File size: 22,663 Bytes
c5d818e
f2f112a
c5d818e
f2f112a
6feb3b2
f2f112a
c5d818e
 
f2f112a
 
 
 
 
c5d818e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f2f112a
 
 
c5d818e
856cffb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c5d818e
 
856cffb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f2f112a
c5d818e
f2f112a
c5d818e
 
 
 
 
 
 
 
f2f112a
 
 
 
 
 
c5d818e
f2f112a
 
 
 
 
 
 
 
 
 
 
 
 
 
c5d818e
f2f112a
 
 
 
 
 
c5d818e
 
 
f2f112a
c5d818e
6feb3b2
 
c5d818e
f2f112a
 
 
c5d818e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6feb3b2
c5d818e
6feb3b2
c5d818e
6feb3b2
 
c5d818e
6feb3b2
c5d818e
 
 
6feb3b2
c5d818e
 
 
f2f112a
c5d818e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f2f112a
 
c5d818e
f2f112a
 
 
 
 
 
 
c5d818e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6feb3b2
f2f112a
 
 
 
c5d818e
 
 
f2f112a
 
 
c5d818e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f2f112a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c5d818e
f2f112a
 
 
 
 
 
 
 
 
 
 
c5d818e
 
 
 
 
 
 
 
 
 
f2f112a
 
 
 
 
 
c5d818e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f2f112a
 
 
 
 
 
 
 
 
 
 
 
044bb04
 
 
 
f2f112a
c5d818e
 
f2f112a
c5d818e
 
 
 
 
 
 
 
e4172fe
f2f112a
 
 
c5d818e
f2f112a
 
 
 
c5d818e
f2f112a
 
 
 
 
 
c5d818e
f2f112a
 
c5d818e
f2f112a
 
c5d818e
 
f2f112a
 
 
c5d818e
6feb3b2
f2f112a
 
 
 
 
 
 
 
 
c5d818e
 
 
 
 
 
 
f2f112a
c5d818e
 
 
 
 
 
 
 
 
 
f2f112a
c5d818e
f2f112a
 
c5d818e
 
f2f112a
 
c5d818e
f2f112a
 
c5d818e
 
 
 
 
af40191
 
 
 
 
 
 
 
f2f112a
 
c5d818e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# runner.py
import json
import sys
from pathlib import Path
from typing import Dict, Any, List, Tuple, Optional

from threading import RLock

import numpy as np
import torch
from PIL import Image, ImageDraw

from ultralytics import YOLO, SAM
import open_clip

# --- ImageBind import: robust for both "pip install -e ImageBind" and local folder ---
try:
    # preferred: ImageBind installed as "imagebind"
    from imagebind import data
    from imagebind.models import imagebind_model
    from imagebind.models.imagebind_model import ModalityType
except ModuleNotFoundError:
    # fallback: repo has ./ImageBind/imagebind/
    REPO_ROOT = Path(__file__).resolve().parents[1]  # repo/
    sys.path.insert(0, str(REPO_ROOT / "ImageBind"))
    from imagebind import data
    from imagebind.models import imagebind_model
    from imagebind.models.imagebind_model import ModalityType

from models.TaskCLIP import TaskCLIP


def _draw_boxes_pil(
    img: Image.Image,
    boxes_xyxy: np.ndarray,
    color: Tuple[int, int, int],
    width: int = 3,
) -> Image.Image:
    out = img.copy()
    draw = ImageDraw.Draw(out)
    if boxes_xyxy is None or len(boxes_xyxy) == 0:
        return out
    for (x0, y0, x1, y1) in boxes_xyxy.tolist():
        draw.rectangle([x0, y0, x1, y1], outline=color, width=width)
    return out


def _crop_pil(img: Image.Image, bbox_list: List[List[float]]) -> Tuple[List[Image.Image], List[int]]:
    """Return list of cropped PIL images + indices mapping back to bbox_list."""
    W, H = img.size
    crops: List[Image.Image] = []
    idxs: List[int] = []
    for i, (x0, y0, x1, y1) in enumerate(bbox_list):
        x0 = max(0, min(W, int(x0)))
        y0 = max(0, min(H, int(y0)))
        x1 = max(0, min(W, int(x1)))
        y1 = max(0, min(H, int(y1)))
        if x1 <= x0 or y1 <= y0:
            continue
        crops.append(img.crop((x0, y0, x1, y1)))
        idxs.append(i)
    return crops, idxs


