File size: 31,893 Bytes
91f863a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
"""
model.py - ChemBERTa ๊ธฐ๋ฐ˜ Multi-Task ๋ชจ๋ธ ์ •์˜
"""
import os, sys
import pandas as pd
import numpy as np

import torch
import torch.nn as nn
import torch.nn.functional as F
import pytorch_lightning as pl

from transformers.models.auto.modeling_auto import AutoModel
from transformers.models.auto.configuration_auto import AutoConfig
from transformers.optimization import get_linear_schedule_with_warmup

from sklearn.metrics import (
    accuracy_score, precision_recall_fscore_support, roc_auc_score,
    mean_squared_error, mean_absolute_error, r2_score
)
from typing import Dict, List, Optional

# ํ”„๋กœ์ ํŠธ ๋ฃจํŠธ ๊ฒฝ๋กœ ์ถ”๊ฐ€
_current_dir = os.path.dirname(os.path.abspath(__file__))
_project_root = os.path.dirname(os.path.dirname(_current_dir))
if _project_root not in sys.path:
    sys.path.insert(0, _project_root)

from admet_ft._modules.scaler import (reverse_scaling, 
                                      reverse_scaling_power, 
                                      reverse_scaling_minmax,
                                      reverse_scaling_adaptive)


class ChemBERTaMultiTask(nn.Module):
    """ChemBERTa ๊ธฐ๋ฐ˜ Multi-Task Learning ๋ชจ๋ธ"""
    
    def __init__(
        self,
        model_name: Optional[str],
        filter_cols: List[str],
        task_list: List[str],
        task_types: Dict[str, str],
        num_classes: Optional[Dict[str, int]] = None,
        hidden_dim: List[int] = [128, 256, 128, 64],
        data_type: str = 'normal'  # data_type ํŒŒ๋ผ๋ฏธํ„ฐ ์ถ”๊ฐ€
    ):
        """
        Args:
            model_name: ChemBERTa ๋ชจ๋ธ ์ด๋ฆ„
            filter_cols: ๋ฒ”์ฃผํ˜• ํŠน์„ฑ ์ปฌ๋Ÿผ ๋ชฉ๋ก
            task_list: ์˜ˆ์ธกํ•  ํƒœ์Šคํฌ ๋ชฉ๋ก
            task_types: ๊ฐ ํƒœ์Šคํฌ์˜ ์œ ํ˜• (classification ๋˜๋Š” regression / multi_layer_regression)
            num_classes: ๊ฐ ๋ถ„๋ฅ˜ ํƒœ์Šคํฌ์˜ ํด๋ž˜์Šค ์ˆ˜ (dict)
            hidden_dim: ์€๋‹‰์ธต ์ฐจ์› ๋ฆฌ์ŠคํŠธ (์˜ˆ: [128, 256, 128, 64])
            use_attention: ์–ดํ…์…˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜ ์‚ฌ์šฉ ์—ฌ๋ถ€
            data_type: ๋ฐ์ดํ„ฐ ์œ ํ˜• (normal, didb_reduce ๋“ฑ)
        """
        super().__init__()

        self.model_name = model_name
        self.filter_cols = filter_cols
        self.task_list = task_list
        self.task_types = task_types
        self.hidden_dim_list = hidden_dim if isinstance(hidden_dim, list) else [hidden_dim]
        self.final_hidden_dim = self.hidden_dim_list[-1]  # ๋งˆ์ง€๋ง‰ ์ฐจ์›์ด ์ตœ์ข… ์ถœ๋ ฅ ์ฐจ์›
        self.data_type = data_type  # data_type ์ €์žฅ

        # ChemBERTa ๋ชจ๋ธ ์„ค์ •
        self.config = AutoConfig.from_pretrained(model_name)
        self.encoder = AutoModel.from_pretrained(model_name, config=self.config)
        self.encoder_hidden = self.config.hidden_size

        # Task type๋ณ„๋กœ output layers ๊ตฌ์„ฑ ๊ฒฐ์ •
        self.reg_tasks = []
        self.multi_reg_tasks = []
        self.cls_tasks = []

        for task in task_list:
            if task_types[task] == 'regression':
                self.reg_tasks.append(task)
            elif task_types[task] == 'multi_layer_regression':
                self.multi_reg_tasks.append(task)
            elif task_types[task] == 'classification':
                self.cls_tasks.append(task)

