File size: 21,835 Bytes
64eec2b
124c90f
64eec2b
 
6ead530
124c90f
 
6ead530
 
 
 
64eec2b
 
124c90f
64eec2b
 
 
124c90f
64eec2b
 
 
 
124c90f
64eec2b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6ead530
64eec2b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
124c90f
 
 
 
64eec2b
 
124c90f
 
 
 
 
64eec2b
 
 
 
 
 
 
 
 
 
124c90f
64eec2b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
124c90f
64eec2b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6ead530
 
64eec2b
 
6ead530
 
 
64eec2b
 
6ead530
 
 
 
64eec2b
6ead530
 
 
64eec2b
 
6ead530
64eec2b
6ead530
 
64eec2b
6ead530
 
 
64eec2b
6ead530
 
 
64eec2b
6ead530
 
 
 
 
 
 
64eec2b
 
 
6ead530
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
64eec2b
6ead530
 
 
 
 
 
 
64eec2b
 
 
6ead530
 
 
 
 
 
 
 
 
 
 
 
 
64eec2b
6ead530
64eec2b
 
6ead530
 
64eec2b
 
 
 
 
6ead530
 
 
64eec2b
6ead530
 
64eec2b
6ead530
 
 
64eec2b
6ead530
124c90f
6ead530
 
64eec2b
124c90f
58d199f
124c90f
 
 
 
 
 
 
58d199f
124c90f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
58d199f
124c90f
 
 
 
 
 
58d199f
124c90f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
58d199f
124c90f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
64eec2b
 
 
124c90f
64eec2b
 
 
 
 
 
124c90f
 
64eec2b
124c90f
64eec2b
 
 
 
124c90f
64eec2b
124c90f
 
 
 
 
64eec2b
 
 
 
 
 
 
 
 
 
 
 
124c90f
64eec2b
 
 
 
 
 
 
 
 
 
 
 
124c90f
64eec2b
124c90f
64eec2b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
124c90f
64eec2b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
124c90f
 
 
 
64eec2b
 
 
 
 
 
 
124c90f
 
 
 
64eec2b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
58d199f
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
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
"""
DeBERTav2.py
DeBERTaV2 masked language modeling pretraining for polymer SMILES (PSMILES).
"""

from __future__ import annotations

import os
import time
import json
import shutil
import argparse
import warnings
from typing import Optional, List, Tuple

warnings.filterwarnings("ignore")


def set_cuda_visible_devices(gpu: str = "0") -> None:
    """Set CUDA_VISIBLE_DEVICES before importing torch/transformers heavy modules."""
    os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu)


def parse_args() -> argparse.Namespace:
    """CLI arguments for paths and key training/data settings."""
    parser = argparse.ArgumentParser(description="DeBERTaV2 MLM pretraining for polymer pSMILES.")
    parser.add_argument("--gpu", type=str, default="0", help="CUDA_VISIBLE_DEVICES value.")
    parser.add_argument(
        "--csv_file",
        type=str,
        default="/path/to/polymer_structures_unified.csv",
        help="Path to input CSV containing a 'psmiles' column.",
    )
    parser.add_argument("--nrows", type=int, default=5_000_000, help="Number of rows to read from CSV.")
    parser.add_argument(
        "--train_txt",
        type=str,
        default="/path/to/generated_polymer_smiles_5M.txt",
        help="Path to write SentencePiece training text (one SMILES per line).",
    )
    parser.add_argument(
        "--spm_prefix",
        type=str,
        default="/path/to/spm_5M",
        help="SentencePiece model prefix (produces <prefix>.model and <prefix>.vocab).",
    )
    parser.add_argument(
        "--tokenized_dataset_dir",
        type=str,
        default="/path/to/dataset_tokenized_all",
        help="Directory to save/load tokenized HF dataset.",
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default="/path/to/polybert_output_5M",
        help="Trainer output directory (will contain best/).",
    )
    return parser.parse_args()


def load_psmiles_from_csv(csv_file: str, nrows: int) -> List[str]:
    """Load pSMILES strings from CSV."""
    import pandas as pd

