This is the pytorch version of Bert-LWAN which is from paper An Empirical Study on Large-Scale Multi-Label Text Classification Including Few and Zero-Shot Labels

The model architecture is referrenced on CAML which is from paper Explainable Prediction of Medical Codes from ClinicalText

the result is:

mertric	pytorch version	tensorflow version (origin code)	the result of the paper
Harmony	RP@5: 0.784 nDCG@5: 0.810	RP@5: 0.781 nDCG@5: 0.805	RP@5: 0.803 nDCG@5: 0.829
Frequent	RP@5: 0.822 nDCG@5: 0.832	RP@5: 0.821 nDCG@5: 0.830	RP@5: 0.843 nDCG@5: 0.854
Few	RP@5: 0.644 nDCG@5: 0.602	RP@5: 0.648 nDCG@5: 0.613	RP@5: 0.699 nDCG@5: 0.650
Zero	RP@5: 0.011 nDCG@5: 0.006	RP@5: 0.045 nDCG@5: 0.020	RP@5: - nDCG@5: -