        # Task heads ๊ตฌ์„ฑ (๋ชจ๋“  layers๋ฅผ task_heads์— ํ†ตํ•ฉ)
        self.task_heads = nn.ModuleDict()

        # ===== Regression: ๊ณต์œ  layers [128, 256, 128, 64, filter_cols] =====
        if len(self.reg_tasks) > 0:
            merged_layers = nn.ModuleList()
            prev_dim = self.encoder_hidden
            # Hidden layers
            for curr_dim in self.hidden_dim_list:
                merged_layers.append(nn.Linear(prev_dim, curr_dim))
                prev_dim = curr_dim
            # Final output layer
            merged_layers.append(nn.Linear(self.final_hidden_dim, len(self.filter_cols)))

            self.task_heads['merged'] = merged_layers

        # ===== Multi-layer Regression: ๊ฐ task๋ณ„ ๋…๋ฆฝ layers [128, 256, 128, 64, 1] =====
        if len(self.multi_reg_tasks) > 0:
            for task in self.multi_reg_tasks:
                safe_task_name = task.replace('.', '__')
                task_layers = nn.ModuleList()
                prev_dim = self.encoder_hidden
                # Hidden layers
                for curr_dim in self.hidden_dim_list:
                    task_layers.append(nn.Linear(prev_dim, curr_dim))
                    prev_dim = curr_dim
                # Final output layer
                task_layers.append(nn.Linear(self.final_hidden_dim, 1))

                self.task_heads[safe_task_name] = task_layers

        # ===== Classification: ๊ฐ task๋ณ„ ๋…๋ฆฝ layers [128, 256, 128, 64, num_classes] =====
        if len(self.cls_tasks) > 0:
            for task in self.cls_tasks:
                safe_task_name = task.replace('.', '__')
                cls_layers = nn.ModuleList()
                prev_dim = self.encoder_hidden
                # Hidden layers
                for curr_dim in self.hidden_dim_list:
                    cls_layers.append(nn.Linear(prev_dim, curr_dim))
                    prev_dim = curr_dim
                # Final output layer
                cls_layers.append(nn.Linear(
                    self.final_hidden_dim,
                    num_classes[task] if num_classes else 2
                ))

                self.task_heads[safe_task_name] = cls_layers


    def forward(self, input_ids, attention_mask):
        """์ˆœ์ „ํŒŒ"""
        encoder_output = self.encoder(input_ids, attention_mask=attention_mask).pooler_output

        task_outputs = {}

        # ===== Regression: merged layers ํ†ต๊ณผ =====
        if len(self.reg_tasks) > 0:
            x = encoder_output
            for layer in self.task_heads['merged']:
                x = layer(x)
            # ์ตœ์ข… ์ถœ๋ ฅ์„ ๊ฐ task๋ณ„๋กœ ๋ถ„๋ฆฌ
            for i, task in enumerate(self.reg_tasks):
                task_outputs[task] = x[:, i]

        # ===== Multi-layer Regression: ๊ฐ task๋ณ„ ๋…๋ฆฝ layers ํ†ต๊ณผ =====
        if len(self.multi_reg_tasks) > 0:
            for task in self.multi_reg_tasks:
                safe_task_name = task.replace('.', '__')
                x = encoder_output
                # Task๋ณ„ ์ „์ฒด layers ํ†ต๊ณผ
                for layer in self.task_heads[safe_task_name]:
                    x = layer(x)
                task_outputs[task] = x.squeeze(-1)

        # ===== Classification: ๊ฐ task๋ณ„ ๋…๋ฆฝ layers ํ†ต๊ณผ =====
        if len(self.cls_tasks) > 0:
            for task in self.cls_tasks:
                safe_task_name = task.replace('.', '__')
                x = encoder_output
                # Task๋ณ„ ์ „์ฒด layers ํ†ต๊ณผ
                for layer in self.task_heads[safe_task_name]:
                    x = layer(x)
                task_outputs[task] = x

        return task_outputs


class ChemBERTaMultiTaskLightning(pl.LightningModule):
    """PyTorch Lightning ๊ธฐ๋ฐ˜ ChemBERTa Multi-Task Learning ๋ชจ๋ธ"""
    
