######################################################################
+
+def str2bool(x):
+ x = x.lower()
+ if x in {"1", "true", "yes"}:
+ return True
+ elif x in {"0", "false", "no"}:
+ return False
+ else:
+ raise ValueError
+
+
parser = argparse.ArgumentParser(
description="An implementation of GPT with cache.",
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
parser.add_argument("--min_learning_rate", type=float, default=6e-5)
+# legacy
+
+parser.add_argument("--legacy_lr_schedule", type=str2bool, default=True)
+
+parser.add_argument("--legacy_large_lr", type=float, default=1e-4)
+
+parser.add_argument("--legacy_small_lr", type=float, default=2e-5)
+
+parser.add_argument("--legacy_nb_epoch_large_lr", type=float, default=10)
+
########################################
parser.add_argument("--model", type=str, default=None)
######################################################################
-# from nanoGPT
+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_large_lr * it / args.nb_warmup_iter
+ elif n_epoch < args.legacy_nb_epoch_large_lr:
+ return args.legacy_large_lr
+ else:
+ return args.legacy_small_lr
+
+ # from nanoGPT
-def get_lr(it):
# 1) linear warmup for warmup_iter steps
if it < args.nb_warmup_iter:
return args.learning_rate * it / args.nb_warmup_iter
total_loss = loss + (args.rho * inner_loss if args.rho > 0 else 0.0)
it += 1
- lr = get_lr(it)
+ lr = get_lr(n_epoch, it)
for param_group in optimizer.param_groups:
param_group["lr"] = lr