Attention-Based Graph Convolutional Network for Zero-Shot Learning with Pre-Training

Table 3

Top-1 accuracy of different models on the aPY test set. To observe the effect of pre-training classes for ZSL, the ImageNet samples of aPY 20 training classes were used as the training set.

Method	Accuracy (%)
	Original loss		Modified loss
	Classifier	All	Classifier	All
GCN	48.0	43.6	64.8	45.4
DGP	60.0	43.5	64.2	47.5
AGCNZ	66.1	49.9	66.8	50.8
ADGPZ	62.3	45.9	65.6	48.3