File size: 19,907 Bytes
1dda194
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Six multimodal reranker adapters: score and reorder the same top-50 candidate pool."""
import argparse
import json
import math
from pathlib import Path

import utils as data
from evaluate import validate_run


import torch
import torch.nn.functional as F

SYSTEM_PROMPT = (
    'Judge whether the Document meets the requirements based on the Query and '
    'the Instruct provided. Note that the answer can only be "yes" or "no".'
)
DEFAULT_INSTRUCTION = (
    "Given a search query containing a reference product image and the user's "
    "need, retrieve relevant products that can satisfy the user's need."
)


def build_qwen_messages(instruction, q_text, q_img, d_text, d_img, image_kwargs=None):
    img_kw = image_kwargs or {}
    content = [
        {"type": "text", "text": "<Instruct>: " + instruction},
        {"type": "text", "text": "<Query>:"},
    ]
    if q_img:
        content.append({"type": "image", "image": "file://" + q_img, **img_kw})
    if q_text:
        content.append({"type": "text", "text": q_text})
    content.append({"type": "text", "text": "\n<Document>:"})
    if d_img:
        content.append({"type": "image", "image": "file://" + d_img, **img_kw})
    if d_text:
        content.append({"type": "text", "text": d_text})
    return [
        {"role": "system", "content": [{"type": "text", "text": SYSTEM_PROMPT}]},
        {"role": "user", "content": content},
    ]


class BaseReranker:
    def __init__(self, model_path, instruction=DEFAULT_INSTRUCTION, batch_size=4):
        self.model_path = model_path
        self.instruction = instruction
        self.batch_size = batch_size

    def score_pairs(self, pairs):
        """pairs: list of {"query_text","query_image","doc_text","doc_image"}.
        Returns list of float scores, same order."""
        scores = []
        for s in range(0, len(pairs), self.batch_size):
            scores.extend(self._score_batch(pairs[s : s + self.batch_size]))
        return scores

    def _score_batch(self, batch):
        raise NotImplementedError


class _QwenVLYesNo(BaseReranker):
    """Qwen2-VL / Qwen2.5-VL / Qwen3-VL instruct models: P(yes) from the
    last-token yes/no logits (Eq. 3 in the paper)."""

    model_cls_name = None
    image_patch_size = 14
    pixel_factor = 28  # min/max pixels unit: patch_size * merge_size
    per_image_pixels = False  # True -> pass min/max pixels via content dict

    def __init__(self, model_path, instruction=DEFAULT_INSTRUCTION, batch_size=4,
                 max_image_tokens=1280):
        super().__init__(model_path, instruction, batch_size)
        import transformers
        from transformers import AutoProcessor

        cls = getattr(transformers, self.model_cls_name)
        f = self.pixel_factor
        self.min_pixels = 4 * f * f
        self.max_pixels = max_image_tokens * f * f
        self.processor = AutoProcessor.from_pretrained(
            self.model_path,
            padding_side="left",
            min_pixels=self.min_pixels,
            max_pixels=self.max_pixels,
            trust_remote_code=True,
        )
        self.model = cls.from_pretrained(
            self.model_path,
            dtype=torch.bfloat16,
            attn_implementation="sdpa",
            trust_remote_code=True,
        ).cuda().eval()
        vocab = self.processor.tokenizer.get_vocab()
        self.yes_id = vocab["yes"]
        self.no_id = vocab["no"]

    def _image_kwargs(self):
        if self.per_image_pixels:
            return {"min_pixels": self.min_pixels, "max_pixels": self.max_pixels}
        return {}

    @torch.no_grad()
    def _score_batch(self, batch):
        from qwen_vl_utils import process_vision_info

        msgs = [
            build_qwen_messages(
                self.instruction,
                p["query_text"], p["query_image"],
                p["doc_text"], p["doc_image"],
                image_kwargs=self._image_kwargs(),
            )
            for p in batch
        ]
        texts = [
            self.processor.apply_chat_template(m, tokenize=False, add_generation_prompt=True)
            for m in msgs
        ]
        images, _ = process_vision_info(msgs, image_patch_size=self.image_patch_size)
        kwargs = {}
        if self.per_image_pixels:
            # images already resized by process_vision_info (official reranker flow)
            kwargs["do_resize"] = False
        inputs = self.processor(
            text=texts, images=images, padding=True, return_tensors="pt", **kwargs
        )
        inputs = {k: v.cuda() if hasattr(v, "cuda") else v for k, v in inputs.items()}
        logits = self.model(**inputs).logits[:, -1, :]
        pair_logits = logits[:, [self.no_id, self.yes_id]].float()
        return F.softmax(pair_logits, dim=-1)[:, 1].cpu().tolist()


class LycheeRerankMM(_QwenVLYesNo):
    """vec-ai/lychee-rerank-mm (Qwen2.5-VL base), official README flow."""