    def __init__(
        self,
        model_name: Optional[str],
        filter_cols: List[str],
        task_list: List[str],
        task_types: Dict[str, str],
        num_classes: Optional[Dict[str, int]] = None,
        hidden_dim: List[int] = [128, 256, 128, 64],
        learning_rate: float = 2e-5,
        weight_decay: float = 0.01,
        warmup_steps: int = 500,
        task_weights: Optional[Dict[str, float]] = None,
        data_type: str = 'normal',  # data_type ํŒŒ๋ผ๋ฏธํ„ฐ ์ถ”๊ฐ€
        scaling: bool = True,
        scaler_path: Optional[str] = None,
        scaler_type: str = 'power',
        loss_type: str = 'mse'  # mse or huber
    ):
        """
        Args:
            model_name: ChemBERTa ๋ชจ๋ธ ์ด๋ฆ„
            filter_cols: ๋ฒ”์ฃผํ˜• ํŠน์„ฑ ์ปฌ๋Ÿผ ๋ชฉ๋ก
            task_list: ์˜ˆ์ธกํ•  ํƒœ์Šคํฌ ๋ชฉ๋ก
            task_types: ๊ฐ ํƒœ์Šคํฌ์˜ ์œ ํ˜• (classification ๋˜๋Š” regression)
            num_classes: ๊ฐ ๋ถ„๋ฅ˜ ํƒœ์Šคํฌ์˜ ํด๋ž˜์Šค ์ˆ˜ (dict)
            hidden_dim: ์€๋‹‰์ธต ์ฐจ์› ๋ฆฌ์ŠคํŠธ (์˜ˆ: [128, 256, 128, 64])
            learning_rate: ํ•™์Šต๋ฅ 
            weight_decay: ๊ฐ€์ค‘์น˜ ๊ฐ์‡ 
            warmup_steps: ์›Œ๋ฐ์—… ์Šคํ… ์ˆ˜
            task_weights: ๊ฐ ํƒœ์Šคํฌ์˜ ์†์‹ค ๊ฐ€์ค‘์น˜ (dict)
            data_type: ๋ฐ์ดํ„ฐ ์œ ํ˜• (normal, didb_reduce ๋“ฑ)
        """
        super().__init__()
        
        # PyTorch Lightning ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์ €์žฅ
        self.save_hyperparameters()
        
        self.model = ChemBERTaMultiTask(
            model_name=model_name,
            filter_cols=filter_cols,
            task_list=task_list,
            task_types=task_types,
            num_classes=num_classes,
            hidden_dim=hidden_dim,
            data_type=data_type  # data_type ์ „๋‹ฌ
        )
        
        self.task_list = task_list
        self.task_types = task_types
        self.task_weights = task_weights or {task: 1.0 for task in task_list}
        self.learning_rate = learning_rate
        self.weight_decay = weight_decay
        self.warmup_steps = warmup_steps
        
        self.filter_cols = filter_cols
        self.scaling = scaling
        self.scaler_path = scaler_path
        self.scaler_type = scaler_type
        self.loss_type = loss_type

        self.validation_step_outputs = []
        self.test_step_outputs = []
        self.test_results_cache = None  # ํ…Œ์ŠคํŠธ ๊ฒฐ๊ณผ ์ž„์‹œ ์ €์žฅ์šฉ

        # ํ‰๊ฐ€ ์ง€ํ‘œ
        self.metrics = {}
        for task in task_list:
            if task_types[task] == 'classification':
                self.metrics[task] = {
                    'accuracy': accuracy_score,
                    'precision_recall_f1': lambda y_true, y_pred: precision_recall_fscore_support(
                        y_true, y_pred, average='binary', zero_division='0')[:3],
                    'auc': lambda y_true, y_pred_proba: roc_auc_score(
                        y_true, y_pred_proba[:, 1]) if len(np.unique(y_true)) > 1 else 0.5
                }
            else:
                # Regression metrics: loss_type์— ๋”ฐ๋ผ ์ฃผ ์†์‹ค ํ•จ์ˆ˜๊ฐ€ ๋‹ฌ๋ผ์ง€์ง€๋งŒ,
                # ํ‰๊ฐ€๋Š” MSE, MAE, R2๋ฅผ ๋ชจ๋‘ ๊ณ„์‚ฐ
                self.metrics[task] = {
                    'mse': mean_squared_error,
                    'mae': mean_absolute_error,
                    'r2': r2_score
                }

    def _reverse_scale(self, values, column: str) -> np.ndarray:
        """
        ์Šค์ผ€์ผ๋ง๋œ ์˜ˆ์ธก๊ฐ’์„ ์›๋ณธ ์Šค์ผ€์ผ๋กœ ์—ญ๋ณ€ํ™˜

