Custom heterogeneous data can't fit into NeighborLoader.

[darrenleeleelee1]

This is my custom heterogeneous data.

HeteroData(
  y=[864900],
  A={
    x=[103928, 4],
    train_mask=[103928]
  },
  B={ x=[101684, 5] },
  C={ x=[49, 5] },
  (A, related, A)={
    edge_index=[2, 114299],
    edge_attr=[114299, 1]
  },
  (A, related, B)={
    edge_index=[2, 16577],
    edge_attr=[16577, 1]
  },
  (A, related, C)={
    edge_index=[2, 140],
    edge_attr=[140, 1]
  },
  (B, related, B)={
    edge_index=[2, 97424],
    edge_attr=[97424, 1]
  }
)

Then I want to fit into my Custom GAT model.

from torch_geometric.loader import HGTLoader, NeighborLoader
from tqdm import tqdm

from torch_geometric.nn import GATConv, Linear, to_hetero

class GAT(torch.nn.Module):
    def __init__(self, hidden_channels):
        super().__init__()
        self.conv1 = GATConv((-1, -1), hidden_channels, add_self_loops=False)
        self.lin1 = Linear(-1, hidden_channels)
        self.conv2 = GATConv((-1, -1), 128, add_self_loops=False)
        self.lin2 = Linear(-1, 128)
        self.fc = torch.nn.Linear(128, 864900)

    def forward(self, x, edge_index):
        x = self.conv1(x, edge_index) + self.lin1(x)
        x = x.relu()
        x = self.conv2(x, edge_index) + self.lin2(x)
        x = self.fc(x)
        return x

model = GAT(hidden_channels=64)
model = to_hetero(model, data.metadata(), aggr='sum')

device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
model = model.to(device)
data = dataset[0]
data = data.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
criterion = torch.nn.MSELoss()

num_neighbors=[8] * 4
input_nodes=['R', data['R'].train_mask]

loader = NeighborLoader(data, batch_size=8, input_nodes=input_nodes,num_neighbors=num_neighbors)
print(loader)


for epoch in range(100):
    model.train()
    total_examples = total_loss = 0
    for batch in loader:
        optimizer.zero_grad()
        batch = batch.to(device)
        out = model(batch.x_dict, batch.edge_index_dict)
        loss = criterion(out, batch.y)
        loss.backward()
        optimizer.step()
        
        total_examples += batch_size
        total_loss += float(loss) * batch_size

    print(f"Epoch: {epoch}, Loss: {loss.item()}")

But I got the error.

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-21-a707df1ecad0> in <module>
     60         optimizer.zero_grad()
     61         batch = batch.to(device)
---> 62         out = model(batch.x_dict, batch.edge_index_dict)
     63         loss = criterion(out, batch.y)
     64         loss.backward()

~/miniconda3/envs/pytorch_env/lib/python3.6/site-packages/torch/fx/graph_module.py in wrapped_call(self, *args, **kwargs)
    614                     print(generate_error_message(topmost_framesummary),
    615                           file=sys.stderr)
--> 616                 raise e.with_traceback(None)
    617 
    618         cls.__call__ = wrapped_call

RuntimeError: The size of tensor a (8) must match the size of tensor b (0) at non-singleton dimension 0

1. I am wondering is that the Node feature size can't be different? (example A's node features size is 4, other is 5)
2. Am I using the NeighborLoader right? The original graph is too large to put into my GPU memory.
3. And where did I did wrong to cause this error?

[EdisonLeeeee]

There seems an error in your HeteroData, in which y should be existed in a specific node type, e.g.:

data['A'].y = ...

In your example, y is interpreted as a node type with zero nodes. BTW, the node feature size can be different for a hetero graph.

[darrenleeleelee1]

Ok I have transform my data into below, also use T.ToUndirected()

HeteroData(
  A={
    x=[103928, 4],
    train_mask=[103928],
    y=[103928]
  },
  B={ x=[101684, 5] },
  C={ x=[49, 5] },
  (A, connect, A)={
    edge_index=[2, 228598],
    edge_attr=[228598, 1]
  },
  (A, connect, B)={
    edge_index=[2, 16577],
    edge_attr=[16577, 1]
  },
  (A, connect, C)={
    edge_index=[2, 140],
    edge_attr=[140, 1]
  },
  (B, connect, B)={
    edge_index=[2, 194848],
    edge_attr=[194848, 1]
  },
  (B, rev_connect, A)={
    edge_index=[2, 16577],
    edge_attr=[16577, 1]
  },
  (C, rev_connect, A)={
    edge_index=[2, 140],
    edge_attr=[140, 1]
  }
)

But the same error is persist.

from torch_geometric.loader import HGTLoader, NeighborLoader
from tqdm import tqdm

from torch_geometric.nn import GATConv, Linear, to_hetero

class GAT(torch.nn.Module):
    def __init__(self, hidden_channels):
        super().__init__()
        self.conv1 = GATConv((-1, -1), hidden_channels, add_self_loops=False)
        self.lin1 = Linear(-1, hidden_channels)
        self.conv2 = GATConv((-1, -1), 128, add_self_loops=False)
        self.lin2 = Linear(-1, 128)
        self.fc = torch.nn.Linear(128, 864900)

    def forward(self, x, edge_index):
        x = self.conv1(x, edge_index) + self.lin1(x)
        x = x.relu()
        x = self.conv2(x, edge_index) + self.lin2(x)
        x = self.fc(x)
        return x

model = GAT(hidden_channels=64)
model = to_hetero(model, data.metadata(), aggr='sum')

device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
model = model.to(device)
data = dataset[0]
data = data.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
criterion = torch.nn.MSELoss()

num_neighbors=[8] * 4
input_nodes=['R', data['R'].train_mask]
loader = NeighborLoader(data, batch_size=8, input_nodes=['R', None], num_neighbors=[-1])

for epoch in range(100):
    model.train()
    total_examples = total_loss = 0
    for batch in loader:
        optimizer.zero_grad()

        print(batch.x_dict)

        out = model(batch.x_dict, batch.edge_index_dict)
        loss = criterion(out, batch.y)
        loss.backward()
        optimizer.step()
        
        total_examples += batch_size
        total_loss += float(loss) * batch_size

    print(f"Epoch: {epoch}, Loss: {loss.item()}")

