DGL model on Hetrogeneous Graph

[Noorulain]

I am trying to apply the GCN2Conv layers with HetroConv module of dgl, but every time I find the "TypeError: forward() missing 1 required positional argument: ‘feat_0’ ". Can you help to fix? The code which I have used is shown below:

import dgl
import torch
from dgl.nn import HeteroGraphConv, GCN2Conv


edges1 = (torch.tensor([0,0,1,1,2]), torch.tensor([2,3,4,4,5]))
edges2 = (torch.tensor([0,1,2,3,4]), torch.tensor([2, 3, 4, 5, 5]))


# random edge masks
mask1 = (torch.rand(5,1)>0.5).float()
mask2 = (torch.rand(5,1)>0.5).float()

#print(type(mask1)) #tenosr
#print(mask1.shape)
#print(mask2)
graph_data = {
       ('user', 'follows', 'user'): edges1,
       ('user', 'plays', 'game'): edges2
       }

# convert the graph data to the hetrogeneous graph
g = dgl.heterograph(graph_data)

#print("g", g)
h1 = {
  'user': torch.randn((g.num_nodes("user"), 5)),
  'game': torch.randn((g.num_nodes("game"), 5))
  }
  

lin1 = torch.nn.Linear(5, 6)
conv1 = HeteroGraphConv({
       'follows': GCN2Conv(6, layer=1, alpha=0.5, project_initial_features=True),
       'plays': GCN2Conv(6, layer=1, alpha=0.5, project_initial_features=True),
       }, aggregate="sum")
conv2 = HeteroGraphConv({
       'follows': GCN2Conv(6, layer=1, alpha=0.5, project_initial_features=True),
       'plays': GCN2Conv(6, layer=1, alpha=0.5, project_initial_features=True),
       }, aggregate="sum")


# pas masks as edge_weight here



inp1 = {key: lin1(x) for key, x in h1.items()}

res = conv1(g, inp1, inp1, mod_kwargs={
                   'follows':{'edge_weight':mask1},
                   'plays':{'edge_weight': mask2}
                   }
                )
res = conv2(g, res, inp1, mod_kwargs={
                   'follows':{'edge_weight':mask1},
                   'plays':{'edge_weight': mask2}
                   }
                )

My error is shown below:

Traceback (most recent call last):
  File "/DGL/GCNII/example1.py", line 49, in <module>
    res = conv1(g, inp1, inp1, mod_kwargs={
  File "/dgl/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl
    return forward_call(*input, **kwargs)
  File "/.conda/envs/dgl/lib/python3.8/site-packages/dgl/nn/pytorch/hetero.py", line 185, in forward
    dstdata = self.mods[etype](
  File "/.conda/envs/dgl/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl
    return forward_call(*input, **kwargs)
TypeError: forward() missing 1 required positional argument: 'feat_0'

[minjie]

The HeteroGraphConv has the following restriction on the graph convolution modules passed in:

• The first argument must be an DGLGraph or an MFG.
• The second argument must be a node feature tensor.

However, GCN2Conv needs two input node feature tensors, feat and feat_0.

[Noorulain]

@minjie , it means, with DGL graph we cannot apply GCN2Conv module. I can successfully apply GCN2Conv from DGL on homogenous graph. But, on heterogeneous graph, I find an error of feat_0.

[minjie]

That’s right. Probably need to modify the GCN2Conv to simplify it’s arguments.

[Noorulain]

@minjie, can you tell me if I modify the code within hetero function of dgl. Through which way, I will utilize, should I need to run every thing from scratch. Does dgl allow to modify the code and then use the updated code?

[minjie]

Yes, just copy out the corresponding source code and make your desired changes.

[Noorulain]

@minjie , I want to modify this code which is present in this link https://github.com/dmlc/dgl/blob/master/python/dgl/nn/pytorch/network_emb.py, thorugh which way we can import the things which are written under from …,

from …base import NID
from …convert import to_heterogeneous, to_homogeneous
from …random import choice
from …sampling import random_walk. I want to import these functions from dgl. But, thorugh which way, I will import NID from …base.