Seed does not work for ShaDowKHopSampler

[CongWeilin]

I am trying to generate the same subgraphs for validation using the following code

all_loader = ShaDowKHopSampler(data,
                               depth=2,
                               num_neighbors=args.hop_neighbors,
                               batch_size=2048,
                               num_workers=10,
                               shuffle=False)

seed_everything(args.seed)
# also tried this >>> torch.manual_seed(12345)
for subgraph_data in all_loader:
    break
subgraph_data

I run the code twice, but gives me different results:

# first time
DataBatch(batch=[642233], ptr=[2049], root_n_id=[2048], adj_t=[642233, 642233, nnz=15303073], x=[642233, 101], y=[2048, 1], y_one_hot_train=[642233, 48], node_inds=[642233])

# second time
DataBatch(batch=[641999], ptr=[2049], root_n_id=[2048], adj_t=[641999, 641999, nnz=15296681], x=[641999, 101], y=[2048, 1], y_one_hot_train=[641999, 48], node_inds=[641999])

[rusty1s]

Yes, sorry about that. Currently, our custom C++ sampling code does not respect any seeds. One workaround is to calculate the subgraphs once, hold them in a data_list and re-use them across epochs.