mwer

mwer copied to clipboard

请问mwe lossr应该怎么做梯度回传呢？

2

看了论文也不太明白，是和CE loss加权吗？但是weighted_relative_edit_errors最后应该会出现负数的情况吧？负数怎么作为loss回传呢？

Alpaca0904

Question

2

It seems that your mwer loss implementation needs prior beam search for inputs for mwer_loss function. We can get 'seq_logprobs' during beam search for each hypothesis, however, your implementation seems...

jtkim-kaist

Have you reproduced results mentioned in the paper

1

thanks for sharing implementation of mwer, Have you reproduced results mentioned in the paper of Rohit Prabhavalkar etc. MINIMUM WORD ERROR RATE TRAINING FOR ATTENTION-BASED SEQUENCE-TO-SEQUENCE MODELS

shaojunyao