Hi, I am referring to the following implementation of pinSAGE :https://github.com/dmlc/dgl/tree/master/examples/pytorch/pinsage
I have the same question about predict new node(same feature nature) in Pinsage model. The question detail is same as https://discuss.dgl.ai/t/exporting-embeddings-for-current-pinsage-implementation/1132/7
I do remove the lines 48-49https://github.com/dmlc/dgl/blob/master/examples/pytorch/pinsage/model.py#L48-L49.
But I found https://github.com/dmlc/dgl/blob/master/examples/pytorch/pinsage/layers.py#L173-L174 here, when the train model was first constructed, the bias field in ItemToItemScorer was created as the same size as the input item size in the train graph. So when I predict a larger graph, the train model became incompatible…
I have two question:
1、Can I simply remove the bias field and add_bias method? If I do remove them, will the result become worst dramatically?
2、Is there any other places in current PinSAGE implementation where it will be obstacles for inductive predicting?
Thanks!