training loss function - Githubissues

chickenbestlover / RNN-Time-series-Anomaly-Detection

RNN based Time-series Anomaly detector model implemented in Pytorch.

Apache License 2.0

1.17k stars 316 forks source link

training loss function #29

Open jchenWill opened 5 years ago

jchenWill commented 5 years ago

Hi Could you explain the three training loss you defined, free running, teacher forcing, and simplified professor force? or point out related papers? Thanks

chickenbestlover commented 4 years ago

A useful blog post about 'free running' (using outputs as inputs) & 'teacher forcing' : post

Professor forcing paper: NIPS 2016

Simplified professor forcing is my idea which is a simplified version of professor forcing. While processor forcing requires a discriminator, simplified processor forcing does not require it.