    df = pd.read_csv(csv_file, nrows=nrows, engine="python")
    return df["psmiles"].astype(str).tolist()


def train_val_split(psmiles_list: List[str], test_size: float = 0.2, random_state: int = 42):
    """Split pSMILES into train/val lists."""
    from sklearn.model_selection import train_test_split

    return train_test_split(psmiles_list, test_size=test_size, random_state=random_state)


def write_sentencepiece_training_text(train_psmiles: List[str], train_txt: str) -> None:
    """Write one pSMILES per line for SentencePiece training."""
    os.makedirs(os.path.dirname(os.path.abspath(train_txt)), exist_ok=True)
    with open(train_txt, "w", encoding="utf-8") as f:
        for s in train_psmiles:
            f.write(s.strip() + "\n")


def get_special_tokens() -> List[str]:
    """
    Special tokens + element symbols (upper and lower case) used as user-defined symbols
    for SentencePiece.
    """
    elements = [
        "H","He","Li","Be","B","C","N","O","F","Ne","Na","Mg","Al","Si","P","S","Cl","Ar","K","Ca","Sc","Ti","V","Cr","Mn",
        "Fe","Co","Ni","Cu","Zn","Ga","Ge","As","Se","Br","Kr","Rb","Sr","Y","Zr","Nb","Mo","Tc","Ru","Rh","Pd","Ag","Cd",
        "In","Sn","Sb","Te","I","Xe","Cs","Ba","La","Hf","Ta","W","Re","Os","Ir","Pt","Au","Hg","Tl","Pb","Bi","Po","At",
        "Rn","Fr","Ra","Ac","Rf","Db","Sg","Bh","Hs","Mt","Ds","Rg","Cn","Nh","Fl","Mc","Lv","Ts","Og","Ce","Pr","Nd","Pm",
        "Sm","Eu","Gd","Tb","Dy","Ho","Er","Tm","Yb","Lu","Th","Pa","U","Np","Pu","Am","Cm","Bk","Cf","Es","Fm","Md","No","Lr"
    ]
    small_elements = [i.lower() for i in elements]

    special_tokens = [
        "<pad>",
        "<mask>",
        "[*]",
        "(", ")", "=", "@", "#",
        "0", "1", "2", "3", "4", "5", "6", "7", "8", "9",
        "-", "+",
        "/", "\\",
        "%", "[", "]",
    ]
    special_tokens += elements + small_elements
    return special_tokens


def train_sentencepiece_if_needed(train_txt: str, spm_model_prefix: str, vocab_size: int = 265) -> str:
    """
    Train SentencePiece model if <prefix>.model does not exist.
    Returns path to the .model file.
    """
    import sentencepiece as spm

    model_path = spm_model_prefix + ".model"
    os.makedirs(os.path.dirname(os.path.abspath(spm_model_prefix)), exist_ok=True)

    if not os.path.isfile(model_path):
        spm.SentencePieceTrainer.train(
            input=train_txt,
            model_prefix=spm_model_prefix,
            vocab_size=vocab_size,
            input_sentence_size=5_000_000,
            character_coverage=1.0,
            user_defined_symbols=get_special_tokens(),
        )
    return model_path


def build_psmiles_tokenizer(spm_path: str, max_len: int = 128):
    """
    Uses SentencePiece-backed DebertaV2Tokenizer.
    """
    from transformers import DebertaV2Tokenizer

    tok = DebertaV2Tokenizer(vocab_file=spm_path, do_lower_case=False)
    tok.add_special_tokens({"pad_token": "<pad>", "mask_token": "<mask>"})
    # store max_len for convenience (not required by HF)
    tok.model_max_length = max_len
    return tok


def tokenize_and_save_dataset(train_psmiles: List[str], val_psmiles: List[str], tokenizer, save_dir: str) -> None:
    """Tokenize train/val and persist the DatasetDict to disk."""
    from datasets import Dataset, DatasetDict

    hf_train = Dataset.from_dict({"text": train_psmiles})
    hf_val = Dataset.from_dict({"text": val_psmiles})