    model_cls_name = "Qwen2_5_VLForConditionalGeneration"
    image_patch_size = 14
    pixel_factor = 28
    per_image_pixels = False


class Qwen3VLInstruct(_QwenVLYesNo):
    """Qwen/Qwen3-VL-8B-Instruct used as a yes/no reranker."""

    model_cls_name = "Qwen3VLForConditionalGeneration"
    image_patch_size = 16
    pixel_factor = 32
    per_image_pixels = False


class Qwen2VLInstruct(_QwenVLYesNo):
    """Qwen/Qwen2-VL-7B-Instruct used as a yes/no reranker."""

    model_cls_name = "Qwen2VLForConditionalGeneration"
    image_patch_size = 14
    pixel_factor = 28
    per_image_pixels = False


class Qwen3VLReranker(_QwenVLYesNo):
    """Qwen/Qwen3-VL-Reranker-8B, official script flow (Eq. 4 in the paper):
    sigmoid over the last hidden state projected by (w_yes - w_no)."""

    model_cls_name = "Qwen3VLForConditionalGeneration"
    image_patch_size = 16
    pixel_factor = 32
    per_image_pixels = True

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        lm_head = self.model.lm_head.weight.data
        w_diff = (lm_head[self.yes_id] - lm_head[self.no_id]).float()
        self.score_linear = torch.nn.Linear(w_diff.size(0), 1, bias=False)
        with torch.no_grad():
            self.score_linear.weight[0] = w_diff
        self.score_linear = self.score_linear.cuda().to(self.model.dtype)
        self.inner = self.model.model

    @torch.no_grad()
    def _score_batch(self, batch):
        from qwen_vl_utils import process_vision_info

        msgs = [
            build_qwen_messages(
                self.instruction,
                p["query_text"], p["query_image"],
                p["doc_text"], p["doc_image"],
                image_kwargs=self._image_kwargs(),
            )
            for p in batch
        ]
        texts = [
            self.processor.apply_chat_template(m, tokenize=False, add_generation_prompt=True)
            for m in msgs
        ]
        images, _ = process_vision_info(msgs, image_patch_size=self.image_patch_size)
        inputs = self.processor(
            text=texts, images=images, padding=True, return_tensors="pt",
            do_resize=False,
        )
        inputs = {k: v.cuda() if hasattr(v, "cuda") else v for k, v in inputs.items()}
        hidden = self.inner(**inputs).last_hidden_state[:, -1]
        scores = torch.sigmoid(self.score_linear(hidden).float().squeeze(-1))
        return scores.cpu().tolist()


# ---------------- InternVL3.5 ----------------

IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)
INTERNVL_IMAGE_SIZE = 448


def _find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float("inf")
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio


def _dynamic_preprocess(image, min_num=1, max_num=6, image_size=448, use_thumbnail=True):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height
    target_ratios = set(
        (i, j)
        for n in range(min_num, max_num + 1)
        for i in range(1, n + 1)
        for j in range(1, n + 1)
        if i * j <= max_num and i * j >= min_num
    )
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
    target_aspect_ratio = _find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size
    )
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size,
        )
        processed_images.append(resized_img.crop(box))
    if use_thumbnail and len(processed_images) != 1:
        processed_images.append(image.resize((image_size, image_size)))
    return processed_images


class InternVLYesNo(BaseReranker):
    """OpenGVLab/InternVL3_5-8B-Instruct: P(yes) from last-token logits."""

    def __init__(self, model_path, instruction=DEFAULT_INSTRUCTION, batch_size=4,
                 max_num_tiles=6):
        super().__init__(model_path, instruction, batch_size)
        import numpy as np
        from PIL import Image
        from transformers import AutoModel, AutoTokenizer
        self.max_num_tiles = max_num_tiles
        mean = torch.tensor(IMAGENET_MEAN).view(3, 1, 1)
        std = torch.tensor(IMAGENET_STD).view(3, 1, 1)

        def transform(img):
            img = img.convert("RGB").resize(
                (INTERNVL_IMAGE_SIZE, INTERNVL_IMAGE_SIZE), Image.BICUBIC)
            arr = np.asarray(img, dtype=np.float32) / 255.0
            t = torch.from_numpy(arr).permute(2, 0, 1)
            return (t - mean) / std

