Files
light-petal-recognition/pc_training/average_hls_qcnn_checkpoints.py

84 lines
2.8 KiB
Python

"""Average HLSQConvNet checkpoints to reduce validation jitter."""
from __future__ import annotations
import argparse
import json
from pathlib import Path
import torch
from hls_qcnn import create_hls_qcnn_from_checkpoint
from train_hls_qcnn import configure_stdio
configure_stdio()
def main() -> int:
parser = argparse.ArgumentParser(description="Average compatible HLS QCNN checkpoints")
parser.add_argument("--checkpoints", nargs="+", required=True)
parser.add_argument("--output", default="../data/models/hls_qcnn_averaged.pth")
parser.add_argument("--metrics", default="../results/hls_qcnn_averaged_manifest.json")
args = parser.parse_args()
paths = [Path(item) for item in args.checkpoints]
missing = [str(path) for path in paths if not path.exists()]
if missing:
print("Missing checkpoints:")
for item in missing:
print(f" {item}")
return 2
checkpoints = [torch.load(path, map_location="cpu") for path in paths]
base_config = checkpoints[0].get("model_config")
for path, checkpoint in zip(paths, checkpoints):
if checkpoint.get("model_config") != base_config:
print(f"Model config mismatch: {path}")
return 2
avg_state = {}
state_keys = checkpoints[0]["model_state_dict"].keys()
for key in state_keys:
tensors = [checkpoint["model_state_dict"][key] for checkpoint in checkpoints]
if torch.is_floating_point(tensors[0]):
avg_state[key] = torch.stack([tensor.float() for tensor in tensors], dim=0).mean(dim=0)
avg_state[key] = avg_state[key].to(dtype=tensors[0].dtype)
else:
avg_state[key] = tensors[-1].clone()
model = create_hls_qcnn_from_checkpoint({"model_config": base_config})
model.load_state_dict(avg_state)
output = Path(args.output)
output.parent.mkdir(parents=True, exist_ok=True)
best_acc = max(float(checkpoint.get("best_acc", 0.0)) for checkpoint in checkpoints)
torch.save(
{
"epoch": max(int(checkpoint.get("epoch", 0)) for checkpoint in checkpoints),
"best_acc": best_acc,
"model_state_dict": avg_state,
"model_config": base_config,
"averaged_from": [str(path) for path in paths],
},
output,
)
metrics = {
"output": str(output),
"averaged_from": [str(path) for path in paths],
"source_best_acc": best_acc,
"model_config": base_config,
}
metrics_path = Path(args.metrics)
metrics_path.parent.mkdir(parents=True, exist_ok=True)
with metrics_path.open("w", encoding="utf-8") as f:
json.dump(metrics, f, indent=2, ensure_ascii=False)
print(f"Wrote averaged checkpoint: {output}")
print(f"Wrote manifest: {metrics_path}")
return 0
if __name__ == "__main__":
raise SystemExit(main())