Learning rate start to decay before start_decay_steps

[witchinghour (Roman)]

Why lr starts to decay after warmup_steps=8000, when start_decay_step is set to 15000?

[guillaumekln (Guillaume Klein)]

Can you post the complete configuration file and the training command line?

[witchinghour (Roman)]

Unfortunately this is the only one available at the moment

[guillaumekln (Guillaume Klein)]

We indeed see “Start decay steps” in this screenshot, but you should check that the parameter is actually set in the training configuration and in the correct place.