File size: 11,519 Bytes
f5ab8aa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Offline DePlot (google/deplot) batch pipeline for ChartQA visual_fact_deplot."""
from __future__ import annotations

import json
import os
from typing import Any, Optional

from data_utils.paths import resolve_image_path

DEFAULT_MODEL_ID = "google/deplot"
DEFAULT_PROMPT = "Generate underlying data table of the figure below:"
PLACEHOLDER_SOURCE = "deplot_placeholder"
REAL_SOURCE = "google/deplot"


def _parse_vf(raw: Any) -> Optional[dict[str, Any]]:
    if raw is None:
        return None
    if isinstance(raw, dict):
        return raw
    if isinstance(raw, str):
        text = raw.strip()
        if not text:
            return None
        try:
            data = json.loads(text)
        except json.JSONDecodeError:
            return None
        return data if isinstance(data, dict) else None
    return None


def is_deplot_placeholder(vf: Any) -> bool:
    data = _parse_vf(vf)
    if data is None:
        return False
    return data.get("source") == PLACEHOLDER_SOURCE


def has_real_deplot(vf: Any) -> bool:
    data = _parse_vf(vf)
    if data is None:
        return False
    if data.get("source") == PLACEHOLDER_SOURCE:
        return False
    table = (data.get("parsed_table") or "").strip()
    return bool(table) and data.get("source") in (REAL_SOURCE, "google/deplot", "deplot")


def format_deplot_for_teacher(vf: Any) -> str:
    """Teacher-facing text from visual_fact_deplot; empty if missing/placeholder."""
    data = _parse_vf(vf)
    if data is None:
        return ""
    if data.get("source") == PLACEHOLDER_SOURCE:
        return ""
    table = (data.get("parsed_table") or "").strip()
    if table:
        return table
    return ""


def placeholder_deplot_table(entry: dict[str, Any], error: Optional[str] = None) -> str:
    question = entry.get("question", entry.get("question_wo_prompt", ""))
    payload: dict[str, Any] = {
        "source": PLACEHOLDER_SOURCE,
        "question": question,
        "parsed_table": {"note": "DePlot unavailable or image missing"},
    }
    if error:
        payload["error"] = error
    return json.dumps(payload, ensure_ascii=False)


def build_deplot_visual_fact(
    entry: dict[str, Any],
    parsed_table: str,
    *,
    model_id: str = DEFAULT_MODEL_ID,
) -> str:
    question = entry.get("question", entry.get("question_wo_prompt", ""))
    payload = {
        "source": REAL_SOURCE,
        "model_id": model_id,
        "question": question,
        "parsed_table": parsed_table.strip(),
    }
    return json.dumps(payload, ensure_ascii=False)


def cache_key_for_entry(entry: dict[str, Any]) -> str:
    image = entry.get("image", "")
    return os.path.abspath(resolve_image_path(image)) if image else ""


def load_deplot_cache(path: str) -> dict[str, str]:
    if not path or not os.path.isfile(path):
        return {}
    try:
        with open(path, encoding="utf-8") as f:
            data = json.load(f)
        if isinstance(data, dict):
            return {str(k): str(v) for k, v in data.items()}
    except (json.JSONDecodeError, OSError):
        pass
    return {}


def save_deplot_cache(path: str, cache: dict[str, str]) -> None:
    if not path:
        return
    os.makedirs(os.path.dirname(os.path.abspath(path)), exist_ok=True)
    tmp = f"{path}.tmp"
    with open(tmp, "w", encoding="utf-8") as f:
        json.dump(cache, f, ensure_ascii=False, indent=2)
    os.replace(tmp, path)


def needs_deplot_processing(
    entry: dict[str, Any],
    *,
    replace_placeholder: bool = True,
    only_missing: bool = False,
) -> bool:
    vf = entry.get("visual_fact_deplot")
    if not vf:
        return True
    if is_deplot_placeholder(vf):
        return replace_placeholder or only_missing
    if has_real_deplot(vf):
        return replace_placeholder and not only_missing
    return only_missing or replace_placeholder


class DePlotRunner:
    """Lazy-loaded batched DePlot inference."""

    def __init__(
        self,
        model_id: str = DEFAULT_MODEL_ID,
        device: Optional[str] = None,
        dtype: Optional[str] = None,
        prompt: str = DEFAULT_PROMPT,
        max_new_tokens: int = 384,
    ):
        self.model_id = model_id
        self.prompt = prompt
        self.max_new_tokens = max_new_tokens
        self._device = device
        self._dtype = dtype
        self._processor = None
        self._model = None

    def _resolve_device(self):
        import torch

        if self._device and self._device != "auto":
            return torch.device(self._device)
        if torch.cuda.is_available():
            return torch.device("cuda")
        return torch.device("cpu")

    def _resolve_dtype(self, device):
        import torch

        if self._dtype == "float32":
            return torch.float32
        if self._dtype == "float16":
            return torch.float16
        if self._dtype == "bfloat16":
            return torch.bfloat16
        if device.type == "cuda":
            return torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
        return torch.float32

    def load(self) -> bool:
        if self._model is not None:
            return True
        try:
            import torch
            from transformers import Pix2StructForConditionalGeneration, Pix2StructProcessor