def overlay_masks(
    img: Image.Image,
    masks: np.ndarray,
    alpha: float = 0.40,
    color: Tuple[int, int, int] = (255, 0, 0),
) -> Image.Image:
    if masks is None or len(masks) == 0:
        return img

    base = np.array(img).astype(np.float32)
    union = np.any(masks.astype(bool), axis=0)  # (H, W)
    if not np.any(union):
        return img

    overlay = base.copy()
    overlay[union] = overlay[union] * 0.2 + np.array(color, dtype=np.float32) * 0.8
    out = base * (1 - alpha) + overlay * alpha
    return Image.fromarray(np.clip(out, 0, 255).astype(np.uint8))


class ModelRunner:
    """
    WebUI runner:
      - YOLO detects bboxes
      - VLM (ImageBind or OpenCLIP) embeds text prompts and crops (+ global image)
      - TaskCLIP scores and selects bboxes
      - optionally visualize bbox or SAM masks
    """

    def __init__(
        self,
        project_root: str,
        device: str = "cuda:0",
        yolo_ckpt: str = "./.checkpoints/yolo12x.pt",
        sam_ckpt: str = "./.checkpoints/sam2.1_l.pt",
        imagebind_ckpt: Optional[str] = None,  # optional local weights path
        id2task_name_file: str = "./id2task_name.json",
        task2prompt_file: str = "./task20.json",
        threshold: float = 0.01,
        forward: bool = True,
        cluster: bool = True,
        forward_thre: float = 0.1,
    ):
        self.root = Path(project_root).resolve()
        self.device = device
        self.threshold = float(threshold)
        self.forward = bool(forward)
        self.cluster = bool(cluster)
        self.forward_thre = float(forward_thre)

        # load task metadata
        self.id2task_name_path = (self.root / id2task_name_file).resolve()
        self.task2prompt_path = (self.root / task2prompt_file).resolve()
        self.id2task_name = json.loads(self.id2task_name_path.read_text())
        self.task2prompt = json.loads(self.task2prompt_path.read_text())

        # caches
        self._vlm_cache: Dict[str, Dict[str, Any]] = {}
        self._yolo_cache: Dict[str, YOLO] = {}
        self._taskclip_cache: Dict[Tuple[Any, ...], TaskCLIP] = {}

        # default ckpt paths (not required; YOLO is cached per-run ckpt)
        self.yolo_ckpt_path = (self.root / yolo_ckpt).resolve() if str(yolo_ckpt).startswith(".") else Path(yolo_ckpt)

        # SAM loaded once
        sam_ckpt_path = (self.root / sam_ckpt).resolve() if str(sam_ckpt).startswith(".") else Path(sam_ckpt)
        self.sam = SAM(str(sam_ckpt_path))

        # ImageBind weights path (optional)
        self.imagebind_ckpt = imagebind_ckpt

        # lock for single-GPU servers
        self._lock = RLock()

    def _get_yolo(self, ckpt_path: str) -> YOLO:
        ckpt_abs = str((self.root / ckpt_path).resolve()) if str(ckpt_path).startswith(".") else str(ckpt_path)
        if ckpt_abs not in self._yolo_cache:
            self._yolo_cache[ckpt_abs] = YOLO(ckpt_abs)
        return self._yolo_cache[ckpt_abs]

    def _load_imagebind(self) -> Any:
        """
        Load ImageBind once and cache it.
        - If self.imagebind_ckpt provided and exists: load pretrained=False then load_state_dict
        - Else: pretrained=True (may download)
        """
        if "imagebind" in self._vlm_cache:
            return self._vlm_cache["imagebind"]["model"]