{'A': tensor([[1.0000e+00, 1.0000e+00, 0.0000e+00, 0.0000e+00],
        [1.0000e+00, 1.0000e+00, 4.0000e+03, 0.0000e+00],
        [1.0000e+00, 1.0000e+00, 4.8000e+03, 0.0000e+00],
        [1.0000e+00, 1.0000e+00, 9.6000e+03, 0.0000e+00],
        [1.0000e+00, 1.0000e+00, 1.4400e+04, 0.0000e+00],
        [1.0000e+00, 1.0000e+00, 1.9200e+04, 0.0000e+00],
        [1.0000e+00, 1.0000e+00, 2.4000e+04, 0.0000e+00],
        [1.0000e+00, 1.0000e+00, 2.8800e+04, 0.0000e+00],
        [1.0000e+00, 4.0000e+00, 4.0000e+03, 0.0000e+00],
        [1.0000e+00, 1.0000e+00, 3.2000e+04, 0.0000e+00]], device='cuda:1'), 'B': tensor([], device='cuda:1', size=(0, 5)), 'C': tensor([], device='cuda:1', size=(0, 5))}
Traceback (most recent call last):
  File "/DATA/darren/miniconda3/envs/pytorch_env/lib/python3.6/site-packages/torch/fx/graph_module.py", line 608, in wrapped_call
    return super(type(self), self).__call__(*args, **kwargs)
  File "/DATA/darren/miniconda3/envs/pytorch_env/lib/python3.6/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "<eval_with_key>.5", line 22, in forward
    conv1__I = torch.add(conv1__I1, conv1__I2);  conv1__I1 = conv1__I2 = None
RuntimeError: The size of tensor a (10) must match the size of tensor b (0) at non-singleton dimension 0

Call using an FX-traced Module, line 22 of the traced Module's generated forward function:
    conv1__R = torch.add(conv1__R3, conv1__R4);  conv1__R3 = conv1__R4 = None
    conv1__I = torch.add(conv1__I1, conv1__I2);  conv1__I1 = conv1__I2 = None

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
    lin1__R = self.lin1.R(x__R);  x__R = None

    lin1__I = self.lin1.I(x__I);  x__I = None

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-7-37287993aeba> in <module>
     55         print(batch.x_dict.keys)
     56         print(batch.x_dict)
---> 57         out = model(batch.x_dict, batch.edge_index_dict)
     58         loss = criterion(out, batch.y)
     59         loss.backward()

~/miniconda3/envs/pytorch_env/lib/python3.6/site-packages/torch/fx/graph_module.py in wrapped_call(self, *args, **kwargs)
    614                     print(generate_error_message(topmost_framesummary),
    615                           file=sys.stderr)
--> 616                 raise e.with_traceback(None)
    617 
    618         cls.__call__ = wrapped_call

RuntimeError: The size of tensor a (10) must match the size of tensor b (0) at non-singleton dimension 0

[EdisonLeeeee]

You were inputting type R nodes in your NeighborLoader, but I couldn't find such a type of node in your HeteroData.

[darrenleeleelee1]

Sorry, I just replace the original node type to upload here. The original HeteroData would be like this.

HeteroData(
  R={
    x=[103928, 4],
    train_mask=[103928],
    y=[103928]
  },
  I={ x=[101684, 5] },
  V={ x=[49, 5] },
  (R, connect, R)={
    edge_index=[2, 228598],
    edge_attr=[228598, 1]
  },
  (R, connect, I)={
    edge_index=[2, 16577],
    edge_attr=[16577, 1]
  },
  (R, connect, V)={
    edge_index=[2, 140],
    edge_attr=[140, 1]
  },
  (I, connect, I)={
    edge_index=[2, 194848],
    edge_attr=[194848, 1]
  },
  (I, rev_connect, R)={
    edge_index=[2, 16577],
    edge_attr=[16577, 1]
  },
  (V, rev_connect, R)={
    edge_index=[2, 140],
    edge_attr=[140, 1]
  }
)

[darrenleeleelee1]

And I am using Python3.6, torch versions are.

$ pip3 list | grep 'torch'
torch                 1.10.1+cu111
torch-cluster         1.5.9
torch-geometric       2.0.3
torch-scatter         2.0.9
torch-sparse          0.6.12
torch-spline-conv     1.2.1
torchaudio            0.10.1+cu111
torchvision           0.11.2+cu111

The server is old so I can upgrade the python and the torch package.

[akihironitta]

I'd strongly suggest updating your Python version if possible because Python 3.7 also ended its life last month https://devguide.python.org/versions/.

[darrenleeleelee1]

Okay, I will try to upgrade the Python version. If the error persists, I will inform you. Thanks.