        Args:
            values: ์Šค์ผ€์ผ๋ง๋œ ์˜ˆ์ธก๊ฐ’ (Tensor ๋˜๋Š” ndarray)
            column: ์ปฌ๋Ÿผ ์ด๋ฆ„

        Returns:
            ์›๋ณธ ์Šค์ผ€์ผ๋กœ ์—ญ๋ณ€ํ™˜๋œ numpy array
        """
        if isinstance(values, torch.Tensor):
            values_np = values.detach().cpu().numpy()
        else:
            values_np = np.asarray(values)

        if not self.scaling or not self.scaler_path:
            return values_np.astype(np.float32)

        if os.path.exists(self.scaler_path):
            data = {col: np.full_like(values_np, np.nan, dtype=np.float32) for col in self.filter_cols}
            data[column] = values_np
            df = pd.DataFrame(data)

            # Scaler ํƒ€์ž…์— ๋”ฐ๋ผ ์ ์ ˆํ•œ ์—ญ๋ณ€ํ™˜ ํ•จ์ˆ˜ ์„ ํƒ
            if self.scaler_type == 'adapt':
                # Adaptive scaler: feature-specific reverse scaling
                reverted_values = reverse_scaling_adaptive(values_np, column, self.scaler_path)
                return reverted_values.astype(np.float32)
            elif self.scaler_type == 'power':
                reverted = reverse_scaling_power(df, self.scaler_path)
            elif self.scaler_type == 'minmax':
                reverted = reverse_scaling_minmax(df, self.scaler_path)
            else:  # zscore (default)
                reverted = reverse_scaling(df, self.scaler_path)

            return reverted[column].to_numpy(dtype=np.float32)

        return values_np.astype(np.float32)

    def _compute_regression_loss(self, preds: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
        """
        Regression loss ๊ณ„์‚ฐ (MSE ๋˜๋Š” Huber)

        Args:
            preds: ์˜ˆ์ธก๊ฐ’ (์›๋ณธ ์Šค์ผ€์ผ)
            labels: ์‹ค์ œ๊ฐ’ (์›๋ณธ ์Šค์ผ€์ผ)

        Returns:
            Loss ๊ฐ’
        """
        if self.loss_type == 'huber':
            return F.huber_loss(preds, labels, delta=1.0)
        else:  # mse (default)
            return F.mse_loss(preds, labels)

    def forward(self, input_ids, attention_mask):
        """๋ชจ๋ธ ์ˆœ์ „ํŒŒ"""
        return self.model(input_ids, attention_mask)
    
    def training_step(self, batch, batch_idx):
        """ํ•™์Šต ๋‹จ๊ณ„ - Scaled ๊ณต๊ฐ„์—์„œ loss ๊ณ„์‚ฐ"""
        labels = batch['labels']
        input_ids, attention_mask = batch['input_ids'], batch['attention_mask']
        task_outputs = self(input_ids, attention_mask)

        losses = {}
        for i, task in enumerate(self.task_list):
            task_labels = labels[:, i]
            mask = ~torch.isnan(task_labels)
            if mask.sum() > 0:
                valid_labels = task_labels[mask]
                valid_outputs = task_outputs[task][mask]
                if self.task_types[task] == 'classification':
                    task_loss = F.cross_entropy(valid_outputs, valid_labels.long())
                else:
                    # Regression: Scaled ๊ณต๊ฐ„์—์„œ loss ๊ณ„์‚ฐ (MSE ๋˜๋Š” Huber)
                    task_loss = self._compute_regression_loss(valid_outputs, valid_labels.float())
                losses[task] = task_loss

        if len(losses) > 0:
            total_loss = sum(losses.values())
            # PyTorch Lightning ๋กœ๊น…
            self.log('train_loss', total_loss, sync_dist=True, on_step=True, on_epoch=True, prog_bar=True)
            for task, loss in losses.items():
                self.log(f'train_{task}_loss', loss, sync_dist=True, on_step=True, on_epoch=True)
            return total_loss
        else:
            # ๋ชจ๋“  ํƒœ์Šคํฌ๊ฐ€ ๊ฒฐ์ธก์ด๋ฉด None ๋ฐ˜ํ™˜(์ด step์— ๋Œ€ํ•ด ํ•™์ŠตX)
            return None
    
