APPNP

[DuanhaoranCC]

from ogb.nodeproppred import PygNodePropPredDataset
import torch_geometric.transforms as T
from torch_geometric.nn import APPNP

dataset = PygNodePropPredDataset('ogbn-arxiv', root='./', transform=T.ToUndirected())
data = dataset[0]
print(data)
appnp = APPNP(K=20, alpha=0.2)
id = th.eye(data.num_nodes).float()
diff_adj = appnp(x=id, edge_index=data.edge_index)

Hello, I am using the APPNP propagation algorithm on a large image, but the following error is reported, I want to get the diffusion matrix of the large image.

Data(num_nodes=169343, edge_index=[2, 2315598], x=[169343, 128], node_year=[169343, 1], y=[169343, 1])
Traceback (most recent call last):
  File "1.py", line 196, in <module>
    process_dataset_appnp(0.01)
  File "1.py", line 186, in process_dataset_appnp
    diff_adj = appnp(x=id, edge_index=data.edge_index)
  File "/home/kglab/.conda/envs/kg/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/kglab/.conda/envs/kg/lib/python3.8/site-packages/torch_geometric/nn/conv/appnp.py", line 118, in forward
    x = self.propagate(edge_index, x=x, edge_weight=edge_weight,
  File "/home/kglab/.conda/envs/kg/lib/python3.8/site-packages/torch_geometric/nn/conv/message_passing.py", line 309, in propagate
    coll_dict = self.__collect__(self.__user_args__, edge_index,
  File "/home/kglab/.conda/envs/kg/lib/python3.8/site-packages/torch_geometric/nn/conv/message_passing.py", line 202, in __collect__
    data = self.__lift__(data, edge_index, dim)
  File "/home/kglab/.conda/envs/kg/lib/python3.8/site-packages/torch_geometric/nn/conv/message_passing.py", line 172, in __lift__
    return src.index_select(self.node_dim, index)
RuntimeError: [enforce fail at alloc_cpu.cpp:73] . DefaultCPUAllocator: can't allocate memory: you tried to allocate 1683229455052 bytes. Error code 12 (Cannot allocate memory)

[rusty1s]

You would first need to apply an MLP before putting it into APPNP. Otherwise, you will create a num_nodes x num_nodes 20 times (K=20), which probably will not fit into memory.