        self.transform = transform
        self.tokenizer = AutoTokenizer.from_pretrained(
            model_path, trust_remote_code=True)
        self.tokenizer.padding_side = "left"
        if self.tokenizer.pad_token is None:
            self.tokenizer.pad_token = self.tokenizer.eos_token
        self.model = AutoModel.from_pretrained(
            model_path,
            torch_dtype=torch.bfloat16,
            trust_remote_code=True,
            use_flash_attn=False,
        ).cuda().eval()
        self.img_context_token = "<IMG_CONTEXT>"
        self.model.img_context_token_id = self.tokenizer.convert_tokens_to_ids(
            self.img_context_token)
        vocab = self.tokenizer.get_vocab()
        self.yes_id = vocab["yes"]
        self.no_id = vocab["no"]
        self.num_image_token = self.model.num_image_token

    def _load_tiles(self, path):
        from PIL import Image

        image = Image.open(path).convert("RGB")
        tiles = _dynamic_preprocess(
            image, min_num=1, max_num=self.max_num_tiles,
            image_size=INTERNVL_IMAGE_SIZE, use_thumbnail=True,
        )
        return [self.transform(t) for t in tiles]

    def _build_sample(self, pair):
        """Returns (input_ids, pixel_values list) for one query-doc pair."""
        user_parts = ["<Instruct>: " + self.instruction, "<Query>:"]
        n_q = n_d = 0
        q_tiles, d_tiles = [], []
        if pair["query_image"]:
            q_tiles = self._load_tiles(pair["query_image"])
            n_q = len(q_tiles)
            user_parts.append("<image>")
        if pair["query_text"]:
            user_parts.append(pair["query_text"])
        user_parts.append("\n<Document>:")
        if pair["doc_image"]:
            d_tiles = self._load_tiles(pair["doc_image"])
            n_d = len(d_tiles)
            user_parts.append("<image>")
        if pair["doc_text"]:
            user_parts.append(pair["doc_text"])
        user_text = "\n".join(user_parts)

        messages = [
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user", "content": user_text},
        ]
        prompt = self.tokenizer.apply_chat_template(
            messages, tokenize=False, add_generation_prompt=True)
        for n in (n_q, n_d):
            if n:
                image_tokens = (
                    "<img>" + self.img_context_token * (self.num_image_token * n)
                    + "</img>"
                )
                prompt = prompt.replace("<image>", image_tokens, 1)
        assert "<image>" not in prompt
        input_ids = self.tokenizer(prompt, return_tensors="pt").input_ids[0]
        return input_ids, q_tiles + d_tiles

    @torch.no_grad()
    def _score_batch(self, batch):
        ids_list, tiles_list = [], []
        for p in batch:
            ids, tiles = self._build_sample(p)
            ids_list.append(ids)
            tiles_list.append(tiles)
        enc = self.tokenizer.pad(
            {"input_ids": ids_list}, padding=True, return_tensors="pt")
        input_ids = enc.input_ids.cuda()
        attention_mask = enc.attention_mask.cuda()
        all_tiles = [t for tiles in tiles_list for t in tiles]
        if all_tiles:
            pixel_values = torch.stack(all_tiles).to(
                dtype=torch.bfloat16, device="cuda")
            image_flags = torch.ones(
                pixel_values.size(0), 1, dtype=torch.long, device="cuda")
        else:
            pixel_values = torch.zeros(
                (1, 3, INTERNVL_IMAGE_SIZE, INTERNVL_IMAGE_SIZE),
                dtype=torch.bfloat16, device="cuda")
            image_flags = torch.zeros(1, 1, dtype=torch.long, device="cuda")
        logits = self.model(
            pixel_values=pixel_values,
            input_ids=input_ids,
            attention_mask=attention_mask,
            image_flags=image_flags,
        ).logits[:, -1, :]
        pair_logits = logits[:, [self.no_id, self.yes_id]].float()
        return F.softmax(pair_logits, dim=-1)[:, 1].cpu().tolist()


class LlavaYesNo(BaseReranker):
    """llava-hf/llava-1.5-7b-hf: P(yes) from last-token logits.
    Single-image-trained model used with two images; weak-baseline contrast."""

    def __init__(self, model_path, instruction=DEFAULT_INSTRUCTION, batch_size=8,
                 **_ignored):
        super().__init__(model_path, instruction, batch_size)
        from transformers import AutoProcessor, LlavaForConditionalGeneration

        self.processor = AutoProcessor.from_pretrained(
            model_path, padding_side="left")
        self.model = LlavaForConditionalGeneration.from_pretrained(
            model_path, dtype=torch.float16, attn_implementation="sdpa",
        ).cuda().eval()
        vocab = self.processor.tokenizer.get_vocab()
        # after "ASSISTANT:" the next token carries a leading space (▁)
        self.yes_ids = [vocab[t] for t in ("▁yes", "▁Yes", "yes", "Yes") if t in vocab]
        self.no_ids = [vocab[t] for t in ("▁no", "▁No", "no", "No") if t in vocab]

    def _build_messages(self, pair, n_images):
        content = [{"type": "text", "text": "<Instruct>: " + self.instruction},
                   {"type": "text", "text": "<Query>:"}]
        if pair["query_image"]:
            content.append({"type": "image"})
        if pair["query_text"]:
            content.append({"type": "text", "text": pair["query_text"]})
        content.append({"type": "text", "text": "\n<Document>:"})
        if pair["doc_image"]:
            content.append({"type": "image"})
        if pair["doc_text"]:
            content.append({"type": "text", "text": pair["doc_text"]})
        return [
            {"role": "system", "content": [{"type": "text", "text": SYSTEM_PROMPT}]},
            {"role": "user", "content": content},
        ]

