrix is here: https://bit.ly/2SIKPgH, which can be converted to edge list as follows as per @rusty1s suggestion:

import scipy.io
adjt8 = scipy.io.loadmat('adjt8.mat')
adjt8 = adjt8['Expression1']
from scipy import sparse
adj81 = sparse.coo_matrix(adjt8)
row = torch.from_numpy(adj81.row).long()
col = torch.from_numpy(adj81.col).long()
edge_index = torch.stack([row,col],dim=0)

#Importing MNIST Data
!pip install spektral
from spektral.datasets import mnist
X_train, y_train, X_val, y_val, X_test, y_test, A = mnist.load_data()

Creating MNIST dataset:

from torch_geometric.data import InMemoryDataset

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

    @property
    def raw_file_names(self):
        return []

    @property
    def processed_file_names(self):
        return ['/tmp/mdata.pt']

    def download(self):
        pass

    def process(self):

        data_list= []
        for i in range(0, len(X_train)):
            data = Data(x=torch.tensor(X_train[i],dtype=torch.float).view(784,1),edge_index=edge_index1,y=torch.tensor(y_train[i]).view(1).long())
            data_list.append(data)

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

I am to trying to implement something similar to MNIST superpixel graclus example (https://bit.ly/3bdlQc6) that @rusty1s implemented. However, for my simple MNIST there is going to be only one edge_index and no edge_attr. I also do not have pos. For normalized cut and graclus I need pos. The superpixel MNIST dataset comes with pos and edge_attr.

How do I go about creating MNIST dataset for PyG?

[jasperhyp]

Sure, the flattening approach should work better compared to a non-hierachical global pooling variant. You can do this here too, of course. The downside of the flattening approach is that you lose translation invariance, similar to an MLP compared to a CNN.

Hi Matthias, Thanks for creating both PyG and Spektral. I have a question related to implementing Flatten() in PyTorch here… I want to add a Flatten layer instead of some pooling layer after a GCNConv, but since I am using batches to train (train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle = True)), the torch.flatten() function would mix all batches together, and torch.nn.Flatten() (layer) would not actually decrease the dimension of tensor as in Spektral, where 3-D tensor [batch_size, node_num, node_atttribute_num] is converted to 2-D [batch_size, node_num*node_attribute_num]. It might not be a problem directly related to Spketral or PyG (indeed, those are built-in functions of torch and keras), but I am wondering if you can help me with this issue? If I just go with torch.flatten(), the final output would only contain one result for batch_size of training samples… Thanks!

[rusty1s]

See https://github.com/rusty1s/pytorch_spline_conv/issues/25