            device = self._resolve_device()
            dtype = self._resolve_dtype(device)
            self._processor = Pix2StructProcessor.from_pretrained(self.model_id)
            self._model = Pix2StructForConditionalGeneration.from_pretrained(
                self.model_id,
                torch_dtype=dtype,
            ).to(device)
            self._model.eval()
            self._device_obj = device
            return True
        except Exception as exc:
            print(f"[DePlot] model load failed: {exc}")
            self._model = None
            return False

    def generate_batch(self, image_paths: list[str]) -> list[str]:
        if not image_paths:
            return []
        if not self.load():
            return [""] * len(image_paths)

        import torch
        from PIL import Image

        images = []
        valid_indices: list[int] = []
        results: list[str] = [""] * len(image_paths)
        for i, path in enumerate(image_paths):
            if not path or not os.path.isfile(path):
                continue
            try:
                images.append(Image.open(path).convert("RGB"))
                valid_indices.append(i)
            except OSError:
                continue

        if not images:
            return results

        device = self._device_obj
        texts = [self.prompt] * len(images)
        with torch.inference_mode():
            inputs = self._processor(images=images, text=texts, return_tensors="pt")
            inputs = {k: v.to(device) for k, v in inputs.items()}
            outputs = self._model.generate(**inputs, max_new_tokens=self.max_new_tokens)
            decoded = self._processor.batch_decode(outputs, skip_special_tokens=True)

        for idx, text in zip(valid_indices, decoded):
            results[idx] = (text or "").strip()
        return results

    def generate_batch_with_oom_retry(
        self,
        image_paths: list[str],
        batch_size: int = 8,
        max_retries: int = 3,
    ) -> list[str]:
        if not image_paths:
            return []
        import torch

        bs = max(1, batch_size)
        out: list[str] = []
        pos = 0
        retries_left = max_retries
        while pos < len(image_paths):
            chunk_paths = image_paths[pos : pos + bs]
            try:
                chunk_out = self.generate_batch(chunk_paths)
                out.extend(chunk_out)
                pos += len(chunk_paths)
                retries_left = max_retries
            except RuntimeError as exc:
                if "out of memory" not in str(exc).lower() or bs <= 1 or retries_left <= 0:
                    out.extend([""] * len(chunk_paths))
                    pos += len(chunk_paths)
                    continue
                if torch.cuda.is_available():
                    torch.cuda.empty_cache()
                bs = max(1, bs // 2)
                retries_left -= 1
        return out


def enrich_entries_with_deplot(
    entries: list[dict[str, Any]],
    *,
    enabled: bool = True,
    model_id: str = DEFAULT_MODEL_ID,
    batch_size: int = 8,
    max_new_tokens: int = 384,
    cache_path: str = "",
    replace_placeholder: bool = True,
    only_missing: bool = False,
    max_samples: int = 0,
    device: Optional[str] = None,
) -> dict[str, int]:
    """
    Fill visual_fact_deplot on entries in-place.
    Returns stats dict: real, placeholder, skipped, failed, cached.
    """
    stats = {"real": 0, "placeholder": 0, "skipped": 0, "failed": 0, "cached": 0}
    work_entries = entries[:max_samples] if max_samples > 0 else entries

    if not enabled:
        for entry in work_entries:
            if not needs_deplot_processing(
                entry, replace_placeholder=replace_placeholder, only_missing=only_missing
            ):
                stats["skipped"] += 1
                continue
            entry["visual_fact_deplot"] = placeholder_deplot_table(entry, error="deplot_disabled")
            stats["placeholder"] += 1
        return stats

    cache = load_deplot_cache(cache_path)
    runner = DePlotRunner(model_id=model_id, device=device, max_new_tokens=max_new_tokens)
    model_ok = runner.load()

    pending: list[tuple[int, str, str]] = []

    for idx, entry in enumerate(work_entries):
        if not needs_deplot_processing(
            entry, replace_placeholder=replace_placeholder, only_missing=only_missing
        ):
            stats["skipped"] += 1
            continue

        key = cache_key_for_entry(entry)
        if key and key in cache and cache[key].strip():
            entry["visual_fact_deplot"] = build_deplot_visual_fact(entry, cache[key], model_id=model_id)
            stats["cached"] += 1
            stats["real"] += 1
            continue

        if not key or not os.path.isfile(key):
            entry["visual_fact_deplot"] = placeholder_deplot_table(entry, error="image_missing")
            stats["placeholder"] += 1
            continue

        if not model_ok:
            entry["visual_fact_deplot"] = placeholder_deplot_table(entry, error="model_load_failed")
            stats["placeholder"] += 1
            continue

        pending.append((idx, key, key))

    if pending and model_ok:
        bs = max(1, batch_size)
        for start in range(0, len(pending), bs):
            chunk = pending[start : start + bs]
            paths = [p[2] for p in chunk]
            tables = runner.generate_batch_with_oom_retry(paths, batch_size=bs)
            for (entry_idx, key, _), table in zip(chunk, tables):
                entry = work_entries[entry_idx]
                if table:
                    entry["visual_fact_deplot"] = build_deplot_visual_fact(
                        entry, table, model_id=model_id
                    )
                    if key:
                        cache[key] = table
                    stats["real"] += 1
                else:
                    entry["visual_fact_deplot"] = placeholder_deplot_table(
                        entry, error="inference_failed"
                    )
                    stats["failed"] += 1
                    stats["placeholder"] += 1
            if cache_path and cache:
                save_deplot_cache(cache_path, cache)

    return stats