Understanding how to batch multiple graphs with the same number of nodes.

[AlexNiko47]

Hi,

I'm trying to implement a graph classification model and I'm struggling to understand how to properly batch a list of graphs and feed them into the model. All my graphs are stored in a list and are of the form Data(x=[978], edge_index=[2, 956484], edge_attr=[956484], y=[1, 38]) and I'm first trying to implement a one layer linear model to test whether my approach works. Here's the code:

Loader = DataLoader(list_of_graphs,batch_size=2, shuffle=True)
LinearModel(torch.nn.Module): def __init__(self, in_channels: int, out_channels: int): super().__init__() self.linear = Linear(in_channels,out_channels) def forward(self,x): x = self.linear(x) out = torch.sigmoid(x) return out

def train(model, optimizer, train_loader, device, loss): model.train() for data in train_loader: optimizer.zero_grad() data = data.to(device) out = model(data.x) loss_tensor = loss(out, data.y.to(torch.float)) print("Loss is: ", loss_tensor.item()) loss_tensor.backward() optimizer.step()

model = LinearModel(978,38) optimizer = torch.optim.Adam(model.parameters(), lr=0.005, weight_decay=0) loss = torch.nn.MSELoss() train(model,optimizer,Loader,"cpu", loss)

For some reason whenever I try to run this I get the error RuntimeError: mat1 and mat2 shapes cannot be multiplied (1x1956 and 978x38) with 1956 being I'm guessing the two graphs from the batch. What is it I'm donig wrong?

[EdisonLeeeee]

Hi @AlexNiko47

In Data, x must have at least 2 dims, indicating [num_nodes, num_features]. A DataLoader would stack the node feature x at the first dimension from each batch by default. So you got 1x1956 as input with batch size 2.

Now I have a question, does 978 denote the feature dimension or not? If it does, there must something wrong with your data as it has only one node in each batch.

[AlexNiko47]

Hi @EdisonLeeeee
I have a gene expression dataset, where there are 978 genes so I would want to have a [978,1] dimension matrix with each gene being a node with 1 feature (the associated expression value for that gene). I create these Data objects from a pandas DataFrame like this:
list_of_graphs = [] for i in training.index: x = training.iloc[i,:].to_numpy() #x = x[ :, np.newaxis] y = labels.iloc[i,:].values y = y[np.newaxis, :] graph = Data(x = torch.as_tensor(x, dtype=torch.float32), edge_index=coo[0], edge_attr=coo[1],y = torch.as_tensor(y, dtype= torch.float32)) list_of_graphs.append(graph)

From the DataFrame I convert each row corresponding to a single experiment to a (978,) np array. If I add an additional dimension with np.newaxis and make x of shape [978,1] I get the same error but now with the node features stacked in the first dimension: RuntimeError: mat1 and mat2 shapes cannot be multiplied (1956x1 and 978x38) so maybe the way in which I create the DataLoader object might be the problem?

[EdisonLeeeee]

I think you have to add an additional dimension for x and make it of shape [978, 1]. Correspondingly, the in_channels of your model should be 1 instead of 978.

[AlexNiko47]

Hi @EdisonLeeeee
I've just tried this and now I'm getting the following error: RuntimeError: The size of tensor a (1956) must match the size of tensor b (2) at non-singleton dimension 0

The shape of the data.x is torch.Size([1956, 1]) and for data.y is torch.Size([2, 38])

[rusty1s]

If you want to use MSELoss, the shapes of the prediction and targets need to be the same. As such, it looks like you are missing a global pooling function such as torch_geometric.nn.global_mean_pool to convert node-level features into graph-level ones.

[AlexNiko47]

The pooling layer seems to have done the trick! Thanks a lot :)