Currently, during training over multiple autoregressive steps, the overall training loss is aggregated into a single scalar and logged as train_loss. There is no way to monitor the degradation of loss at individual unroll steps during training (unlike validation and testing, which support per-step loss logging using val_loss_unroll{step}).
This enhancement proposes adding a --train_steps_to_log CLI option (similar to --val_steps_to_log) to allow users to log the training loss for specific unroll steps separately (e.g. train_loss_unroll1, train_loss_unroll5, train_loss_unroll10).
Currently, during training over multiple autoregressive steps, the overall training loss is aggregated into a single scalar and logged as
train_loss. There is no way to monitor the degradation of loss at individual unroll steps during training (unlike validation and testing, which support per-step loss logging usingval_loss_unroll{step}).This enhancement proposes adding a
--train_steps_to_logCLI option (similar to--val_steps_to_log) to allow users to log the training loss for specific unroll steps separately (e.g.train_loss_unroll1,train_loss_unroll5,train_loss_unroll10).