######################################################################
+
+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,
# legacy
-parser.add_argument("--legacy_lr_schedule", action="store_true", default=False)
+parser.add_argument("--legacy_lr_schedule", type=str2bool, default=True)
-parser.add_argument("--legacy_learning_rate", type=float, default=1e-4)
+parser.add_argument("--legacy_large_lr", type=float, default=1e-4)
-parser.add_argument("--legacy_min_learning_rate", type=float, default=2e-5)
+parser.add_argument("--legacy_small_lr", type=float, default=2e-5)
-parser.add_argument("--nb_large_lr_epochs", type=float, default=10)
+parser.add_argument("--legacy_nb_epoch_large_lr", type=float, default=10)
########################################
parser.add_argument("--rho", type=float, default=0.0)
-parser.add_argument("--dim_rec_v", type=int, default=None)
-
parser.add_argument("--nb_blocks", type=int, default=None)
parser.add_argument("--dropout", type=float, default=0.1)
parser.add_argument("--no_checkpoint", action="store_true", default=False)
-parser.add_argument("--overwrite_results", action="store_true", default=False)
+parser.add_argument("--continue_training", action="store_true", default=False)
parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
"dim_keys": 32,
"dim_hidden": 32,
"nb_heads": 2,
- "dim_rec_v": 16,
"nb_blocks": 2,
},
"17K-C": {
"nb_heads": 2,
"nb_lines": 16,
"caterpillar_height": 4,
- "dim_rec_v": 16,
"nb_blocks": 2,
},
"4M": {
"dim_keys": 32,
"dim_hidden": 1024,
"nb_heads": 4,
- "dim_rec_v": 64,
"nb_blocks": 6,
},
"4M-C": {
"nb_heads": 4,
"nb_lines": 32,
"caterpillar_height": 4,
- "dim_rec_v": 64, # dim_model / nb_heads
"nb_blocks": 6,
},
"37M": {
"dim_keys": 64,
"dim_hidden": 2048,
"nb_heads": 8,
- "dim_rec_v": 64,
"nb_blocks": 12,
},
"37M-C": {
"nb_heads": 8,
"nb_lines": 256,
"caterpillar_height": 32,
- "dim_rec_v": 64,
"nb_blocks": 12,
},
"122M": {
"dim_keys": 64,
"dim_hidden": 2048,
"nb_heads": 8,
- "dim_rec_v": 96,
"nb_blocks": 24,
},
"122M-C": {
"dim_hidden": 2048,
"nb_heads": 8,
"nb_lines": 128,
- "dim_rec_v": 96,
"nb_blocks": 24,
},
"352M": {
"dim_keys": 64,
"dim_hidden": 2048,
"nb_heads": 8,
- "dim_rec_v": 128,
"nb_blocks": 48,
},
"352M-C": {
"dim_hidden": 2048,
"nb_heads": 8,
"nb_lines": 128,
- "dim_rec_v": 128,
"nb_blocks": 48,
},
}
try:
os.mkdir(args.result_dir)
except FileExistsError:
- if not args.overwrite_results:
+ if not args.continue_training:
print(f"result directory {args.result_dir} already exists")
exit(1)
# 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
+ 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_min_learning_rate
+ return args.legacy_small_lr
# from nanoGPT
nb_heads=args.nb_heads,
nb_lines=args.nb_lines,
caterpillar_height=args.caterpillar_height,
- dim_rec_v=args.dim_rec_v,
nb_blocks=args.nb_blocks,
causal=True,
dropout=args.dropout,