File size: 2,519 Bytes
178f61f | 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 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 | """Quantify the historical strict=False teacher checkpoint key mismatch."""
from pathlib import Path
import json
import sys
ROOT = Path(r"E:\Gaze_estimation")
sys.path.insert(0, str(ROOT / ".codex_deps"))
sys.path.insert(0, str(ROOT))
import torch
from src.models.teacher import L2CS
checkpoint_path = ROOT / "checkpoints" / "resnet50.pt"
state = torch.load(checkpoint_path, map_location="cpu", weights_only=True)
model = L2CS(backbone_name="resnet50", pretrained=False)
result = model.load_state_dict(state, strict=False)
model_keys = set(model.state_dict())
checkpoint_keys = set(state)
loaded_keys = sorted(model_keys & checkpoint_keys)
prefix_map = {
"conv1.": "backbone.0.", "bn1.": "backbone.1.",
"layer1.": "backbone.4.", "layer2.": "backbone.5.",
"layer3.": "backbone.6.", "layer4.": "backbone.7.",
}
remapped = {}
for key, value in state.items():
new_key = key
for source, target in prefix_map.items():
if key.startswith(source):
new_key = target + key[len(source):]
break
remapped[new_key] = value
fixed_model = L2CS(backbone_name="resnet50", pretrained=False)
fixed_result = fixed_model.load_state_dict(remapped, strict=False)
report = {
"checkpoint": str(checkpoint_path),
"checkpoint_tensor_keys": len(checkpoint_keys),
"wrapper_tensor_keys": len(model_keys),
"historical_loader_loaded_key_count": len(loaded_keys),
"historical_loader_loaded_keys": loaded_keys,
"historical_loader_missing_key_count": len(result.missing_keys),
"historical_loader_unexpected_key_count": len(result.unexpected_keys),
"historical_loader_missing_keys": result.missing_keys,
"historical_loader_unexpected_keys": result.unexpected_keys,
"remapped_loader_missing_keys": fixed_result.missing_keys,
"remapped_loader_unexpected_keys": fixed_result.unexpected_keys,
"interpretation": "Historical strict=False loading accepts only identically named keys and silently leaves the ResNet backbone at initialization/pretrained values.",
}
out = ROOT / "artifacts" / "kd-teacher-trap-diagnostic" / "teacher_loader_key_audit.json"
out.write_text(json.dumps(report, indent=2), encoding="utf-8")
print(json.dumps({k: report[k] for k in (
"checkpoint_tensor_keys", "wrapper_tensor_keys", "historical_loader_loaded_key_count",
"historical_loader_loaded_keys", "historical_loader_missing_key_count",
"historical_loader_unexpected_key_count", "remapped_loader_missing_keys",
"remapped_loader_unexpected_keys")}, indent=2))
|