    @torch.no_grad()
    def _score_batch(self, batch):
        from PIL import Image

        texts, images = [], []
        for p in batch:
            msgs = self._build_messages(p, 0)
            texts.append(self.processor.apply_chat_template(
                msgs, tokenize=False, add_generation_prompt=True))
            for key in ("query_image", "doc_image"):
                if p[key]:
                    images.append(Image.open(p[key]).convert("RGB"))
        inputs = self.processor(
            text=texts, images=images or None, padding=True, return_tensors="pt")
        inputs = {k: v.cuda() if hasattr(v, "cuda") else v for k, v in inputs.items()}
        logits = self.model(**inputs).logits[:, -1, :].float()
        probs = F.softmax(logits, dim=-1)
        p_yes = probs[:, self.yes_ids].sum(-1)
        p_no = probs[:, self.no_ids].sum(-1)
        return (p_yes / (p_yes + p_no + 1e-12)).cpu().tolist()


RERANKERS = {
    "qwen3_vl_reranker_8b": ("Qwen/Qwen3-VL-Reranker-8B", Qwen3VLReranker),
    "qwen3_vl_8b_instruct": ("Qwen/Qwen3-VL-8B-Instruct", Qwen3VLInstruct),
    "internvl3_5_8b": ("OpenGVLab/InternVL3_5-8B-Instruct", InternVLYesNo),
    "lychee_rerank_mm": ("vec-ai/lychee-rerank-mm", LycheeRerankMM),
    "llava_1_5_7b": ("llava-hf/llava-1.5-7b-hf", LlavaYesNo),
    "qwen2_vl_7b_instruct": ("Qwen/Qwen2-VL-7B-Instruct", Qwen2VLInstruct),
}





def get_reranker(tag, model_path=None, batch_size=4):
    model_id, cls = RERANKERS[tag]
    return cls(model_path or model_id, batch_size=batch_size)


def rerank(baseline, queries, corpus, root, model):
    query_map = {str(row["idx"]): row for row in queries}
    baseline = validate_run(baseline, query_map, corpus)
    if any(len(row["ranked"]) != 50 for row in baseline):
        raise ValueError("Reranking requires 50 candidates per query")
    output = []
    for row in baseline:
        query = query_map[row["qid"]]
        query_image = data.query_image_path(root, query["image_url"])
        pairs = []
        for hit in row["ranked"]:
            product = corpus[hit["did"]]
            pairs.append({"query_text": query["query"], "query_image": query_image,
                          "doc_text": product["title"],
                          "doc_image": data.product_image_path(root, product["pict_url"])})
        scores = model.score_pairs(pairs)
        if len(scores) != len(pairs) or not all(math.isfinite(score) for score in scores):
            raise ValueError(f"Incomplete or invalid scores for {row['qid']}")
        ranked = [{"did": hit["did"], "score": float(score)} for hit, score in zip(row["ranked"], scores)]
        output.append({"qid": row["qid"], "ranked": sorted(ranked, key=lambda hit: -hit["score"])})
    return output


def main():
    p = argparse.ArgumentParser(description=__doc__)
    p.add_argument("--model", required=True, choices=RERANKERS)
    p.add_argument("--model-path", default=None, help="Optional local model weights")
    p.add_argument("--baseline", default=None)
    p.add_argument("--data", default=None)
    p.add_argument("--output", required=True)
    p.add_argument("--batch-size", type=int, default=4)
    args = p.parse_args()
    if args.batch_size <= 0:
        p.error("batch-size must be positive")
    root = data.data_root(args.data)
    path = Path(args.baseline) if args.baseline else root / "baseline/qwen3_vl_embedding_8b_top50.json"
    baseline = json.loads(path.read_text())
    model = get_reranker(args.model, args.model_path, args.batch_size)
    run = rerank(baseline, data.load_queries(root), data.load_corpus(root), root, model)
    out = Path(args.output)
    out.parent.mkdir(parents=True, exist_ok=True)
    out.write_text(json.dumps(run))
    print(f"Saved {len(run)} query rankings to {out}")


if __name__ == "__main__":
    main()