+def get_lr(n_epoch, it):
+ if args.legacy_lr_schedule:
+ # my crude scheduling to compare to previous baseline, added
+ # warmup though
+
+ if it < args.nb_warmup_iter:
+ return args.legacy_learning_rate * it / args.nb_warmup_iter
+ elif it < args.nb_large_lr_epochs:
+ return args.legacy_learning_rate
+ else:
+ return args.legacy_min_learning_rate
+
+ # from nanoGPT