File size: 19,198 Bytes
a48a226
5e88f2a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a48a226
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5e88f2a
a48a226
 
5e88f2a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a48a226
 
 
5e88f2a
 
 
 
 
 
 
 
 
a48a226
5e88f2a
a48a226
 
 
 
 
 
5e88f2a
 
 
 
 
 
 
 
 
 
 
 
 
 
a48a226
5e88f2a
a48a226
 
 
 
 
5e88f2a
 
a48a226
 
 
5e88f2a
a48a226
 
5e88f2a
 
 
 
 
 
 
 
 
 
 
 
 
 
a48a226
 
 
5e88f2a
 
a48a226
 
 
 
 
 
 
 
 
 
 
5e88f2a
 
 
 
 
 
 
 
 
 
a48a226
 
 
 
 
 
5e88f2a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a48a226
 
5e88f2a
a48a226
5e88f2a
 
a48a226
 
5e88f2a
 
 
a48a226
 
 
 
 
 
 
 
 
5e88f2a
a48a226
 
5e88f2a
 
 
 
a48a226
5e88f2a
 
 
a48a226
5e88f2a
 
a48a226
 
5e88f2a
 
a48a226
 
5e88f2a
a48a226
 
 
 
5e88f2a
 
 
 
 
 
 
 
 
 
 
a48a226
5e88f2a
a48a226
 
5e88f2a
 
 
 
a48a226
 
5e88f2a
 
a48a226
 
 
 
5e88f2a
 
a48a226
 
5e88f2a
a48a226
5e88f2a
 
a48a226
5e88f2a
a48a226
 
 
 
5e88f2a
a48a226
 
57fa990
 
 
 
 
 
 
 
 
 
 
5e88f2a
 
 
a48a226
5e88f2a
 
a48a226
5e88f2a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a48a226
5e88f2a
 
 
 
 
 
a48a226
5e88f2a
 
 
 
 
 
 
a48a226
 
5e88f2a
 
 
 
 
 
 
 
 
 
 
 
 
57fa990
5e88f2a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a48a226
 
 
 
 
 
 
5e88f2a
a48a226
5e88f2a
 
 
a48a226
 
 
5e88f2a
a48a226
 
 
5e88f2a
a48a226
 
5e88f2a
 
a48a226
 
5e88f2a
 
a48a226
 
5e88f2a
 
a48a226
 
5e88f2a
 
a48a226
 
5e88f2a
 
 
 
 
 
a48a226
 
 
5e88f2a
 
 
 
a48a226
5e88f2a
a48a226
 
5e88f2a
 
 
 
 
 
 
 
 
 
a48a226
 
5e88f2a
 
 
 
 
a48a226
5e88f2a
 
 
 
 
 
 
 
 
 
a48a226
57fa990
a48a226
 
5e88f2a
a48a226
 
 
5e88f2a
57fa990
 
 
 
 
 
5e88f2a
 
a48a226
5e88f2a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a48a226
 
 
 
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
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
"""
sen2sr_evaluate.py
==================
Evaluate WEO-SAS/sen2sr variants using the opensr-test benchmark suite.

Metrics computed per variant × dataset:
  - reflectance  (↓)  L1 distance — radiometric fidelity
  - spectral     (↓)  Spectral Angle Distance — colour consistency
  - spatial      (↓)  Phase Correlation — geometric stability
  - synthesis    (↑)  High-frequency detail added
  - hallucination(↓)  False details not in HR
  - omission     (↓)  Real details missing from SR
  - improvement  (↑)  Correct new details introduced

Usage
-----
    pip install opensr-test huggingface_hub sen2sr safetensors rasterio

    # Evaluate everything (RGBN-compatible datasets + variants)
    python sen2sr_evaluate.py

    # Specific variants and/or datasets
    python sen2sr_evaluate.py --variants main mamba-rgbn-x4 --datasets naip spot

    # Skip download if already cached
    python sen2sr_evaluate.py --cache-dir ./model_cache

Notes
-----
- RGBN variants (main, lite-rgbn-x4, mamba-rgbn-x4) are evaluated on all
  opensr-test datasets (NAIP, SPOT, Venus, Spain Crops, Spain Urban).
- Full-pipeline 10-band variants (lite-main, mamba-main) and RSWIR variants
  (lite-rswir-x2, mamba-rswir-x2) require all 10 Sentinel-2 bands.
  opensr-test only provides 4-band RGBN patches, so these variants use the
  4 RGBN bands for input and the remaining 6 channels are zero-padded.
  For a fair evaluation of those variants, use your own 10-band Sentinel-2
  tiles and call evaluate_custom() directly.
"""

from __future__ import annotations

import argparse
import csv
import json
import sys
from pathlib import Path
from typing import Dict, List, Optional

import numpy as np
import torch


# ---------------------------------------------------------------------------
# Variant registry
# ---------------------------------------------------------------------------