    def tokenize_batch(examples):
        return tokenizer(examples["text"], truncation=True, padding="max_length", max_length=tokenizer.model_max_length)

    train_tok = hf_train.map(tokenize_batch, batched=True, batch_size=10_000, num_proc=10)
    val_tok = hf_val.map(tokenize_batch, batched=True, batch_size=10_000, num_proc=10)

    dataset_dict = DatasetDict({"train": train_tok, "test": val_tok})
    os.makedirs(save_dir, exist_ok=True)
    dataset_dict.save_to_disk(save_dir)


def load_tokenized_dataset(tokenized_dir: str):
    """Load tokenized DatasetDict and set torch formats."""
    from datasets import DatasetDict

    dataset_all = DatasetDict.load_from_disk(tokenized_dir)
    dataset_train = dataset_all["train"]
    dataset_test = dataset_all["test"]

    dataset_train.set_format(type="torch", columns=["input_ids", "attention_mask"])
    dataset_test.set_format(type="torch", columns=["input_ids", "attention_mask"])
    return dataset_train, dataset_test


class EpochMetricsCallback:
    """
    TrainerCallback wrapper that:
    - Tracks best validation loss
    - Implements early stopping on val_loss with patience
    - Saves best model + tokenizer.model copy
    - Prints epoch-level stats
    """

    def __init__(self, tokenizer_model_path: str, output_dir: str, patience: int = 10):
        from transformers.trainer_callback import TrainerCallback
        from sentencepiece import SentencePieceProcessor

        class _CB(TrainerCallback):
            def __init__(self, outer):
                super().__init__()
                self.outer = outer

            def on_epoch_end(self, args, state, control, **kwargs):
                self.outer._on_epoch_end(args, state, control, **kwargs)

            def on_evaluate(self, args, state, control, metrics=None, **kwargs):
                self.outer._on_evaluate(args, state, control, metrics=metrics, **kwargs)

            def on_train_end(self, args, state, control, **kwargs):
                self.outer._on_train_end(args, state, control, **kwargs)

        self._cb_cls = _CB
        self._sp = SentencePieceProcessor()
        self._sp.Load(tokenizer_model_path)

        self.tokenizer_model_path = tokenizer_model_path
        self.output_dir = output_dir

        self.best_val_loss = float("inf")
        self.best_epoch = 0
        self.epochs_no_improve = 0
        self.patience = patience

        self.all_epochs = []
        self.best_val_f1 = None
        self.best_val_accuracy = None
        self.best_perplexity = None

        self.trainer_ref = None
        self._last_train_loss = None

    def as_trainer_callback(self):
        return self._cb_cls(self)

    def _save_model(self, trainer_obj, suffix: str) -> None:
        if trainer_obj is None:
            return
        model_dir = os.path.join(self.output_dir, suffix)
        os.makedirs(model_dir, exist_ok=True)
        trainer_obj.model.save_pretrained(model_dir)
        try:
            shutil.copyfile(self.tokenizer_model_path, os.path.join(model_dir, "tokenizer.model"))
        except Exception:
            pass

    def _on_epoch_end(self, args, state, control, **kwargs):
        train_loss = None
        for log in reversed(state.log_history):
            if "loss" in log and float(log.get("loss", 0)) != 0.0:
                train_loss = log["loss"]
                break
        self._last_train_loss = train_loss

    def _on_evaluate(self, args, state, control, metrics=None, **kwargs):
        import numpy as np

        eval_metrics = metrics or {}
        eval_loss = eval_metrics.get("eval_loss")
        eval_f1 = eval_metrics.get("eval_f1")
        eval_accuracy = eval_metrics.get("eval_accuracy", None)

        train_loss = self._last_train_loss

        epoch_data = {
            "epoch": state.epoch,
            "train_loss": train_loss,
            "val_loss": eval_loss,
            "val_f1": eval_f1,
            "val_accuracy": eval_accuracy,
            "perplexity": np.exp(eval_loss) if eval_loss is not None else None,
        }
        self.all_epochs.append(epoch_data)

