mberr / torch-ppr Goto Github PK

related to #18

Dear torch-ppr

I installed torch-ppr on my Mac with python 3.9 and run the example code

>>> import torch
>>> edge_index = torch.as_tensor(data=[(0, 1), (1, 2), (1, 3), (2, 4)]).t()
>>> from torch_ppr import page_rank
>>> page_rank(edge_index)

I got a runtimeerror as

x = torch.sparse.addmm(input=x0, sparse=adj, dense=x, beta=alpha, alpha=beta)
RuntimeError: mat1 and mat2 shapes cannot be multiplied (2x4 and 2x1)

I printed the shape of x0, adj and x

torch.Size([2, 1])
torch.Size([2, 4])
torch.Size([2, 1])

I believe that the shape of adj should be 2x2 or I might be wrong.
I find the define process of adj.

# convert to sparse matrix, shape: (n, n)
adj = edge_index_to_sparse_matrix(edge_index=edge_index, num_nodes=num_nodes)
adj = adj + adj.t()

The adj is symmect.

I wonder how to fix the runtimeError or any suggestions?
Thanks in advanced
meatball1982
12-May-2022 09:54:50

can u provide one simple example , how to use it for bipartite graph for recommending items to for users using this PPR code.

thanks

Hello,

Thank you for this great repository; it is a great, handy package that performs very well! I was wondering however; is it possible to incorporate edge weights into the personalized pagerank method?

Best
Filip

Thank you for this repo!

The reason to request a 'layer' fomulation is to convert the function page_rank to an onnx graph with torch.onnx (only accepts models).

Once I have the onnx model, I can compile it different hardware (other than cuda).

Maybe need just the forward pass, no need for a backward pass although I think the compute will be differentiable.

Thanks.

Suddenly, everything stopped working 😱 presumably because of the changes to torch.sparse.
Particularly, I am on PyTorch 1.10, master branch of PyKEEN and torch-ppr 0.0.5.

Problem 1: the allclose() check does not pass now:

MWE:

import torch
from torch_ppr import page_rank

from pykeen.datasets import FB15k237

dataset = FB15k237(create_inverse_triples=False)
edges = dataset.training.mapped_triples[:, [0, 2]].t()
pr = page_rank(edge_index=torch.cat([edges, edges.flip(0)], dim=-1), num_nodes=dataset.num_entities)

>> ValueError: Invalid column sum: tensor([1.0000, 1.0000, 1.0000,  ..., 1.0000, 1.0000, 1.0000]). expected 1.0

Looking into the debugger:

• adj_sum does sum up to the number of nodes
• the default tolerance fails the check, but if I reduce rtol=1e-4 or atol=1e-4 - the check passes

Problem 2: the signature of torch.sparse.addmm has changed from the one used in power_iteration so the API call fails with the unknown kwarg error.

In fact, I can't find where those kwargs input, sparse, dense come from because the current signature has less readable mat, mat1, mat2. I traced to the very Torch 1.3.0 and still can't find where those originated from. Where does this signature come from? 😅

My test env

torch                 1.10.0
torch-ppr             0.0.5