    def validation_step(self, batch, batch_idx):
        """๊ฒ€์ฆ ๋‹จ๊ณ„ - Scaled ๊ณต๊ฐ„์—์„œ loss ๊ณ„์‚ฐ"""
        labels = batch['labels']
        input_ids, attention_mask = batch['input_ids'], batch['attention_mask']
        task_outputs = self(input_ids, attention_mask)

        losses = {}
        predictions = {}

        for i, task in enumerate(self.task_list):
            task_labels = labels[:, i]
            mask = ~torch.isnan(task_labels)
            if mask.sum() > 0:
                valid_labels = task_labels[mask]
                valid_outputs = task_outputs[task][mask]
                if self.task_types[task] == 'classification':
                    task_loss = F.cross_entropy(valid_outputs, valid_labels.long())
                    task_preds = torch.argmax(valid_outputs, dim=1)
                    task_probs = F.softmax(valid_outputs, dim=1)
                    predictions[task] = {
                        'labels': valid_labels.detach().cpu(),
                        'preds': task_preds.detach().cpu(),
                        'probs': task_probs.detach().cpu()
                    }
                else:
                    # Regression: Scaled ๊ณต๊ฐ„์—์„œ loss ๊ณ„์‚ฐ (MSE ๋˜๋Š” Huber)
                    task_loss = self._compute_regression_loss(valid_outputs, valid_labels.float())
                    predictions[task] = {
                        'labels': valid_labels.detach().cpu(),
                        'preds': valid_outputs.detach().cpu()
                    }
                losses[task] = task_loss
            else:
                continue

        if len(losses) > 0:
            total_loss = sum(losses.values())
            # PyTorch Lightning ๋กœ๊น…
            self.log('val_loss', total_loss, prog_bar=True, sync_dist=True, on_step=True, on_epoch=True)
            for task, loss in losses.items():
                self.log(f'val_{task}_loss', loss, sync_dist=True, on_step=True, on_epoch=True)
        else:
            total_loss = None

        output = {'val_loss': total_loss, 'predictions': predictions}
        self.validation_step_outputs.append(output)
        return output
    
    def on_validation_epoch_end(self):
        """๊ฒ€์ฆ ์—ํฌํฌ ์ข…๋ฃŒ ์‹œ ์ฒ˜๋ฆฌ"""
        # ๋ชจ๋“  ๋ฐฐ์น˜์˜ ์˜ˆ์ธก ๊ฒฐ๊ณผ ์ˆ˜์ง‘
        all_predictions = {task: {'labels': [], 'preds': [], 'probs': []} 
                        for task in self.task_list if self.task_types[task] == 'classification'}
        
        all_predictions.update({task: {'labels': [], 'preds': []} 
                            for task in self.task_list if self.task_types[task] == 'regression' or self.task_types[task] == 'multi_layer_regression'})
        
        for output in self.validation_step_outputs:
            for task in self.task_list:
                if task not in output['predictions']:
                    continue  # ํ•ด๋‹น ๋ฐฐ์น˜์— ์˜ˆ์ธก๊ฐ’์ด ์—†์œผ๋ฉด ๊ฑด๋„ˆ๋œ€
                task_preds = output['predictions'][task]
                
                if self.task_types[task] == 'classification':
                    all_predictions[task]['labels'].extend(task_preds['labels'])
                    all_predictions[task]['preds'].extend(task_preds['preds'])
                    all_predictions[task]['probs'].extend(task_preds['probs'])
                else:
                    all_predictions[task]['labels'].extend(task_preds['labels'])
                    all_predictions[task]['preds'].extend(task_preds['preds'])
        
