Closed hawklucky closed 6 years ago
You don't have to do that. Since for the evaluation on AwA2, we just fetch the corresponding predictions on ImageNet nodes, since AwA2 nodes is a subset of ImageNet nodes.
For the induce_parents function, it BFS from the query nodes "s" until reach the base nodes "stop_set", intuitively attach "s" to "stop_set", which is mainly due to that xml-nodes does not contain the all ImageNet-nodes for testing.
I want to know how to construct awa2 graph for training? Would you give me some help?
You don't have to do that. It utilizes the exact same graph as imagenet-21k task, also the same model which is trained on imagenet-21k.
Thanks, I'll try it.
Can I ask is the results for GCNZ reported in the paper also trained on the exact same graph as imagenet-21k task, also the same model which is trained on imagenet-21k?
Can I ask is the results for GCNZ reported in the paper also trained on the exact same graph as imagenet-21k task, also the same model which is trained on imagenet-21k?
All the results of GCNZ are from the same model.
Did I make awa2_induced_graph by replace input and output arguments? And what the meaning of induce_parents function? Also I notice that you are construct graph with all (train + test) categories together and Would that obtain test information during training?