Argparse Argument for "Patience" -- and Early Exit without Updates for "patience" number of steps

ReaLLMASIC / nanoGPT

The simplest, fastest repository for training/finetuning medium-sized GPTs.

MIT License

23 stars 17 forks source link

Closed mmoffatt2 closed 7 months ago

mmoffatt2 commented 8 months ago

-added patience argparse parameter to train.py -if self.best_val_loss doesn't improve after "patience" number of steps, then train.py will exit early

Note: self.best_val_loss is checked after each step, not iteration

gkielian commented 7 months ago

Looks good :+1: merging