File size: 8,856 Bytes
f894a3c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# /// script
# requires-python = ">=3.11"
# dependencies = ["datasets>=4.0.0", "huggingface-hub", "pillow", "matplotlib"]
# ///
"""
Local visual inspection for the output of qwen3vl-detect.py.

Given an output dataset (produced by qwen3vl-detect.py) and optionally its
source dataset (with ground-truth bboxes), renders side-by-side PNGs of
predicted detections vs ground truth, one per row.

Output: a /tmp/<slug>-viz/ directory of PNGs and a per-row text summary
(detection counts, label histograms, bbox value ranges).

Usage:
    uv run inspect-detections.py OUTPUT_DATASET [--split SPLIT] [--source SOURCE_DATASET]

If --source is provided, the script will also load the source dataset, pull
ground-truth `objects` (bbox + category), auto-discover class names from the
ClassLabel feature, and overlay GT boxes on a second panel for comparison.
"""

import argparse
import json
import os
from collections import Counter

import matplotlib.patches as mpatches
import matplotlib.pyplot as plt
from datasets import load_dataset


def render_box(ax, x1, y1, x2, y2, color, lw=1.2, label=None, fontsize=7):
    rect = mpatches.Rectangle(
        (x1, y1), x2 - x1, y2 - y1, fill=False, edgecolor=color, linewidth=lw
    )
    ax.add_patch(rect)
    if label:
        ax.text(
            x1,
            y1,
            label,
            fontsize=fontsize,
            color="white",
            bbox=dict(
                boxstyle="round,pad=0.15",
                facecolor=color,
                alpha=0.75,
                edgecolor="none",
            ),
            va="bottom",
            ha="left",
        )


def discover_class_names(ds) -> list[str]:
    """Pull class names from an `objects.category` or `objects.category_id`
    ClassLabel feature on a HF dataset. Returns [] if not found."""
    feats = ds.features
    if "objects" not in feats:
        return []
    obj_feat = feats["objects"]
    # objects may be List[Dict[...]] or Sequence(Dict[...]) — both expose `feature`
    inner = getattr(obj_feat, "feature", None) or obj_feat
    if not hasattr(inner, "keys"):
        return []
    for key in ("category", "category_id"):
        if key in inner:
            cat_feat = inner[key]
            # Sequence(ClassLabel) or ClassLabel
            cl = getattr(cat_feat, "feature", None) or cat_feat
            names = getattr(cl, "names", None)
            if names:
                return list(names)
    return []


def main() -> None:
    parser = argparse.ArgumentParser(
        description="Render Qwen3-VL detection output side-by-side with ground truth."
    )
    parser.add_argument("output_dataset", help="HF dataset ID with detections column")
    parser.add_argument(
        "--split", default="train", help="Split of output_dataset (default: train)"
    )
    parser.add_argument(
        "--source",
        default=None,
        help="Optional source dataset with ground-truth `objects` for GT overlay panel.",
    )
    parser.add_argument(
        "--source-split", default=None, help="Split of source dataset (default: same as --split)"
    )
    parser.add_argument(
        "--max-rows", type=int, default=None, help="Render only the first N rows"
    )
    parser.add_argument(
        "--out-dir",
        default=None,
        help="Output directory (default: /tmp/<output-slug>-viz/)",
    )
    args = parser.parse_args()

    slug = args.output_dataset.split("/")[-1]
    out_dir = args.out_dir or f"/tmp/{slug}-viz"
    os.makedirs(out_dir, exist_ok=True)

    print(f"Loading {args.output_dataset} split={args.split}…")
    ds = load_dataset(args.output_dataset, split=args.split)
    if args.max_rows:
        ds = ds.select(range(min(args.max_rows, len(ds))))
    print(f"rows={len(ds)}  cols={ds.column_names}")

    # qwen3vl-detect preserves the source `objects` column on each output row,
    # so GT is co-located with detections (no shuffled-index join needed).
    # The source dataset (if provided) is used only to recover ClassLabel names
    # since push_to_hub strips the ClassLabel typing on round-trip.
    gt_names: list[str] = []
    if args.source:
        gt_split = args.source_split or args.split
        print(f"Loading source {args.source} split={gt_split} for class names…")
        gt_src = load_dataset(args.source, split=gt_split)
        gt_names = discover_class_names(gt_src)
        print(f"GT classes ({len(gt_names)}): {gt_names}")
    elif "objects" in ds.column_names:
        gt_names = discover_class_names(ds)
        if gt_names:
            print(f"GT classes from output features: {gt_names}")

    for i in range(len(ds)):
        row = ds[i]
        img = row["image"]
        W, H = img.size
        info = json.loads(row["inference_info"])
        dets = row["detections"]