        # ๊ฐ ํƒœ์Šคํฌ๋ณ„ ํ‰๊ฐ€ ์ง€ํ‘œ ๊ณ„์‚ฐ
        for task in self.task_list:
            task_preds = all_predictions[task]
            # ๋น„์–ด์žˆ๊ฑฐ๋‚˜, ํ…์„œ๊ฐ€ ์•„๋‹Œ ๊ฐ’์ด ๋“ค์–ด์žˆ์œผ๋ฉด metric ๊ณ„์‚ฐ ๊ฑด๋„ˆ๋œ€
            if not task_preds['labels'] or not task_preds['preds']:
                continue
            if not all(isinstance(x, torch.Tensor) for x in task_preds['labels']):
                continue
            if not all(isinstance(x, torch.Tensor) for x in task_preds['preds']):
                continue
            
            if self.task_types[task] == 'classification':
                # ๋ถ„๋ฅ˜ ํƒœ์Šคํฌ ํ‰๊ฐ€
                labels_np = torch.stack(task_preds['labels']).numpy()
                preds_np = torch.stack(task_preds['preds']).numpy()
                
                acc = self.metrics[task]['accuracy'](labels_np, preds_np)
                prec, rec, f1 = self.metrics[task]['precision_recall_f1'](labels_np, preds_np)
                
                # ๋กœ๊น…
                self.log(f'val_{task}_acc', acc, sync_dist=True)
                self.log(f'val_{task}_precision', prec, sync_dist=True)
                self.log(f'val_{task}_recall', rec, sync_dist=True)
                self.log(f'val_{task}_f1', f1, sync_dist=True)
                
                # AUC ๊ณ„์‚ฐ (๊ฐ€๋Šฅํ•œ ๊ฒฝ์šฐ)
                try:
                    probs = torch.stack(task_preds['probs']).numpy()
                    if probs.shape[1] >= 2:  # ์ด์ง„ ๋ถ„๋ฅ˜ ์ด์ƒ์ธ ๊ฒฝ์šฐ
                        auc = self.metrics[task]['auc'](labels_np, probs)
                        self.log(f'val_{task}_auc', auc, sync_dist=True)
                except:
                    pass
            else:
                # ํšŒ๊ท€ ํƒœ์Šคํฌ ํ‰๊ฐ€
                labels_np = torch.stack(task_preds['labels']).numpy()
                preds_np = torch.stack(task_preds['preds']).numpy()
                # === NaN ๋งˆ์Šคํ‚น ์ถ”๊ฐ€ ===
                valid_mask = (~np.isnan(labels_np)) & (~np.isnan(preds_np))
                labels_np_valid = labels_np[valid_mask]
                preds_np_valid = preds_np[valid_mask]

                if len(preds_np_valid) < 2:  # ์ƒ˜ํ”Œ์ด ๋ถ€์กฑํ•œ ๊ฒฝ์šฐ R2, MSE ๊ณ„์‚ฐ ์ƒ๋žต
                    continue
                mse = self.metrics[task]['mse'](labels_np_valid, preds_np_valid)
                mae = self.metrics[task]['mae'](labels_np_valid, preds_np_valid)
                r2 = self.metrics[task]['r2'](labels_np_valid, preds_np_valid)
                # ๋กœ๊น…
                self.log(f'val_{task}_mse', mse, sync_dist=True)
                self.log(f'val_{task}_mae', mae, sync_dist=True)
                self.log(f'val_{task}_r2', r2, sync_dist=True)
        
        # ๋‹จ๊ณ„๊ฐ€ ๋๋‚˜๋ฉด ์ถœ๋ ฅ ๋ชฉ๋ก ์ดˆ๊ธฐํ™”
        self.validation_step_outputs.clear()
    
    def test_step(self, batch, batch_idx):
        labels = batch['labels']
        input_ids, attention_mask = batch['input_ids'], batch['attention_mask']
        task_outputs = self(input_ids, attention_mask)

        batch_size = labels.shape[0]
        losses = {}
        predictions = {}

        # SMILES ์ •๋ณด ์ถ”์ถœ
        batch_smiles = batch.get('smiles', None)

        for i, task in enumerate(self.task_list):
            task_labels = labels[:, i]
            mask = ~torch.isnan(task_labels)

            # ์ „์ฒด ๋ฐฐ์น˜ ํฌ๊ธฐ ์œ ์ง€ํ•˜๋ฉด์„œ NaN์œผ๋กœ ์ดˆ๊ธฐํ™”
            full_preds = torch.full((batch_size,), float('nan'), dtype=torch.float32)
            full_labels = task_labels.detach().cpu().float()

            if mask.sum() > 0:
                valid_labels = task_labels[mask]
                valid_outputs = task_outputs[task][mask]
                if self.task_types[task] == 'classification':
                    task_loss = F.cross_entropy(valid_outputs, valid_labels.long())
                    task_preds = torch.argmax(valid_outputs, dim=1).float()
                    task_probs = F.softmax(valid_outputs, dim=1)