        if eval_loss is not None and eval_loss < self.best_val_loss - 1e-6:
            self.best_val_loss = eval_loss
            self.best_epoch = state.epoch
            self.epochs_no_improve = 0
            self.best_val_f1 = eval_f1
            self.best_val_accuracy = eval_accuracy
            self.best_perplexity = np.exp(eval_loss) if eval_loss is not None else None
            self._save_model(self.trainer_ref, "best")
        else:
            self.epochs_no_improve += 1

        if self.epochs_no_improve >= self.patience:
            print(f"Early stopping: no improvement in val_loss for {self.patience} epochs.")
            control.should_training_stop = True

        total_params = sum(p.numel() for p in self.trainer_ref.model.parameters()) if self.trainer_ref is not None else 0
        trainable_params = sum(p.numel() for p in self.trainer_ref.model.parameters() if p.requires_grad) if self.trainer_ref is not None else 0

        print(f"\n=== Epoch {int(state.epoch)}/{args.num_train_epochs} ===")
        print(f"Train Loss: {train_loss:.4f}" if train_loss is not None else "Train Loss: None")
        print(f"Validation Loss: {eval_loss:.4f}" if eval_loss is not None else "Validation Loss: None")
        print(f"Validation F1: {eval_f1:.4f}" if eval_f1 is not None else "Validation F1: None")
        if eval_accuracy is not None:
            print(f"Validation Accuracy:{eval_accuracy:.4f}")
        if eval_loss is not None:
            print(f"Perplexity: {np.exp(eval_loss):.2f}")
        print(f"Best Val Loss: {self.best_val_loss:.4f} (epoch {int(self.best_epoch)})")
        print(f"Total Params: {total_params}")
        print(f"Trainable Params: {trainable_params}")
        print(f"No improvement count:{self.epochs_no_improve}/{self.patience}")

    def _on_train_end(self, args, state, control, **kwargs):
        print("\n=== Model saved ===")
        print(f"Best model (epoch {int(self.best_epoch)}, val_loss={self.best_val_loss:.4f}): {os.path.join(self.output_dir, 'best')}/")


def compute_metrics(eval_pred):
    """Metrics for MLM: accuracy + weighted F1 computed only on masked (-100 excluded) positions."""
    import numpy as np
    from sklearn.metrics import f1_score

    logits, labels = eval_pred
    flat_logits = logits.reshape(-1, logits.shape[-1])
    flat_labels = labels.reshape(-1)
    mask = flat_labels != -100

    if mask.sum() == 0:
        return {"eval_f1": 0.0, "eval_accuracy": 0.0}

    masked_logits = flat_logits[mask]
    masked_labels = flat_labels[mask]
    preds = np.argmax(masked_logits, axis=-1)

    f1 = f1_score(masked_labels, preds, average="weighted")
    accuracy = float(np.mean(masked_labels == preds))
    return {"eval_f1": f1, "eval_accuracy": accuracy}


# =============================================================================
# Encoder wrapper for MLM training
# =============================================================================

class PSMILESDebertaEncoder:
    """
    Dual-use wrapper:
    - For MLM training (HF Trainer):
        forward(input_ids, attention_mask, labels) -> HF outputs (with .loss, .logits)
    - token_logits(...) helper for reconstruction 
    """

    def __init__(
        self,
        model_dir_or_name: Optional[str] = None,
        hidden_size: int = 600,
        num_hidden_layers: int = 12,
        num_attention_heads: int = 12,
        intermediate_size: int = 512,
        vocab_size: Optional[int] = None,
        pad_token_id: int = 0,
        emb_dim: int = 600,
    ):
        import torch
        import torch.nn as nn
        from transformers import DebertaV2Config, DebertaV2ForMaskedLM

        self.torch = torch
        self.nn = nn

        if model_dir_or_name is not None:
            self.model = DebertaV2ForMaskedLM.from_pretrained(model_dir_or_name)
        else:
            if vocab_size is None:
                vocab_size = 265  # fallback; will be resized by caller if tokenizer provided
            config = DebertaV2Config(
                vocab_size=vocab_size,
                hidden_size=hidden_size,
                num_attention_heads=num_attention_heads,
                num_hidden_layers=num_hidden_layers,
                intermediate_size=intermediate_size,
                pad_token_id=pad_token_id,
            )
            self.model = DebertaV2ForMaskedLM(config)