        if self.imagebind_ckpt:
            ckpt_path = (self.root / self.imagebind_ckpt).resolve() if str(self.imagebind_ckpt).startswith(".") else Path(self.imagebind_ckpt)
            if ckpt_path.exists():
                m = imagebind_model.imagebind_huge(pretrained=False).to(self.device).eval()
                state = torch.load(str(ckpt_path), map_location="cpu")
                # common wrappers
                if isinstance(state, dict) and "model" in state and isinstance(state["model"], dict):
                    state = state["model"]
                if isinstance(state, dict) and "state_dict" in state and isinstance(state["state_dict"], dict):
                    state = state["state_dict"]
                m.load_state_dict(state, strict=False)
                self._vlm_cache["imagebind"] = {"kind": "imagebind", "model": m}
                return m

        m = imagebind_model.imagebind_huge(pretrained=True).to(self.device).eval()
        self._vlm_cache["imagebind"] = {"kind": "imagebind", "model": m}
        return m

    def _get_vlm(self, vlm_model: str) -> Dict[str, Any]:
        if vlm_model in self._vlm_cache and vlm_model != "imagebind":
            return self._vlm_cache[vlm_model]

        if vlm_model == "imagebind":
            m = self._load_imagebind()
            return {"kind": "imagebind", "model": m}

        if vlm_model == "vit-b":
            m, _, preprocess = open_clip.create_model_and_transforms(
                "ViT-B-32", pretrained="laion2b_s34b_b79k"
            )
            m = m.to(self.device).eval()
            tokenizer = open_clip.get_tokenizer("ViT-B-32")
            pack = {"kind": "openclip", "model": m, "preprocess": preprocess, "tokenizer": tokenizer}
            self._vlm_cache[vlm_model] = pack
            return pack

        if vlm_model == "vit-l":
            m, _, preprocess = open_clip.create_model_and_transforms(
                "ViT-L-14", pretrained="laion2b_s32b_b82k"
            )
            m = m.to(self.device).eval()
            tokenizer = open_clip.get_tokenizer("ViT-L-14")
            pack = {"kind": "openclip", "model": m, "preprocess": preprocess, "tokenizer": tokenizer}
            self._vlm_cache[vlm_model] = pack
            return pack

        raise ValueError(f"Unknown vlm_model: {vlm_model}")

    def _encode_vlm(
        self,
        vlm_model: str,
        prompt_use: List[str],
        seg_list: List[Image.Image],
        full_img_pil: Image.Image,
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        pack = self._get_vlm(vlm_model)

        with torch.inference_mode():
            if pack["kind"] == "imagebind":
                m = pack["model"]
                input_pack = {
                    ModalityType.TEXT: data.load_and_transform_text(prompt_use, self.device),
                    ModalityType.VISION: data.read_and_transform_vision_data(seg_list, self.device),
                }
                emb = m(input_pack)
                text_embeddings = emb[ModalityType.TEXT]
                bbox_embeddings = emb[ModalityType.VISION]

                input_pack2 = {ModalityType.VISION: data.read_and_transform_vision_data([full_img_pil], self.device)}
                emb2 = m(input_pack2)
                image_embedding = emb2[ModalityType.VISION].squeeze(0)
                return text_embeddings, bbox_embeddings, image_embedding

            # openclip branch
            m = pack["model"]
            preprocess = pack["preprocess"]
            tokenizer = pack["tokenizer"]

            # text
            text = tokenizer(prompt_use).to(self.device)
            text_embeddings = m.encode_text(text).float()
            text_embeddings = text_embeddings / text_embeddings.norm(dim=-1, keepdim=True)

            # bbox crops
            crop_tensors = [preprocess(im) for im in seg_list]
            crop_batch = torch.stack(crop_tensors, dim=0).to(self.device)
            bbox_embeddings = m.encode_image(crop_batch).float()
            bbox_embeddings = bbox_embeddings / bbox_embeddings.norm(dim=-1, keepdim=True)

            # global image
            img_tensor = preprocess(full_img_pil).unsqueeze(0).to(self.device)
            image_embedding = m.encode_image(img_tensor).float().squeeze(0)
            image_embedding = image_embedding / image_embedding.norm(dim=-1, keepdim=True)

            return text_embeddings, bbox_embeddings, image_embedding

    def list_task_ids(self) -> List[int]:
        ids: List[int] = []
        for k in self.id2task_name.keys():
            try:
                ids.append(int(k))
            except Exception:
                pass
        return sorted(ids)

