Custom negative sampler

Questions
#1 
class CustomNegativeSampler:
    def __init__(self, k=1):
        self.k = k
        neg_g, _ = dgl.load_graphs(neg_graph.dgl')
        self.neg_g = neg_g[0]

    def __call__(self, g, eids, exclude_eids=None):
        src, dst = self.neg_g.edges()
        num_samples = len(eids) * self.k
        random_indices = torch.randint(0, len(src), (num_samples,))
        neg_src = src[random_indices]
        neg_dst = dst[random_indices]

this is my custom negative sampler im trying to use and i get this error:

TypeError: This sampler does not support edge or link prediction; please add anoptional third argument for edge IDs to exclude in its sample() method.

can anyone explain what i must modify in my code ?

#2

Hi @angelosmath, how do you invoke this sampler?

#3

I’m using it here:

    test_sampler = as_edge_prediction_sampler(
        sampler,
        exclude = "reverse_id" if args.exclude_edges else None,
        reverse_eids = reverse_eids if args.exclude_edges else None,
        negative_sampler =  CustomNegativeSampler(k=1),
    )

my reference for the negative sampler is this example:

https://docs.dgl.ai/en/0.8.x/guide/minibatch-link.html#guide-minibatch-link-classification-sampler

thanks for the reply if you need more information feel free to ask.

#4

Could you try the latest tutorial? 6.3 Training GNN for Link Prediction with Neighborhood Sampling — DGL 2.0 documentation

#5

I’m trying but i dont know what is wrong with graphbolt, i get the error:

AttributeError: module 'dgl' has no attribute 'graphbolt'

Αlso, do I need to use the DataLoader you present in 6.3 or can I use the 6.3 negative sampler with the dataloader I have?

test_sampler = NeighborSampler(
        [15, 10, 5],
        prefetch_node_feats = ["feat"],
        fused = fused_sampling,
    ) 

    test_sampler = as_edge_prediction_sampler(
        sampler,
        exclude = "reverse_id" if args.exclude_edges else None,
        reverse_eids = reverse_eids if args.exclude_edges else None,
        negative_sampler =  CustomNegativeSampler(k=1)
    )

    test_dataloader = DataLoader(
        test_pos_graph,
        seed_edges,
        sampler,
        device=device,
        batch_size=args.train_batch_size,
        shuffle=True,
        drop_last=False,
        num_workers=0,
        use_uva=use_uva,
    )
#6

It is just release on last Friday in DGL 2.0 for Linux, please update your DGL version and retry it, we are happy to hear your feedback.

#7

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.