        # Use hidden size from config if available
        hs = int(getattr(self.model.config, "hidden_size", hidden_size))
        self.pool_proj = nn.Linear(hs, emb_dim)
        self._device = None

    # ---- nn.Module-like API ----
    def to(self, device):
        self.model.to(device)
        self.pool_proj.to(device)
        self._device = device
        return self

    def train(self, mode: bool = True):
        self.model.train(mode)
        self.pool_proj.train(mode)
        return self

    def eval(self):
        return self.train(False)

    def parameters(self):
        for p in self.model.parameters():
            yield p
        for p in self.pool_proj.parameters():
            yield p

    def state_dict(self):
        sd = {"model": self.model.state_dict(), "pool_proj": self.pool_proj.state_dict()}
        return sd

    def load_state_dict(self, state_dict, strict: bool = False):
        if isinstance(state_dict, dict) and "model" in state_dict and "pool_proj" in state_dict:
            self.model.load_state_dict(state_dict["model"], strict=strict)
            self.pool_proj.load_state_dict(state_dict["pool_proj"], strict=strict)
        else:
            # allow loading a raw HF state_dict (best-effort)
            try:
                self.model.load_state_dict(state_dict, strict=strict)
            except Exception:
                # ignore if incompatible
                pass
        return self

    def __call__(self, input_ids, attention_mask=None, labels=None):
        return self.forward(input_ids=input_ids, attention_mask=attention_mask, labels=labels)

    # ---- Core helpers ----
    def _pool_hidden(self, last_hidden_state, attention_mask=None):
        """
        Pool token embeddings -> sequence embedding.
        Use attention-masked mean pooling (robust).
        """
        import torch

        if attention_mask is None:
            return last_hidden_state.mean(dim=1)

        mask = attention_mask.to(last_hidden_state.device).unsqueeze(-1).float()
        denom = mask.sum(dim=1).clamp(min=1.0)
        pooled = (last_hidden_state * mask).sum(dim=1) / denom
        return pooled

    def forward(self, input_ids, attention_mask=None, labels=None):
        """
        If labels is provided -> MLM mode: return HF outputs (Trainer compatible).
        Else -> encoder mode: return pooled embedding.
        """
        if labels is not None:
            return self.model(input_ids=input_ids, attention_mask=attention_mask, labels=labels)

        out = self.model.deberta(input_ids=input_ids, attention_mask=attention_mask, return_dict=True)
        last_hidden = out.last_hidden_state
        pooled = self._pool_hidden(last_hidden, attention_mask=attention_mask)
        return self.pool_proj(pooled)

    def token_logits(self, input_ids, attention_mask=None, labels=None):
        """
        - If labels provided: returns loss tensor from HF MLM forward
        - Else: returns token logits (B, L, V)
        """
        outputs = self.model(input_ids=input_ids, attention_mask=attention_mask, labels=labels)
        if labels is not None:
            return outputs.loss
        return outputs.logits


def build_model_and_trainer(tokenizer, dataset_train, dataset_test, spm_model_path: str, output_dir: str):
    """Construct model, training args, callback, and Trainer."""
    import torch
    from transformers import Trainer, TrainingArguments, DataCollatorForLanguageModeling

    data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=True, mlm_probability=0.15)

    vocab_size = len(tokenizer)
    pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else 0

    model = PSMILESDebertaEncoder(
        model_dir_or_name=None,
        vocab_size=vocab_size,
        pad_token_id=pad_token_id,
        hidden_size=600,
        num_attention_heads=12,
        num_hidden_layers=12,
        intermediate_size=512,
        emb_dim=600,
    )
    # resize HF embeddings
    try:
        model.model.resize_token_embeddings(len(tokenizer))
    except Exception:
        pass