VARIANTS: Dict[str, dict] = {
    "main": {
        "repo_id":     "WEO-SAS/sen2sr",
        "revision":    None,
        "in_channels": 4,
        "scale":       4,
        "note":        "SEN2SRLite RGBN 4x (CNN)",
    },
    "lite-rswir-x2": {
        "repo_id":     "WEO-SAS/sen2sr",
        "revision":    "lite-rswir-x2",
        "in_channels": 10,
        "scale":       2,
        "note":        "SEN2SRLite RSWIR 2x (CNN) — zero-pads channels 4-9",
    },
    "lite-main": {
        "repo_id":     "WEO-SAS/sen2sr",
        "revision":    "lite-main",
        "in_channels": 10,
        "scale":       4,
        "note":        "SEN2SRLite full 10-band 4x (CNN) — zero-pads channels 4-9",
    },
    "mamba-rgbn-x4": {
        "repo_id":     "WEO-SAS/sen2sr",
        "revision":    "mamba-rgbn-x4",
        "in_channels": 4,
        "scale":       4,
        "note":        "SEN2SR RGBN 4x (Mamba)",
    },
    "mamba-rswir-x2": {
        "repo_id":     "WEO-SAS/sen2sr",
        "revision":    "mamba-rswir-x2",
        "in_channels": 10,
        "scale":       2,
        "note":        "SEN2SR RSWIR 2x (Swin2SR) — zero-pads channels 4-9",
    },
    "mamba-main": {
        "repo_id":     "WEO-SAS/sen2sr",
        "revision":    "mamba-main",
        "in_channels": 10,
        "scale":       4,
        "note":        "SEN2SR full 10-band 4x (Mamba+Swin) — zero-pads channels 4-9",
    },
    "srresnet": {
        "repo_id":     "WEO-SAS/srresnet",
        "revision":    None,
        "in_channels": 4,
        "scale":       4,
        "note":        "SRResNet RGBN→RGB 4x (baseline)",
    },
}

DATASETS = ["naip", "spot", "venus", "spain_crops", "spain_urban"]

# Canonical output column names → actual opensr_test.Metrics key
METRIC_MAP = {
    "reflectance":  "reflectance",
    "spectral":     "spectral",
    "spatial":      "spatial",
    "synthesis":    "synthesis",
    "hallucination": "ha_metric",
    "omission":     "om_metric",
    "improvement":  "im_metric",
}
METRIC_COLS = list(METRIC_MAP.keys())


# ---------------------------------------------------------------------------
# Model loading
# ---------------------------------------------------------------------------

def load_model(variant: str, cache_dir: str, local_models_dir: Optional[str] = None):
    """Load a WEO-SAS model variant from a local dir or by downloading from HF Hub."""
    if local_models_dir:
        local_dir = str(Path(local_models_dir) / variant)
        if not Path(local_dir).is_dir():
            raise FileNotFoundError(f"Model dir not found: {local_dir}")
    else:
        from huggingface_hub import snapshot_download
        repo_id  = VARIANTS[variant].get("repo_id", "WEO-SAS/sen2sr")
        revision = VARIANTS[variant]["revision"]
        kwargs   = dict(repo_id=repo_id, local_dir=f"{cache_dir}/{variant}")
        if revision:
            kwargs["revision"] = revision
        local_dir = snapshot_download(**kwargs)

    sys.path.insert(0, local_dir)

    # Clear any cached module from a previous variant
    for mod in ["model", "sen2sr_pt", "predictor", "base"]:
        sys.modules.pop(mod, None)

    from model import Model  # noqa: PLC0415
    return Model(local_dir=local_dir)


# ---------------------------------------------------------------------------
# Inference helpers
# ---------------------------------------------------------------------------

def _pad_to_multiple(arr: np.ndarray, multiple: int) -> tuple:
    """Pad (C, H, W) to the next multiple of `multiple`; return (padded, orig_h, orig_w)."""
    _, h, w = arr.shape
    h_pad = ((h + multiple - 1) // multiple) * multiple
    w_pad = ((w + multiple - 1) // multiple) * multiple
    if h_pad == h and w_pad == w:
        return arr, h, w
    padded = np.zeros((arr.shape[0], h_pad, w_pad), dtype=arr.dtype)
    padded[:, :h, :w] = arr
    return padded, h, w


def run_sr(model, lr_np: np.ndarray, in_channels: int, scale: int = 4,
           patch_size: int = 128) -> np.ndarray:
    """
    Run SR on a single LR patch.