    @staticmethod
    def _unwrap_state_dict(obj: Any) -> Dict[str, torch.Tensor]:
        # supports {"state_dict": ...} style checkpoints
        if isinstance(obj, dict) and "state_dict" in obj and isinstance(obj["state_dict"], dict):
            return obj["state_dict"]
        if isinstance(obj, dict):
            return obj
        raise TypeError(f"Unsupported checkpoint format: {type(obj)}")

    def _infer_ckpt_flags(self, state: Dict[str, torch.Tensor]) -> Tuple[bool, bool, int]:
        """
        Infer:
          - is_hdc: whether checkpoint contains HDC submodule keys
          - has_cross_attention: whether checkpoint contains cross-attn keys
          - ckpt_d_model: best-effort inferred d_model from weights
        """
        keys = list(state.keys())
        is_hdc = any(k.startswith("ScoreFunction.HDReason.") for k in keys)
        # NOTE: adjust this if your TaskCLIP names cross-attn differently
        has_cross = any("cross_attn" in k or "cross_attn_text" in k for k in keys)

        if "decoder_norm.weight" in state:
            ckpt_d_model = int(state["decoder_norm.weight"].shape[0])
        elif "ScoreFunction.norm.weight" in state:
            ckpt_d_model = int(state["ScoreFunction.norm.weight"].shape[0])
        else:
            ckpt_d_model = -1

        return is_hdc, has_cross, ckpt_d_model

    def _get_taskclip(
        self,
        ckpt_path: str,
        d_model: int,
        n_words: int,
        score_function: str,
        hdv_dim: int,
        cross_attention: bool,
    ) -> TaskCLIP:
        ckpt_abs = str((self.root / ckpt_path).resolve()) if str(ckpt_path).startswith(".") else str(ckpt_path)
        if not Path(ckpt_abs).exists():
            raise FileNotFoundError(f"TaskCLIP checkpoint not found: {ckpt_abs}")

        eff_hdv_dim = int(hdv_dim) if score_function == "HDC" else 0

        # cache key must include cross_attention + score_function + dimensions
        key = (ckpt_abs, int(d_model), int(n_words), str(score_function), int(eff_hdv_dim), bool(cross_attention))
        if key in self._taskclip_cache:
            return self._taskclip_cache[key]

        state_raw = torch.load(ckpt_abs, map_location="cpu")
        state = self._unwrap_state_dict(state_raw)

        ckpt_is_hdc, ckpt_has_cross, ckpt_d_model = self._infer_ckpt_flags(state)

        # Validate score_function against checkpoint
        if score_function == "HDC" and not ckpt_is_hdc:
            raise RuntimeError(f"Checkpoint is NOT HDC but score_function=HDC was selected. ckpt={ckpt_abs}")
        if score_function != "HDC" and ckpt_is_hdc:
            raise RuntimeError(f"Checkpoint IS HDC but score_function=default was selected. ckpt={ckpt_abs}")

        # Validate d_model against checkpoint (if inferred)
        if ckpt_d_model != -1 and int(d_model) != int(ckpt_d_model):
            raise RuntimeError(
                f"d_model mismatch: VLM produced d_model={int(d_model)} but checkpoint expects d_model={int(ckpt_d_model)}. ckpt={ckpt_abs}"
            )

        model_config = {
            "num_layers": 8,
            "norm": None,
            "return_intermediate": False,
            "d_model": int(d_model),
            "nhead": 4,
            "dim_feedforward": 2048,
            "dropout": 0.1,
            "N_words": int(n_words),
            "activation": "gelu",
            "normalize_before": False,
            "device": self.device,
            "ratio_text": 0.3,
            "ratio_image": 0.3,
            "ratio_glob": 0.3,
            "norm_before": True,
            "norm_after": False,
            "MIN_VAL": 10.0,
            "MAX_VAL": 30.0,
            "cross_attention": bool(cross_attention),
            "score_function": "HDC" if score_function == "HDC" else "default",
            "HDV_D": int(eff_hdv_dim),
        }

