A Distinguishable Pseudo-Feature Synthesis Method for Generalized Zero-Shot Learning

Algorithm 1

DPFS training algorithm

	Input: training dataset , class-attribute matrices of seen classes and unseen ones and , learning rate , and max-epochs of the embedding module pretraining and the classifier training and
	Initialize: set of the weight parameters of the embedding module and the preclassifier , classifier weight parameter
(1)	Build attribute projection matrices with and by, and equations (1), (2), and (9) for unseen classes
(2)	Compute the base vectors with the matrices and by equations (10) and (11)
(3)	for step = 0, …, do
(4)	Set and
(5)	Compute base class prototype with by equation (3)
(6)	Build prototype query set with and by equations (5) and (6)
(7)	Compute by equation (8)
(8)	Update
(9)	end for
(10)	for step = 0, …, do
(11)	Synthesize candidate pseudo-features for unseen classes by equation (12)
(12)	Dispose of the outliers of candidate pseudo-features by equation (13)
(13)	Select a certain number of samples
(14)	Train the classifier with the selected samples to update while fine-tune .
(15)	end for
	Output: Embedding network and classifier