File size: 1,756 Bytes
11d8066 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 | """Reassemble a full text-classification state_dict from a ParaGen-Bench pool checkpoint and its base checkpoint.
Pool files (ckpt/text_classification/<arch>__<dataset>/<model_id>/collect_XXX.pth) omit the vocabulary token-embedding
matrix, which was frozen during collection and is therefore identical to the base model's. It is stored once, in the
base checkpoint zoo/<dataset>/<arch>/<model_id>.pth. Every pool file lists the omitted keys under "frozen_keys".
from assemble import assemble
sd = assemble("zoo/agnews/bertbase/bertbase_010.pth",
"ckpt/text_classification/bertbase__agnews/bertbase_010/collect_000.pth")
model.load_state_dict(sd, strict=True)
"""
import torch
def assemble(base_ckpt_path, pool_ckpt_path, map_location="cpu"):
"""Return the full model state_dict: pool weights + the frozen keys copied from the base's model_state_dict."""
pool = torch.load(pool_ckpt_path, map_location=map_location, weights_only=False)
base = torch.load(base_ckpt_path, map_location=map_location, weights_only=False)
sd = dict(pool["model_state_dict"])
base_sd = base["model_state_dict"]
for k in pool["frozen_keys"]:
if k in sd:
raise ValueError(f"frozen key {k!r} is unexpectedly present in the pool file")
if k not in base_sd:
raise KeyError(f"frozen key {k!r} not found in the base checkpoint {base_ckpt_path}")
sd[k] = base_sd[k].to(torch.float32)
return sd
if __name__ == "__main__":
import sys
if len(sys.argv) != 3:
raise SystemExit("usage: python assemble.py <base_ckpt.pth> <collect_XXX.pth>")
sd = assemble(sys.argv[1], sys.argv[2])
print(f"{len(sd)} tensors, {sum(v.numel() for v in sd.values()):,} values")
|