        m = TaskCLIP(model_config, normalize_before=model_config["normalize_before"], device=model_config["device"])
        m.load_state_dict(state, strict=True)
        m = m.to(self.device).eval()

        self._taskclip_cache[key] = m
        return m

    def _find_same_class(self, predict_res, score, visited, i, classes, confs, forward_thre):
        cls_i = classes[i]
        for j in range(len(score)):
            if visited[j] == 1:
                continue
            if classes[j] == cls_i and float(score[j]) > forward_thre:
                visited[j] = 1
                predict_res[j]["category_id"] = 1
                predict_res[j]["score"] = float(score[j])

    def _sam_masks_from_bboxes(self, image_path: str, bbox_list: List[List[float]], img_h: int, img_w: int) -> np.ndarray:
        if not bbox_list:
            return np.zeros((0, img_h, img_w), dtype=bool)

        bboxes = [[float(x0), float(y0), float(x1), float(y1)] for x0, y0, x1, y1 in bbox_list]

        try:
            res = self.sam(image_path, bboxes=bboxes)
            r0 = res[0]
            if r0.masks is None:
                return np.zeros((0, img_h, img_w), dtype=bool)
            masks = r0.masks.data.detach().cpu().numpy().astype(bool)
            return masks
        except Exception:
            # fallback per-box
            masks_list = []
            for bb in bboxes:
                rr = self.sam(image_path, bboxes=bb)[0]
                if rr.masks is None:
                    continue
                m = rr.masks.data.detach().cpu().numpy().astype(bool)
                masks_list.append(m[0])
            if len(masks_list) == 0:
                return np.zeros((0, img_h, img_w), dtype=bool)
            return np.stack(masks_list, axis=0)

    def run(
        self,
        image_path: str,
        task_id: int,
        vlm_model: str = "imagebind",
        od_model: str = "yolo",
        yolo_ckpt: str = "./.checkpoints/yolo12x.pt",
        score_function: str = "default",
        hdv_dim: int = 256,
        taskclip_ckpt: str = "./test_model/default/decoder.pt",
        viz_mode: str = "bbox",
        hw_noise_dist: str = "none",
        hw_noise_width: int = 0,
        hw_noise_strength: int = 0,
        hdc_bits: int = 32
    ) -> Dict[str, Any]:
        if vlm_model not in ["imagebind", "vit-b", "vit-l"]:
            raise ValueError(f"Unknown vlm_model: {vlm_model}")
        if od_model != "yolo":
            raise ValueError("Currently only od_model='yolo' is supported.")
        if viz_mode not in ["bbox", "mask"]:
            raise ValueError(f"Unknown viz_mode={viz_mode}")

        # Training convention you stated:
        # - default => cross_attention True
        # - HDC     => cross_attention False
        # If your actual training differs, change this rule OR pass it from app.py.
        cross_attention = (score_function != "HDC")

        with self._lock:
            img = Image.open(image_path).convert("RGB")

            task_name = self.id2task_name[str(task_id)]
            prompt_words = self.task2prompt[task_name]
            prompt_use = ["The item is " + w for w in prompt_words]

            # YOLO detect
            yolo = self._get_yolo(yolo_ckpt)
            outputs = yolo(image_path)
            bbox_list = outputs[0].boxes.xyxy.tolist()
            classes = outputs[0].boxes.cls.tolist()
            confidences = outputs[0].boxes.conf.tolist()

            H, W = img.size[1], img.size[0]
            all_boxes = np.asarray(bbox_list, dtype=np.float32)

            # visualize all detections
            if viz_mode == "bbox":
                img_yolo = _draw_boxes_pil(img, all_boxes, color=(0, 255, 0), width=3)
                all_masks = None
            else:
                all_masks = self._sam_masks_from_bboxes(image_path, bbox_list, img_h=H, img_w=W)
                img_yolo = overlay_masks(img, all_masks, alpha=0.35, color=(0, 255, 0))

            # crop bboxes
            seg_list, _ = _crop_pil(img, bbox_list)
            if len(seg_list) == 0:
                return {
                    "task_id": task_id,
                    "task_name": task_name,
                    "bbox_list": bbox_list,
                    "selected_indices": [],
                    "images": {"original": img, "yolo": img_yolo, "selected": img.copy()},
                }

