search

unilight / seq2seq-vc

A sequence-to-sequence voice conversion toolkit.
MIT License
84 stars 10 forks source link

use_guided_attn_loss #11

Closed leelee724 closed 5 months ago

leelee724 commented 5 months ago

Hello author, I want to add guided_attention_loss during training. I have added "use_guided_attn_loss: true" to the config, but the following error message will appear. How to solve it? Thank you!

The error message is as follows: Traceback (most recent call last): File "/home/user/000_VC/seq2seq-vc/egs/0412/stage4A_2_v10_guide_attn_loss/../../../seq2seq_vc/bin/vc_train.py", line 500, in main() File "/home/user/000_VC/seq2seq-vc/egs/0412/stage4A_2_v10_guide_attn_loss/../../../seq2seq_vc/bin/vc_train.py", line 411, in main criterion = { File "/home/user/000_VC/seq2seq-vc/egs/0412/stage4A_2_v10_guide_attn_loss/../../../seq2seq_vc/bin/vc_train.py", line 412, in criterion_class: getattr(seq2seq_vc.losses, criterion_class)( AttributeError: module 'seq2seq_vc.losses' has no attribute 'guided_attn'

unilight commented 5 months ago

Hi @leelee724, unfortunately the guided attention loss is not supported in this repo... also based on my informal experiments, it does not have a significant impact on the convergence speed nor the quality of the generated speech.