                    # ์œ ํšจํ•œ ์œ„์น˜์—๋งŒ ์˜ˆ์ธก๊ฐ’ ์ฑ„์šฐ๊ธฐ
                    full_preds[mask] = task_preds.detach().cpu()

                    predictions[task] = {
                        'labels': full_labels,
                        'preds': full_preds,
                        'probs': task_probs.detach().cpu()
                    }
                else:
                    # Regression: ์—ญ๋ณ€ํ™˜ ํ›„ ์›๋ณธ ์Šค์ผ€์ผ์—์„œ loss ๊ณ„์‚ฐ
                    preds_orig_np = self._reverse_scale(valid_outputs, task)
                    preds_orig = torch.from_numpy(preds_orig_np).float()
                    labels_orig = valid_labels.detach().cpu().float()  # Test labels are already in original scale

                    # ์›๋ณธ ์Šค์ผ€์ผ์—์„œ loss ๊ณ„์‚ฐ (MSE ๋˜๋Š” Huber)
                    task_loss = self._compute_regression_loss(
                        preds_orig.to(valid_outputs.device),
                        labels_orig.to(valid_outputs.device)
                    )

                    # ์œ ํšจํ•œ ์œ„์น˜์—๋งŒ ์˜ˆ์ธก๊ฐ’ ์ฑ„์šฐ๊ธฐ
                    full_preds[mask] = preds_orig.cpu()

                    predictions[task] = {
                        'labels': full_labels,
                        'preds': full_preds
                    }
                losses[task] = task_loss
            else:
                # ๋ชจ๋“  ๊ฐ’์ด NaN์ธ ๊ฒฝ์šฐ์—๋„ ๊ตฌ์กฐ ์œ ์ง€
                predictions[task] = {
                    'labels': full_labels,
                    'preds': full_preds
                }

        if len(losses) > 0:
            total_loss = sum(losses.values())
            # PyTorch Lightning ๋กœ๊น… (WandB ํ˜ธํ™˜)
            self.log('test_loss', total_loss, sync_dist=True, on_step=True, on_epoch=True, prog_bar=True)
            for task, loss in losses.items():
                self.log(f'test_{task}_loss', loss, sync_dist=True, on_step=True, on_epoch=True)
        else:
            total_loss = None

        output = {'test_loss': total_loss, 'predictions': predictions, 'smiles': batch_smiles}  # SMILES ์ถ”๊ฐ€
        self.test_step_outputs.append(output)
        return output

    def test_step_end(self, outputs):
        """ํ…Œ์ŠคํŠธ ๋‹จ๊ณ„ ์ข…๋ฃŒ ์‹œ ์ฒ˜๋ฆฌ (WandB ํ˜ธํ™˜)"""
        # step๋ณ„ ๋ฉ”ํŠธ๋ฆญ ๊ณ„์‚ฐ ๋ฐ ๋กœ๊น…
        if outputs['test_loss'] is not None:
            # step๋ณ„ loss ๋กœ๊น…
            self.log('test_step_loss', outputs['test_loss'], on_step=True, on_epoch=False)
            
            # ๊ฐ task๋ณ„ step๋ณ„ loss ๋กœ๊น…
            for task in self.task_list:
                if task in outputs['predictions']:
                    task_preds = outputs['predictions'][task]
                    if self.task_types[task] == 'classification':
                        # ๋ถ„๋ฅ˜ ๋ฉ”ํŠธ๋ฆญ ๊ณ„์‚ฐ
                        labels = task_preds['labels']
                        preds = task_preds['preds']
                        if len(labels) > 0 and len(preds) > 0:
                            acc = (labels == preds).float().mean()
                            self.log(f'test_{task}_step_acc', acc, on_step=True, on_epoch=False)
                    else:
                        # ํšŒ๊ท€ ๋ฉ”ํŠธ๋ฆญ ๊ณ„์‚ฐ
                        labels = task_preds['labels']
                        preds = task_preds['preds']
                        if len(labels) > 0 and len(preds) > 0:
                            labels_np = labels.detach().cpu().numpy()
                            preds_np = preds.detach().cpu().numpy()
                            valid_mask = (~np.isnan(labels_np)) & (~np.isnan(preds_np))
                            labels_np = labels_np[valid_mask]
                            preds_np = preds_np[valid_mask]
                            if len(preds_np) == 0:
                                continue
                            mse_val = mean_squared_error(labels_np, preds_np)
                            mse_tensor = torch.tensor(mse_val, device=labels.device, dtype=torch.float32)
                            self.log(f'test_{task}_step_mse', mse_tensor, on_step=True, on_epoch=False)
        
