File size: 4,948 Bytes
a62fa5d | 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 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 | """Validate every model config actually builds + runs a forward pass at IMG_SIZE=128
with ImageNet weights -- for BOTH strategy 2 (plain SMP) and strategy 3 (refinement stack).
Run this on the GPU box BEFORE launching the full matrix:
python validate_models.py
For each model it reports:
* whether strategy 2 builds/forwards (and at which encoder_depth: tries 5, falls back to 4)
* the encoder's out_channels (what the strategy-3 projection consumes)
* whether strategy 3 (PixelDRLMG_WithDecoder) builds/forwards
Anything marked FAIL must be fixed (or the model swapped) before you burn compute.
"""
from __future__ import annotations
import importlib.util
import json
import pathlib
import sys
import traceback
import torch
import segmentation_models_pytorch as smp
REPO = pathlib.Path(__file__).resolve().parent
IMG = 128
BATCH = 2
MANIFEST = json.loads((REPO / "params" / "models_manifest.json").read_text(encoding="utf-8"))
def load_runner():
"""Import the simplified runner module (module-level only; main() is guarded)."""
path = REPO / "foldsrunner_simplified_after_ablation.py"
spec = importlib.util.spec_from_file_location("foldsrunner_simplified_after_ablation", path)
mod = importlib.util.module_from_spec(spec)
sys.modules[spec.name] = mod # register BEFORE exec so @dataclass can resolve its module
spec.loader.exec_module(mod)
return mod
def short(exc: BaseException) -> str:
return f"{type(exc).__name__}: {str(exc)[:180]}"
def main() -> int:
runner = load_runner()
x = torch.randn(BATCH, 3, IMG, IMG)
results = []
for key, cfg in MANIFEST.items():
arch, enc, proj = cfg["arch"], cfg["encoder"], cfg["proj_dim"]
row = {"model": key, "arch": arch, "encoder": enc, "proj_dim": proj,
"depth": None, "out_channels": None, "s2": "FAIL", "s3": "FAIL"}
# ---- strategy 2: plain SMP model -------------------------------------
for depth in (5, 4):
try:
m2 = smp.create_model(arch=arch, encoder_name=enc, encoder_weights="imagenet",
encoder_depth=depth, in_channels=3, classes=1)
m2.eval()
with torch.no_grad():
y = m2(x)
row["depth"] = depth
row["out_channels"] = list(m2.encoder.out_channels)
row["s2"] = f"OK out={tuple(y.shape)}"
del m2
break
except Exception as exc: # noqa: BLE001
row["s2"] = f"FAIL(depth={depth}) {short(exc)}"
# ---- strategy 3: refinement stack ------------------------------------
if row["depth"] is not None:
try:
m3 = runner.PixelDRLMG_WithDecoder(
arch=arch, encoder_name=enc, encoder_weights="imagenet",
encoder_depth=row["depth"], proj_dim=proj, dropout_p=0.0,
)
m3.eval()
with torch.no_grad():
ctx = m3.prepare_refinement_context(x)
state = m3.forward_refinement_state(
ctx["base_features"], ctx["decoder_prob"].detach(), ctx["decoder_prob"],
ctx["mc_variance"], ctx["pred_entropy"],
encoder_features=ctx.get("encoder_features"),
)
policy, _ = m3.forward_from_state(state)
row["s3"] = f"OK state={tuple(state.shape)} policy={tuple(policy.shape)}"
del m3
except Exception as exc: # noqa: BLE001
row["s3"] = f"FAIL {short(exc)}"
traceback.print_exc()
results.append(row)
# ---- report --------------------------------------------------------------
print("\n" + "=" * 110)
print(f"MODEL VALIDATION @ {IMG}x{IMG}, ImageNet weights")
print("=" * 110)
for r in results:
ok = r["s2"].startswith("OK") and r["s3"].startswith("OK")
print(f"\n[{'PASS' if ok else 'FAIL'}] {r['model']} arch={r['arch']} encoder={r['encoder']}")
print(f" encoder_depth : {r['depth']}")
print(f" out_channels : {r['out_channels']}")
print(f" strategy 2 : {r['s2']}")
print(f" strategy 3 : {r['s3']}")
failed = [r["model"] for r in results if not (r["s2"].startswith("OK") and r["s3"].startswith("OK"))]
depth4 = [r["model"] for r in results if r["depth"] == 4]
print("\n" + "=" * 110)
if depth4:
print(f"NOTE: these need SMP_ENCODER_DEPTH = 4 (not 5): {depth4}")
print(" -> add -e 's/^SMP_ENCODER_DEPTH = .*/SMP_ENCODER_DEPTH = 4/' to that model's command.")
if failed:
print(f"FAILED: {failed} -- fix/swap these before running the matrix.")
return 1
print("ALL MODELS PASS.")
return 0
if __name__ == "__main__":
raise SystemExit(main())
|