How to use GraphSAINTRandomWalkSampler for BipartiteData

[freemanhm]

Hi all,
I followed the torch_geometric tutorial and I am using Bipartite Graph for link prediction.

class BipartiteData(Data):
    def __init__(self, edge_index, x_s, x_t):
        super(BipartiteData, self).__init__()
        self.edge_index = edge_index
        self.x_s = x_s
        self.x_t = x_t

    def __inc__(self, key, value):
        if key == 'edge_index':
            return torch.tensor([[self.x_s.size(0)], [self.x_t.size(0)]])
        else:
            return super().__inc__(key, value)

we add some attributes to the data(BipartiteData). The size of x_s =[130000,10],x_t[100000,12], It is too big to use a single GPU for training, so I want to split this graph with GraphSAINTRandomWalkSampler().

data= BipartiteData(edge_index=edge_index,x_s = x_s,x_t=x_t)
data.train_pos_edge_index = train_pos_edge_index
data.train_neg_edge_index = train_neg_edge_index
data.test_pos_edge_index = test_pos_edge_index
data.user_emb_nums = x_t_emb_num.int()
data.item_emb_nums = x_s_emb_num.int()
data.item_title_x = item_title_x
data.item_tags_x = item_tags_x
data.user_jobroles_x = user_jobroles_x
data.words_num = words_num

loader = GraphSAINTRandomWalkSampler(data, batch_size=6000, walk_length=2, num_steps=5, sample_coverage=100,)
for data in loader:
    print(data.x_s.size())

but I get the following problems:

  Traceback (most recent call last):
    File "/root/miniconda3/envs/myconda/lib/python3.8/site-packages/IPython/core/interactiveshell.py", line 3418, in run_code
      exec(code_obj, self.user_global_ns, self.user_ns)
    File "<ipython-input-1-d07b1f27fae1>", line 1, in <module>
      for data in loader:
    File "/root/miniconda3/envs/myconda/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 435, in __next__
      data = self._next_data()
    File "/root/miniconda3/envs/myconda/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 475, in _next_data
      data = self._dataset_fetcher.fetch(index)  # may raise StopIteration
    File "/root/miniconda3/envs/myconda/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 47, in fetch
      return self.collate_fn(data)
    File "/root/miniconda3/envs/myconda/lib/python3.8/site-packages/torch_geometric/data/graph_saint.py", line 102, in __collate__
      data = self.data.__class__()
  TypeError: __init__() missing 3 required positional arguments: 'edge_index', 'x_s', and 'x_t'

[rusty1s]

I see. The error comes from requiring the 3 input arguments of BipartiteData. Dropping the __init__ function completely should fix this issue. However, I don't think that GraphSAINT is well-defined on bipartite graph-structured data but instead expects to operate in a homogeneous graph. For this, it's best to convert your graph into a homogeneous one first before running GraphSAINT samplers (e.g., via edge_index[:, 1] += 100000)

[freemanhm]

OK, I'll try it according to your suggestion. Thanks a lot for your answer and this awesome repository 👍

[freemanhm]

However, there are two types of nodes with different features in my Bipartite Graph. How can I convert the graph into a homogeneous one?

[rusty1s]

Yes, this is a bit tricky in that case and we are working on better support for that. One possible workaround is to keep both x_s and x_t outside data, and instead add index attributes to data that denote to which type each node belongs, and what's it's global index in either x_s or x_t, e.g.:

data.n_type = torch.cat([torch.zeros(130000, dtype=torch.long), torch.ones(100000, dtype=torch.long)])
data.n_id = torch.cat([torch.arange(130000), torch.arange(100000)])

which can then be used after sampling to reconstruct mini-batched x_s and x_t.

[freemanhm]

ok，I will try.

[freemanhm]

It works when I revised my codes per your suggestions. Thanks a lot.
Now, I have another question about how I should set the parameters' values in the "GraphSAINTRandomWalkSampler(data, batch_size=6000, walk_length=2, num_steps=5, sample_coverage=100,)". In other words, I am not sure about the relationship between the "batch_size" and "num_steps". In my opinion, I would set the "num_steps=data.num_nodes/batch_size", is that right?