    lr_np   : (C_avail, H, W) float32 in [0, 1]   — opensr-test provides C=4 (RGBN)
    Returns : (C_out, H*scale, W*scale) float32, cropped to exact expected size
    """
    C_avail = lr_np.shape[0]

    if in_channels == C_avail:
        inp = lr_np
    elif in_channels > C_avail:
        pad = np.zeros((in_channels - C_avail,) + lr_np.shape[1:], dtype=np.float32)
        inp = np.concatenate([lr_np, pad], axis=0)
    else:
        inp = lr_np[:in_channels]

    # Pre-pad to patch_size so HardConstraint sees consistent LR↔SR sizes
    orig_h, orig_w = inp.shape[1], inp.shape[2]
    inp, _, _ = _pad_to_multiple(inp, patch_size)

    sr = model.predict(inp)

    # Crop to exact expected size based on original (unpadded) LR dimensions
    h_out = orig_h * scale
    w_out = orig_w * scale
    return sr[:, :h_out, :w_out]


# ---------------------------------------------------------------------------
# Per-dataset evaluation
# ---------------------------------------------------------------------------

def _save_comparison(
    lr: np.ndarray,
    sr: np.ndarray,
    hr: np.ndarray,
    path: Path,
    title: str,
    variant: str,
) -> None:
    try:
        import matplotlib
        matplotlib.use("Agg")
        import matplotlib.pyplot as plt
        from skimage.transform import resize as sk_resize

        def to_rgb(arr):
            rgb = np.clip(arr[:3].transpose(1, 2, 0), 0, 1)
            return (rgb * 255).astype(np.uint8)

        hr_h, hr_w = hr.shape[1], hr.shape[2]
        lr_big = sk_resize(to_rgb(lr), (hr_h, hr_w), order=1, preserve_range=True).astype(np.uint8)
        sr_rgb = to_rgb(sr)
        hr_rgb = to_rgb(hr)

        fig, axes = plt.subplots(1, 3, figsize=(12, 4))
        for ax, img, label in zip(
            axes,
            [lr_big, sr_rgb, hr_rgb],
            ["LR (bicubic)", f"SR ({variant})", "HR (reference)"],
        ):
            ax.imshow(img)
            ax.set_title(label, fontsize=10)
            ax.axis("off")
        fig.suptitle(f"{variant}{title}", fontsize=12, fontweight="bold")
        plt.tight_layout()
        path.parent.mkdir(parents=True, exist_ok=True)
        plt.savefig(path, dpi=100, bbox_inches="tight")
        plt.close(fig)
        print(f"    Saved image: {path.name}")
    except Exception as e:
        print(f"    [WARN] Could not save image: {e}")


def evaluate_dataset(
    model,
    variant:     str,
    dataset_name: str,
    max_samples: Optional[int] = None,
    save_images_dir: Optional[Path] = None,
) -> Dict[str, float]:
    """
    Run a variant against one opensr-test dataset and return mean metrics.

    Returns a dict mapping metric name → mean value, or empty dict on error.
    """
    try:
        import opensr_test
    except ImportError:
        raise ImportError("pip install opensr-test")

    try:
        dataset = opensr_test.load(dataset_name)
    except Exception as e:
        print(f"    [WARN] Could not load dataset '{dataset_name}': {e}")
        return {}

    # opensr-test dataset is a dict: {"L2A": (N,C,H,W) uint16, "HRharm": (N,C,H,W) uint16}
    lr_all = dataset["L2A"]
    hr_all = dataset["HRharm"]

    metrics_obj = opensr_test.Metrics()
    vinfo       = VARIANTS[variant]
    in_ch       = vinfo["in_channels"]
    scale       = vinfo["scale"]
    accum: Dict[str, list] = {m: [] for m in METRIC_COLS}
    n = lr_all.shape[0] if max_samples is None else min(max_samples, lr_all.shape[0])
    saved_image = False

    for i in range(n):
        lr = lr_all[i].astype(np.float32) / 10000.0  # (C, H, W) → [0, 1]
        hr = hr_all[i].astype(np.float32) / 10000.0

        try:
            sr = run_sr(model, lr, in_ch, scale)
        except Exception as e:
            print(f"    [WARN] SR failed on sample {i}: {e}")
            continue

