search

auspicious3000 / contentvec

speech self-supervised representations
MIT License
434 stars 32 forks source link

Contrastive loss weights #8

Open jingru-lin opened 1 year ago

jingru-lin commented 1 year ago

According to the paper, the contrastive loss weight is set to 1e-5 * num_train_steps.

However, in the given config, the 'loss_weights: [10,1]', where 1 is the contrastive loss weight.

May I know why?

auspicious3000 commented 1 year ago

The run.sh will override the config.