renqincai / c2ae_tensorflow

Tensorflow implementation of the paper 'Learning Deep Latent Spaces for Multi-Label Classifications' in AAAI 2017

dataset

dataset	labels	instances	feature	cardinality
mirflickr	38	25000	1000	4.7

Performance

type	dataset	Micro F1	Macro F1
this code	mirflickr	0.540699	0.39643
paper	mirflickr	0.54	0.39

Parameter

• alpha(for output loss para): 0.5
• learning rate: 0.0001
• learning rate decay: 0.98
• momentum: 0.99
• optimizer decay: 0.9
• l2penalty：0.001
• maxepoch: 50
• lagrange para:0.5
• batch size: 500
• hidden units: 512
• latent embedding units: (0, 1) of label dims, default 0.8

Description

• learning rate: decay by new_lr = lr_init * decay^epoch

• custom optimizer: similar to RMSProp, decay 0.9, momentum 0.99

  init rrr = 0, delta = 0 rrr=sqrt((rrr.^2)0.9+(grad.^2)0.1) grad=grad/rrr delta=momentumdelta-etagrad new_weight=old_weight + delta

• lagrange para: for caculate embedding loss

Requrements

• tensorflow 1.12.0
• numpy 1.16.2

Thanks

the paper's author code matlab https://github.com/chihkuanyeh/C2AE

other implement(I think there some mistakes in this code) https://github.com/dhruvramani/C2AE-Multilabel-Classification

Reference

Yeh, C. K., Wu, W. C., Ko, W. J., & Wang, Y. C. F. (2017). Learning deep latent space for multi-label classification. In AAAI (pp. 2838–2844).