| |
| """Frozen evaluator for the published proxy_v4.pt. |
| |
| Deterministically reconstructs the validation split the training recipe |
| defines (the three held-out panels: labeled rows routed by panel, the |
| val-panel fiber negatives, and the seed-0 procedural background negatives) |
| and scores the published checkpoint on it. |
| |
| Expected output, verified 2026-07-11 against the released proxy_v4.pt: |
| |
| val: 81 positives / 123 negatives |
| AUROC 0.9853457794 |
| |
| Usage: |
| python eval_checkpoint.py --checkpoint /path/to/proxy_v4.pt \ |
| --maps-dir /path/to/s1_official_panels --crops-dir ./crops |
| |
| Crops for the val rows must exist (run generate_crops.py on both |
| train_labels.jsonl and fiber_negatives_50.jsonl first); background-negative |
| crops are generated on the fly from --maps-dir, same as train.py. |
| """ |
| import argparse |
| import os |
| import sys |
|
|
| import numpy as np |
| import torch |
| import torchvision |
| from PIL import Image |
|
|
| HERE = os.path.dirname(os.path.abspath(__file__)) |
| sys.path.insert(0, HERE) |
| from train import (VAL_PANELS, WIN_S1, SZ, CropDataset, auroc, load_rows, |
| sample_background_negatives) |
| from torch.utils.data import DataLoader |
| import torch.nn as nn |
|
|
| Image.MAX_IMAGE_PIXELS = None |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser(description=__doc__) |
| ap.add_argument("--checkpoint", default="proxy_v4.pt") |
| ap.add_argument("--labels", default=os.path.join(HERE, "train_labels.jsonl")) |
| ap.add_argument("--fiber", default=os.path.join(HERE, "fiber_negatives_50.jsonl")) |
| ap.add_argument("--maps-dir", required=True) |
| ap.add_argument("--crops-dir", default="./crops") |
| args = ap.parse_args() |
|
|
| def crop_path(r): |
| p = os.path.join(args.crops_dir, f"{r['id']}.png") |
| if not os.path.exists(p): |
| raise FileNotFoundError(f"missing crop {p} -- run generate_crops.py first") |
| return p |
|
|
| val_items = [] |
| pos_by_panel = {} |
| for r in load_rows(args.labels): |
| if r["scroll"] != "s1" or r["split"] != "train": |
| continue |
| if r["label"] == "positive": |
| pos_by_panel.setdefault(r["panel_or_segment"], []).append((r["y"], r["x"])) |
| if r["panel_or_segment"] in VAL_PANELS: |
| val_items.append((crop_path(r), 1 if r["label"] == "positive" else 0, r["weight"])) |
| for r in load_rows(args.fiber): |
| if r["panel_or_segment"] in VAL_PANELS: |
| val_items.append((crop_path(r), 0, r["weight"])) |
|
|
| for panel, y, x in sample_background_negatives(args.maps_dir, pos_by_panel, seed=0): |
| if panel not in VAL_PANELS: |
| continue |
| crop_out = os.path.join(args.crops_dir, f"bg_{panel}_y{y}_x{x}.png") |
| if not os.path.exists(crop_out): |
| im = np.array(Image.open(os.path.join(args.maps_dir, panel + ".jpg")).convert("L")) |
| c = im[y:y + WIN_S1, x:x + WIN_S1].astype(np.float32) |
| active = c[c > 10] |
| if len(active) > 50: |
| lo, hi = np.percentile(active, [2, 99.5]) |
| c = np.clip((c - lo) / max(hi - lo, 1e-6), 0, 1) |
| else: |
| c = c / 255.0 |
| Image.fromarray((c * 255).astype(np.uint8)).resize((SZ, SZ), Image.BILINEAR).save(crop_out) |
| val_items.append((crop_out, 0, 1.0)) |
|
|
| n_pos = sum(l for _, l, _ in val_items) |
| print(f"val: {n_pos} positives / {len(val_items) - n_pos} negatives") |
|
|
| m = torchvision.models.resnet18() |
| m.fc = nn.Linear(512, 1) |
| m.load_state_dict(torch.load(args.checkpoint, map_location="cpu")) |
| m.eval() |
|
|
| ys, ss = [], [] |
| with torch.no_grad(): |
| for xb, yb, _ in DataLoader(CropDataset(val_items, aug=False), batch_size=64): |
| ss += list(torch.sigmoid(m(xb).squeeze(1)).numpy()) |
| ys += list(yb.numpy()) |
| print(f"AUROC {auroc(ys, ss):.10f}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