            # VLM embeddings
            text_embeddings, bbox_embeddings, image_embedding = self._encode_vlm(
                vlm_model=vlm_model,
                prompt_use=prompt_use,
                seg_list=seg_list,
                full_img_pil=img,
            )

            # Ensure dims are consistent
            if int(bbox_embeddings.shape[-1]) != int(image_embedding.shape[-1]):
                raise RuntimeError(
                    f"Embedding dim mismatch: bbox_embeddings dim={bbox_embeddings.shape[-1]} "
                    f"vs image_embedding dim={image_embedding.shape[-1]}"
                )

            # IMPORTANT: d_model should come from bbox_embeddings (tgt), not global image
            d_model = int(bbox_embeddings.shape[-1])
            n_words = int(text_embeddings.shape[0])

            # TaskCLIP (load correct arch)
            taskclip = self._get_taskclip(
                ckpt_path=taskclip_ckpt,
                d_model=d_model,
                n_words=n_words,
                score_function=score_function,
                hdv_dim=hdv_dim,
                cross_attention=cross_attention,
            )

            # Score
            with torch.inference_mode():
                tgt = bbox_embeddings
                memory = text_embeddings
                image_embedding_2d = image_embedding.view(1, -1)
                _, _, score_res, _ = taskclip(
                    tgt, 
                    memory, 
                    image_embedding_2d,
                    hw_noise_dist=hw_noise_dist,
                    hw_noise_width=int(hw_noise_width),
                    hw_noise_strength=int(hw_noise_strength),
                    hdc_bits=hdc_bits)
                score = score_res.view(-1).detach().cpu().numpy().tolist()

            # post-process
            predict_res = []
            for i in range(len(bbox_list)):
                predict_res.append({"category_id": -1, "score": -1, "class": int(classes[i])})

            visited = [0] * len(score)
            for i, x in enumerate(score):
                if visited[i] == 1:
                    continue
                if float(x) > self.threshold:
                    visited[i] = 1
                    predict_res[i]["category_id"] = 1
                    predict_res[i]["score"] = float(x)
                    if self.forward:
                        self._find_same_class(predict_res, score, visited, i, classes, confidences, self.forward_thre)
                else:
                    predict_res[i]["category_id"] = 0
                    predict_res[i]["score"] = 1.0 - float(x)

            # cluster optimization
            if self.cluster and self.forward and len(seg_list) > 1:
                cluster_scores: Dict[int, List[float]] = {}
                for p in predict_res:
                    if int(p["category_id"]) == 1:
                        c = int(p["class"])
                        cluster_scores.setdefault(c, []).append(float(p["score"]))

                if len(cluster_scores) > 1:
                    cluster_ave = {c: float(np.mean(v)) for c, v in cluster_scores.items()}
                    select_class = max(cluster_ave, key=lambda k: cluster_ave[k])
                    for p in predict_res:
                        if int(p["category_id"]) == 1 and int(p["class"]) != int(select_class):
                            p["category_id"] = 0

            selected_indices = [i for i, p in enumerate(predict_res) if int(p["category_id"]) == 1]
            selected_boxes = (
                all_boxes[selected_indices] if len(selected_indices) > 0 else np.zeros((0, 4), dtype=np.float32)
            )

            # visualize selected
            if viz_mode == "bbox":
                img_selected = _draw_boxes_pil(img, selected_boxes, color=(255, 0, 0), width=4)
            else:
                if all_masks is not None and all_masks.shape[0] > 0 and len(selected_indices) > 0:
                    sel_masks = all_masks[selected_indices]
                else:
                    sel_masks = np.zeros((0, H, W), dtype=bool)
                img_selected = overlay_masks(img, sel_masks, alpha=0.45, color=(255, 0, 0))

            return {
                "task_id": task_id,
                "task_name": task_name,
                "bbox_list": bbox_list,
                "classes": classes,
                "confidences": confidences,
                "scores": score,
                "selected_indices": selected_indices,
                "images": {"original": img, "yolo": img_yolo, "selected": img_selected},
            }