    device = "cuda:0" if torch.cuda.is_available() else "cpu"
    model.to(device)

    training_args = TrainingArguments(
        output_dir=output_dir,
        overwrite_output_dir=True,
        num_train_epochs=25,
        per_device_train_batch_size=16,
        per_device_eval_batch_size=8,
        eval_accumulation_steps=1000,
        gradient_accumulation_steps=4,
        eval_strategy="epoch",
        logging_strategy="steps",
        logging_steps=500,
        logging_first_step=True,
        save_strategy="no",
        learning_rate=1e-4,
        weight_decay=0.01,
        fp16=torch.cuda.is_available(),
        report_to=[],
        disable_tqdm=False,
    )

    callback_wrapper = EpochMetricsCallback(tokenizer_model_path=spm_model_path, output_dir=output_dir, patience=10)

    trainer = Trainer(
        model=model,  # wrapper is Trainer-compatible
        args=training_args,
        train_dataset=dataset_train,
        eval_dataset=dataset_test,
        data_collator=data_collator,
        compute_metrics=compute_metrics,
        callbacks=[callback_wrapper.as_trainer_callback()],
    )
    callback_wrapper.trainer_ref = trainer
    return model, trainer, callback_wrapper


def run_training(csv_file: str, nrows: int, train_txt: str, spm_prefix: str, tokenized_dir: str, output_dir: str) -> None:
    """End-to-end: load data, train tokenizer (if needed), tokenize, train model, print final report."""
    psmiles_list = load_psmiles_from_csv(csv_file, nrows=nrows)
    train_psmiles, val_psmiles = train_val_split(psmiles_list, test_size=0.2, random_state=42)

    write_sentencepiece_training_text(train_psmiles, train_txt)
    spm_model_path = train_sentencepiece_if_needed(train_txt, spm_prefix, vocab_size=265)

    tokenizer = build_psmiles_tokenizer(spm_path=spm_model_path, max_len=128)

    tokenize_and_save_dataset(train_psmiles, val_psmiles, tokenizer, tokenized_dir)
    dataset_train, dataset_test = load_tokenized_dataset(tokenized_dir)

    model, trainer, callback = build_model_and_trainer(
        tokenizer=tokenizer,
        dataset_train=dataset_train,
        dataset_test=dataset_test,
        spm_model_path=spm_model_path,
        output_dir=output_dir,
    )

    start_time = time.time()
    train_output = trainer.train()
    total_time = time.time() - start_time

    # Final report
    total_params = sum(p.numel() for p in model.parameters())
    trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
    non_trainable_params = total_params - trainable_params

    print(f"\n=== Final Results ===")
    print(f"Total Training Time (s): {total_time:.2f}")
    print(f"Best Validation Loss: {callback.best_val_loss:.4f}")
    print(f"Best Validation F1: {callback.best_val_f1:.4f}" if callback.best_val_f1 is not None else "Best Validation F1: None")
    print(f"Best Validation Accuracy: {callback.best_val_accuracy:.4f}" if callback.best_val_accuracy is not None else "Best Validation Accuracy: None")
    print(f"Best Perplexity: {callback.best_perplexity:.2f}" if callback.best_perplexity is not None else "Best Perplexity: None")
    print(f"Best Model Epoch: {int(callback.best_epoch)}")
    try:
        print(f"Final Training Loss: {train_output.training_loss:.4f}")
    except Exception:
        pass
    print(f"Total Parameters: {total_params}")
    print(f"Trainable Parameters: {trainable_params}")
    print(f"Non-trainable Parameters: {non_trainable_params}")


def main():
    args = parse_args()
    set_cuda_visible_devices(args.gpu)

    run_training(
        csv_file=args.csv_file,
        nrows=args.nrows,
        train_txt=args.train_txt,
        spm_prefix=args.spm_prefix,
        tokenized_dir=args.tokenized_dataset_dir,
        output_dir=args.output_dir,
    )


if __name__ == "__main__":
    main()