Random walk with custom transition probabilities

[jinga-lala]

Is there any way to call random_walk function with custom transition probabilities?

[rusty1s]

Currently no. I think one can implement it via alias sampling.

[jinga-lala]

Are you referring to this function?

pytorch_cluster/torch_cluster/rw.py

Line 8 in 4764b48

def random_walk(row: Tensor, col: Tensor, start: Tensor, walk_length: int,

If yes, I am sorry but I don't understand this function because it seems you are returning the same function again. If possible can you write the exact code snippet?
Any help would be appreciated.

[rusty1s]

torch.ops.torch_cluster.random_walk calls the underlying CPU/CUDA implementation in csrc/cpu/rw_cpu.cpp and csrc/cuda/rw_cuda.cu. These implementations cannot currently handle custom transition probabilities.

[jinga-lala]

Thanks for your reply.
Can you please write a wrapper over this function to handle transition probabilities, probably using alias sampling as you mentioned earlier?

[rusty1s]

I see what I can do :)

[ChainlessCoder]

This would be an extremely useful feature... As of now it's kinda tricky to do weighted random walks for e.g. node embeddings :/