        # For x2 models on 4x datasets: SR is half the HR size — skip metrics
        if sr.shape[1] != hr.shape[1] or sr.shape[2] != hr.shape[2]:
            if i == 0:
                print(f"    [SKIP] SR {sr.shape} != HR {hr.shape} — scale mismatch, skipping dataset")
            continue

        if save_images_dir and not saved_image:
            img_path = save_images_dir / f"{variant}_{dataset_name}.png"
            _save_comparison(lr, sr, hr, img_path, dataset_name, variant)
            saved_image = True

        lr_t = torch.from_numpy(lr)
        sr_t = torch.from_numpy(sr)
        hr_t = torch.from_numpy(hr)

        # Align channels: metrics require lr/sr/hr to have the same count
        min_ch = min(lr_t.shape[0], sr_t.shape[0], hr_t.shape[0])
        lr_t, sr_t, hr_t = lr_t[:min_ch], sr_t[:min_ch], hr_t[:min_ch]

        try:
            result = metrics_obj.compute(lr=lr_t, sr=sr_t, hr=hr_t)
            if not isinstance(result, dict):
                result = vars(result) if hasattr(result, "__dict__") else {}
        except Exception as e:
            print(f"    [WARN] Metrics failed on sample {i}: {e}")
            continue

        for col, api_key in METRIC_MAP.items():
            val = result.get(api_key)
            if val is not None:
                v = float(val.mean()) if hasattr(val, "mean") else float(val)
                accum[col].append(v)

        if (i + 1) % 10 == 0:
            print(f"    {i+1}/{n} samples processed", end="\r")

    print()
    return {m: float(np.mean(vs)) if vs else float("nan") for m, vs in accum.items()}


# ---------------------------------------------------------------------------
# HF output helpers
# ---------------------------------------------------------------------------

def _nan_to_null(obj):
    """Recursively replace float NaN with None so json.dump produces valid JSON."""
    if isinstance(obj, float) and np.isnan(obj):
        return None
    if isinstance(obj, dict):
        return {k: _nan_to_null(v) for k, v in obj.items()}
    if isinstance(obj, list):
        return [_nan_to_null(v) for v in obj]
    return obj


def build_eval_json(rows: list) -> dict:
    """Build eval_results.json dict from accumulated CSV rows."""
    from collections import defaultdict

    per_dataset: dict = {}
    agg: dict = defaultdict(lambda: defaultdict(list))

    for row in rows:
        v   = row["variant"]
        ds  = row["dataset"]
        per_dataset.setdefault(ds, {})
        m_vals = {}
        for m in METRIC_COLS:
            val = row.get(m, float("nan"))
            m_vals[m] = val
            if not (isinstance(val, float) and np.isnan(val)):
                agg[v][m].append(val)
        per_dataset[ds][v] = m_vals

    aggregate = {
        v: {m: float(np.mean(vs)) if vs else float("nan") for m, vs in metrics.items()}
        for v, metrics in agg.items()
    }

    variants_meta = {
        v: {"note": VARIANTS[v]["note"], "in_channels": VARIANTS[v]["in_channels"],
            "scale": VARIANTS[v]["scale"]}
        for v in VARIANTS
        if v in agg
    }

    return {
        "eval_type": "super_resolution",
        "model_name": "SEN2SR",
        "variants": variants_meta,
        "per_dataset": per_dataset,
        "aggregate": aggregate,
    }


def push_to_hf(
    eval_json: dict,
    images_dir: Optional[Path],
    csv_path: str,
    hf_token: str,
    commit_message: str = "eval: update benchmark results",
) -> None:
    from huggingface_hub import HfApi
    api = HfApi(token=hf_token)

    repo_id = "WEO-SAS/sen2sr"

    # Push eval_results.json
    eval_str = json.dumps(_nan_to_null(eval_json), indent=2)
    api.upload_file(
        path_or_fileobj=eval_str.encode(),
        path_in_repo="eval_results.json",
        repo_id=repo_id,
        repo_type="model",
        commit_message=commit_message,
    )
    print("Pushed eval_results.json")

    # Push CSV
    if Path(csv_path).exists():
        api.upload_file(
            path_or_fileobj=csv_path,
            path_in_repo=f"eval/{Path(csv_path).name}",
            repo_id=repo_id,
            repo_type="model",
            commit_message=commit_message,
        )
        print(f"Pushed eval/{Path(csv_path).name}")

    # Push images
    if images_dir and images_dir.exists():
        for img_path in sorted(images_dir.glob("*.png")):
            api.upload_file(
                path_or_fileobj=str(img_path),
                path_in_repo=f"eval_images/{img_path.name}",
                repo_id=repo_id,
                repo_type="model",
                commit_message=commit_message,
            )
            print(f"Pushed eval_images/{img_path.name}")


# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------

def main():
    parser = argparse.ArgumentParser(description="Evaluate WEO-SAS/sen2sr variants")
    parser.add_argument(
        "--variants", nargs="+", default=list(VARIANTS.keys()),
        choices=list(VARIANTS.keys()),
        help="Variants to evaluate (default: all)",
    )
    parser.add_argument(
        "--datasets", nargs="+", default=DATASETS, choices=DATASETS,
        help="Datasets to use (default: all)",
    )
    parser.add_argument(
        "--max-samples", type=int, default=None,
        help="Cap samples per dataset (useful for a quick smoke-test)",
    )
    parser.add_argument(
        "--cache-dir", default="./sen2sr_model_cache",
        help="Directory to cache downloaded model weights",
    )
    parser.add_argument(
        "--local-models-dir", default=None,
        help="Use pre-downloaded models instead of HF Hub (subdir per variant: main/, lite-main/, etc.)",
    )
    parser.add_argument(
        "--output", default="sen2sr_eval_results.csv",
        help="Output CSV path",
    )
    parser.add_argument(
        "--images-dir", default="./eval_images",
        help="Directory for visual comparison PNG files",
    )
    parser.add_argument(
        "--hf-token", default=None,
        help="HuggingFace write token (or set HF_TOKEN env var)",
    )
    parser.add_argument(
        "--no-push", action="store_true",
        help="Skip HF push (dry-run)",
    )
    args = parser.parse_args()

    import os
    hf_token = args.hf_token or os.environ.get("HF_TOKEN")
    images_dir = Path(args.images_dir)
    images_dir.mkdir(parents=True, exist_ok=True)

    Path(args.cache_dir).mkdir(parents=True, exist_ok=True)
    rows = []

    for variant in args.variants:
        print(f"\n{'='*60}")
        print(f"Variant: {variant}  ({VARIANTS[variant]['note']})")
        print(f"{'='*60}")

        try:
            print("  Loading model...")
            model = load_model(variant, args.cache_dir, args.local_models_dir)
        except Exception as e:
            print(f"  [ERROR] Could not load model: {e}")
            continue

        for ds in args.datasets:
            print(f"  Dataset: {ds}")
            metrics = evaluate_dataset(model, variant, ds, args.max_samples, images_dir)
            if not metrics:
                continue

            row = {"variant": variant, "dataset": ds}
            row.update(metrics)
            rows.append(row)

            # Pretty-print
            print(f"  {'Metric':<16} {'Value':>10}")
            print(f"  {'-'*28}")
            for m in METRIC_COLS:
                arrow = "↑" if m in ("synthesis", "improvement") else "↓"
                print(f"  {m:<16} {metrics.get(m, float('nan')):>9.4f} {arrow}")

    # Save CSV + eval_results.json
    if rows:
        fieldnames = ["variant", "dataset"] + METRIC_COLS
        with open(args.output, "w", newline="") as f:
            writer = csv.DictWriter(f, fieldnames=fieldnames)
            writer.writeheader()
            writer.writerows(rows)
        print(f"\nResults saved to: {args.output}")

        eval_json = build_eval_json(rows)
        json_path = Path(args.output).parent / "eval_results.json"
        with open(json_path, "w") as f:
            json.dump(_nan_to_null(eval_json), f, indent=2)
        print(f"Results saved to: {json_path}")
    else:
        print("\nNo results to save.")

    # Summary table
    if rows:
        print("\n" + "="*60)
        print("SUMMARY — mean across all datasets")
        print("="*60)
        from collections import defaultdict
        agg: dict = defaultdict(lambda: defaultdict(list))
        for row in rows:
            for m in METRIC_COLS:
                v = row.get(m, float("nan"))
                if not np.isnan(v):
                    agg[row["variant"]][m].append(v)

        header = f"{'Variant':<20}" + "".join(f"{m[:8]:>11}" for m in METRIC_COLS)
        print(header)
        print("-" * len(header))
        for variant in args.variants:
            if variant not in agg:
                continue
            vals = "".join(
                f"{np.mean(agg[variant].get(m, [float('nan')])):>11.4f}"
                for m in METRIC_COLS
            )
            print(f"{variant:<20}{vals}")

    # Push to HF
    if rows and not args.no_push:
        if not hf_token:
            print("\n[WARN] No HF token — skipping push. Pass --hf-token or set HF_TOKEN.")
        else:
            print("\nPushing results to HuggingFace...")
            push_to_hf(eval_json, images_dir, args.output, hf_token)
            print("Done.")


if __name__ == "__main__":
    main()