File size: 11,216 Bytes
857c2e9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Quick sanity check for VLAC trajectory values on the toy dataset.

The dataset is produced by ``testing/prepare_vlac_test_data.py`` in
``task_progress`` mode.  Each entry already includes the image paths (relative
``images/``) together with ground-truth progress numbers in ``[0, 1]``.

This script keeps things intentionally small and prints a short report for a set
of frame/reference configurations (e.g., 4×4, 4×8, 8×4, 8×8) so we can inspect
MAE, final-frame accuracy, and latency versus sequence length.
"""

from __future__ import annotations

import argparse
import base64
import itertools
import json
import sys
import time
from tqdm import tqdm
from io import BytesIO
from pathlib import Path
from typing import Dict, Iterable, List, Optional, Sequence

import numpy as np
import requests
from PIL import Image

# ---------------------------------------------------------------------------
# Helpers
# ---------------------------------------------------------------------------


def read_manifest(dataset_dir: Path, json_name: str) -> List[Dict]:
    manifest_path = dataset_dir / json_name
    images_dir = dataset_dir / "images"
    if not manifest_path.is_file():
        raise FileNotFoundError(f"Metadata JSON not found: {manifest_path}")
    if not images_dir.is_dir():
        raise FileNotFoundError(f"Images directory not found: {images_dir}")

    with manifest_path.open("r", encoding="utf-8") as f:
        raw_entries = json.load(f)

    entries: List[Dict] = []
    for entry in raw_entries:
        frames = entry.get("frames") or []
        if not frames:
            continue
        for frame in frames:
            frame["abs_path"] = str(images_dir / frame["path"])
        entry["reference"] = [str(images_dir / rel) for rel in entry.get("reference", [])]
        entries.append(entry)
    return entries


def image_to_base64(path: Path) -> str:
    with Image.open(path) as img:
        img = img.convert("RGB")
        buffer = BytesIO()
        img.save(buffer, format="JPEG", quality=95)
    return base64.b64encode(buffer.getvalue()).decode("utf-8")


def encode_images(paths: Iterable[str]) -> List[str]:
    return [image_to_base64(Path(p)) for p in paths]


def sample_fixed_interval_frames(image_list, num_frames):
    # sample num_frames frames from image_list
    # sample with equal interval while also ensuring the first and the last frames are included
    if len(image_list) == 0:
        raise ValueError("image_list is empty")
    elif len(image_list) == 1:
        return [image_list[0]] * num_frames
    elif num_frames == 2:
        return [image_list[0]] * (num_frames//2) + [image_list[-1]] * (num_frames//2)
    elif num_frames == 3:
        return [image_list[0]] + [image_list[1]] * (num_frames-2) + [image_list[-1]]
    else:
        total_frames = len(image_list)
        indices = np.linspace(start=0, stop=total_frames - 1, num=num_frames, dtype=int)
        sampled_frames = [image_list[i] for i in indices]
    return sampled_frames


def call_trajectory_critic(
    session: requests.Session,
    base_url: str,
    task: str,
    frames_b64: List[str],
    reference_b64: Optional[List[str]],
    timeout: float,
) -> Dict:
    payload = {
        "task": task,
        "frames": frames_b64,
        "reference": reference_b64,
        "ref_num": len(reference_b64 or []),
        "skip": 1,
        "batch_size": min(len(frames_b64), 8),
        "think": False,
        "return_video": False,
    }
    start = time.time()
    resp = session.post(f"{base_url.rstrip('/')}/trajectory-critic", json=payload, timeout=timeout)
    resp.raise_for_status()
    result = resp.json()
    result["latency_sec"] = time.time() - start
    return result


# ---------------------------------------------------------------------------
# Evaluation
# ---------------------------------------------------------------------------


def evaluate_combo(
    manifest: Sequence[Dict],
    base_url: str,
    timeout: float,
    frame_limit: int,
    ref_limit: int,
    done_threshold_list: list,
) -> Dict[str, float]:
    session = requests.Session()
    mae_values: List[float] = []
    latencies: List[float] = []
    total_frames = 0
    pred_last_value_list = []
    pred_mid_value_list = []

    for entry in tqdm(manifest):
        frames = entry["frames"]
        if len(frames) <= frame_limit:
            selected_frames = frames
        else:
            selected_frames = sample_fixed_interval_frames(frames, frame_limit)
        selected_frames_paths = [frame["abs_path"] for frame in selected_frames]
        frames_b64 = encode_images(selected_frames_paths)

        reference_paths = entry["reference"]
        if len(reference_paths) <= ref_limit:
            selected_reference_paths = reference_paths            
        else:
            selected_reference_paths = sample_fixed_interval_frames(reference_paths, ref_limit)
        reference_b64 = encode_images(selected_reference_paths)

        gt = np.array([frame["progress"] for frame in selected_frames], dtype=np.float32)

        try:
            result = call_trajectory_critic(
                session=session,
                base_url=base_url,
                task=entry.get("task", ""),
                frames_b64=frames_b64,
                reference_b64=reference_b64,
                timeout=timeout,
            )
        except requests.RequestException as exc:
            print(f"[warn] request failed for demo {entry.get('demo_id')}: {exc}")
            continue

        preds = np.array(result.get("value_list", []), dtype=np.float32)
        if preds.size == 0:
            continue

