search

yinboc / DGP

Rethinking Knowledge Graph Propagation for Zero-Shot Learning, in CVPR 2019
MIT License
320 stars 57 forks source link

Questions about evaluation on awa2 #17

Open CallMeNumber1 opened 3 years ago

CallMeNumber1 commented 3 years ago

Q1：If i want to evaluate on awa2，should i still train on imagenet？

CallMeNumber1 commented 3 years ago

Q2：By the way, 24 out of 40 of the training classes in awa2 overlap with the ImageNet 2012 dataset. Does this mean that the rest of the training classes do not have corresponding ground truth classifier wights ?

lil-wayne-0319 commented 2 years ago

yes, you can. cause awa2 is the subset of imageNet+22K

lil-wayne-0319 commented 2 years ago

i think the ground truth weight of the unseen class is constructed by Resnet50 model and inputs from imageNet+22k, not your training set imageNet+1k