        # GT extraction from the output row itself (qwen3vl-detect preserves
        # all input columns including `objects`). Handles both list-of-dicts
        # and dict-of-lists shapes.
        gt_cats, gt_bbox = [], []
        if "objects" in row and row["objects"] is not None:
            objs = row["objects"]
            if isinstance(objs, dict):
                gt_cats = objs.get("category", objs.get("category_id", []))
                gt_bbox = objs.get("bbox", [])
            else:  # list of dicts
                for o in objs:
                    if "category" in o:
                        gt_cats.append(o["category"])
                    elif "category_id" in o:
                        gt_cats.append(o["category_id"])
                    gt_bbox.append(o["bbox"])

        bbox_vals = [v for d in dets for v in d["bbox"]]
        bbox_max = max(bbox_vals) if bbox_vals else 0
        bbox_min = min(bbox_vals) if bbox_vals else 0
        qwen_hist = Counter(d["label"] for d in dets)
        gt_hist = Counter(
            gt_names[c] if 0 <= c < len(gt_names) else f"#{c}" for c in gt_cats
        ) if gt_names else Counter()

        print(
            f"\n=== row {i}  image_id={row.get('image_id', '?')}  orig={W}x{H}  "
            f"inference_image_size={info['image_size']} ==="
        )
        if gt_cats:
            print(f"  GT: {len(gt_cats)} objects  {dict(gt_hist)}")
        print(f"  Qwen: {len(dets)} detections  {dict(qwen_hist)}")
        print(f"  Qwen bbox range: [{bbox_min:.0f}, {bbox_max:.0f}]  (image {W}x{H})")

        # Render
        n_panels = 2 if gt_cats else 1
        fig, axes = plt.subplots(1, n_panels, figsize=(10 * n_panels, 12))
        if n_panels == 1:
            axes = [axes]
        for ax in axes:
            ax.imshow(img)
            ax.set_xlim(0, W)
            ax.set_ylim(H, 0)
            ax.set_aspect("equal")
            ax.axis("off")

        # v1.1+ output stores bbox in pixel coords; v1 stored 0-1000 normalised.
        needs_denorm = bbox_max <= 1001 and max(W, H) > 1001
        sx = (W / 1000.0) if needs_denorm else 1.0
        sy = (H / 1000.0) if needs_denorm else 1.0
        title_extra = "(0-1000 → pixels)" if needs_denorm else "(pixel coords)"

        axes[0].set_title(
            f"Detections (n={len(dets)}) {title_extra}  range [{bbox_min:.0f}, {bbox_max:.0f}]",
            fontsize=11,
        )
        for d in dets:
            b = d["bbox"]
            if len(b) == 4:
                render_box(axes[0], b[0] * sx, b[1] * sy, b[2] * sx, b[3] * sy, "#E03030", lw=1.0)
        for d in dets[:10]:
            b = d["bbox"]
            if len(b) == 4:
                render_box(
                    axes[0],
                    b[0] * sx,
                    b[1] * sy,
                    b[2] * sx,
                    b[3] * sy,
                    "#E03030",
                    lw=1.4,
                    label=d["label"][:18],
                    fontsize=8,
                )

        if gt_cats and len(axes) > 1:
            palette = [
                "#1f78b4", "#33a02c", "#ff7f00", "#6a3d9a", "#b15928",
                "#e31a1c", "#fb9a99", "#a6cee3", "#b2df8a", "#fdbf6f",
                "#cab2d6", "#ffff99",
            ]
            color_by_name: dict[str, str] = {}
            axes[1].set_title(f"Ground truth (n={len(gt_cats)})", fontsize=11)
            for c, b in zip(gt_cats, gt_bbox):
                if len(b) != 4:
                    continue
                name = gt_names[c] if 0 <= c < len(gt_names) else f"#{c}"
                color = color_by_name.setdefault(name, palette[len(color_by_name) % len(palette)])
                x, y, w_, h_ = b  # COCO xywh
                render_box(axes[1], x, y, x + w_, y + h_, color, lw=2.0, label=name[:18], fontsize=8)

        plt.tight_layout()
        path = f"{out_dir}/row{i}_id{row.get('image_id', i)}.png"
        plt.savefig(path, dpi=70, bbox_inches="tight")
        plt.close()

    print(f"\nDone. {len(ds)} rows. Open: {out_dir}")


if __name__ == "__main__":
    main()