        # mid_idx = min(len(preds) // 2 + 1, len(preds) - 1)
        mid_idx = -2

        pred_last_value_list.append(preds[-1])
        pred_mid_value_list.append(preds[mid_idx])

        mae_values.append(float(np.mean(np.abs(preds[-1] - gt[-1]))))
        latencies.append(result.get("latency_sec", 0.0))
        total_frames += len(preds)

    accuracy_with_different_thresholds = {}
    for done_threshold in done_threshold_list:
        tp = fp = tn = fn = 0
        for pred_last, pred_mid in zip(pred_last_value_list, pred_mid_value_list):
            # Expected ground truth: last frame is positive, mid frame is negative
            if pred_last >= done_threshold:
                tp += 1
            else:
                fn += 1

            if pred_mid >= done_threshold:
                fp += 1
            else:
                tn += 1

        total = tp + fp + fn + tn
        precision = tp / (tp + fp) if (tp + fp) else float("nan")
        recall = tp / (tp + fn) if (tp + fn) else float("nan")
        if any(np.isnan(value) for value in (precision, recall)) or (precision + recall) == 0:
            f1 = float("nan")
        else:
            f1 = 2 * precision * recall / (precision + recall)
        accuracy = (tp + tn) / total if total else float("nan")

        accuracy_with_different_thresholds[done_threshold] = {
            "accuracy": accuracy,
            "precision": precision,
            "recall": recall,
            "f1": f1,
        }

    if not mae_values:
        return {
            "mae": float("nan"),
            "frames": 0,
            "latency": float("nan"),
            "final_accuracy": {},
        }

    return {
        "mae": float(np.mean(mae_values)),
        "frames": total_frames,
        "latency": float(np.mean(latencies)) if latencies else float("nan"),
        "final_accuracy": accuracy_with_different_thresholds,
    }


# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="VLAC trajectory sanity check")
    parser.add_argument("--dataset-dir", required=True, help="Directory containing images/ and dataset JSON")
    parser.add_argument("--json-name", default="dataset_frame_progress.json", help="Manifest filename")
    parser.add_argument("--base-url", default="http://localhost:8111", help="VLAC service base URL")
    parser.add_argument("--timeout", type=float, default=30.0, help="HTTP timeout in seconds")
    parser.add_argument("--max-demos", type=int, default=None, help="Evaluate only the first N demos")
    parser.add_argument(
        "--frame-counts",
        type=int,
        nargs="+",
        default=[4, 8],
        help="Number of trajectory frames to feed per call (default: 4 8)",
    )
    parser.add_argument(
        "--ref-counts",
        type=int,
        nargs="+",
        default=[4, 8],
        help="Number of reference frames to feed per call (default: 4 8)",
    )
    parser.add_argument(
        "--done-threshold-list",
        type=list,
        default=[50, 55, 60, 65, 70, 75, 80, 85, 90, 95],
        help="Threshold on progress for final-frame accuracy (default: 0.9)",
    )
    return parser.parse_args()


def main() -> int:
    args = parse_args()
    dataset_dir = Path(args.dataset_dir)
    try:
        manifest = read_manifest(dataset_dir, args.json_name)
    except FileNotFoundError as exc:
        print(exc)
        return 1

    if args.max_demos is not None:
        manifest = manifest[: args.max_demos]

    if not manifest:
        print("No demos found in the manifest. Regenerate the dataset with testing/prepare_vlac_test_data.py")
        return 1

    frame_counts = sorted(set(fc for fc in args.frame_counts if fc > 0))
    ref_counts = sorted(set(rc for rc in args.ref_counts if rc > 0))
    if not frame_counts or not ref_counts:
        print("Provide positive frame/reference counts.")
        return 1

    print(f"Loaded {len(manifest)} demos from {dataset_dir}")

    results: Dict[tuple, Dict[str, float]] = {}
    print("Threshold: ", args.done_threshold_list)
    for frame_limit, ref_limit in itertools.product(frame_counts, ref_counts):
        metrics = evaluate_combo(
            manifest=manifest,
            base_url=args.base_url,
            timeout=args.timeout,
            frame_limit=frame_limit,
            ref_limit=ref_limit,
            done_threshold_list=args.done_threshold_list,
        )
        results[(frame_limit, ref_limit)] = metrics

    print("\n=== Results by (frames, reference) ===")
    for (frame_limit, ref_limit), metrics in sorted(results.items()):
        mae = metrics["mae"]
        latency = metrics["latency"]
        final_acc = metrics["final_accuracy"]
        print(f"{frame_limit}x{ref_limit}")
        for threshold, stats in final_acc.items():
            acc = stats["accuracy"]
            precision = stats["precision"]
            recall = stats["recall"]
            f1 = stats["f1"]
            print(
                f"threshold {threshold}: "
                f"accuracy {acc:.3f}, precision {precision:.3f}, "
                f"recall {recall:.3f}, f1 {f1:.3f}"
            )
        print()
        print(
            f"frames={frame_limit:>2}, ref={ref_limit:>2} -> "
            f"MAE {mae:.4f}, avg latency {latency:.2f}s, frames used {metrics['frames']}"
        )
        print()

    return 0


if __name__ == "__main__":
    sys.exit(main())