        return outputs
    
    def on_test_epoch_end(self):
        """ํ…Œ์ŠคํŠธ ์—ํฌํฌ ์ข…๋ฃŒ ์‹œ ์ฒ˜๋ฆฌ (๊ฒ€์ฆ๊ณผ ์œ ์‚ฌ)"""
        # ๋ชจ๋“  ๋ฐฐ์น˜์˜ ์˜ˆ์ธก ๊ฒฐ๊ณผ ์ˆ˜์ง‘
        all_predictions = {task: {'labels': [], 'preds': [], 'probs': []}
                        for task in self.task_list if self.task_types[task] == 'classification'}

        all_predictions.update({task: {'labels': [], 'preds': []}
                            for task in self.task_list if self.task_types[task] == 'regression' or self.task_types[task] == 'multi_layer_regression'})

        # SMILES ์ˆ˜์ง‘์šฉ ๋ฆฌ์ŠคํŠธ ์ถ”๊ฐ€
        all_smiles = []

        for output in self.test_step_outputs:
            # SMILES ์ˆ˜์ง‘
            if 'smiles' in output and output['smiles'] is not None:
                all_smiles.extend(output['smiles'])

            for task in self.task_list:
                if task not in output['predictions']:
                    continue  # ํ•ด๋‹น ๋ฐฐ์น˜์— ์˜ˆ์ธก๊ฐ’์ด ์—†์œผ๋ฉด ๊ฑด๋„ˆ๋œ€
                task_preds = output['predictions'][task]

                if self.task_types[task] == 'classification':
                    all_predictions[task]['labels'].extend(task_preds['labels'])
                    all_predictions[task]['preds'].extend(task_preds['preds'])
                    all_predictions[task]['probs'].extend(task_preds['probs'])
                else:
                    all_predictions[task]['labels'].extend(task_preds['labels'])
                    all_predictions[task]['preds'].extend(task_preds['preds'])
        
        # ๊ฐ ํƒœ์Šคํฌ๋ณ„ ํ‰๊ฐ€ ์ง€ํ‘œ ๊ณ„์‚ฐ ๋ฐ ์ €์žฅ
        results = {}
        for task in self.task_list:
            task_preds = all_predictions[task]
            # ๋น„์–ด์žˆ๊ฑฐ๋‚˜, ํ…์„œ๊ฐ€ ์•„๋‹Œ ๊ฐ’์ด ๋“ค์–ด์žˆ์œผ๋ฉด metric ๊ณ„์‚ฐ ๊ฑด๋„ˆ๋œ€
            if not task_preds['labels'] or not task_preds['preds']:
                continue
            if not all(isinstance(x, torch.Tensor) for x in task_preds['labels']):
                continue
            if not all(isinstance(x, torch.Tensor) for x in task_preds['preds']):
                continue
            
            if self.task_types[task] == 'classification':
                # ๋ถ„๋ฅ˜ ํƒœ์Šคํฌ ํ‰๊ฐ€
                labels_np = torch.stack(task_preds['labels']).numpy()
                preds_np = torch.stack(task_preds['preds']).numpy()
                
                acc = self.metrics[task]['accuracy'](labels_np, preds_np)
                prec, rec, f1 = self.metrics[task]['precision_recall_f1'](labels_np, preds_np)
                
                # ๊ฒฐ๊ณผ ์ €์žฅ
                results[f'{task}_acc'] = acc
                results[f'{task}_precision'] = prec
                results[f'{task}_recall'] = rec
                results[f'{task}_f1'] = f1
                
                # ๋กœ๊น…
                self.log(f'test_{task}_acc', acc, sync_dist=True)
                self.log(f'test_{task}_precision', prec, sync_dist=True)
                self.log(f'test_{task}_recall', rec, sync_dist=True)
                self.log(f'test_{task}_f1', f1, sync_dist=True)
            else:
                # ํšŒ๊ท€ ํƒœ์Šคํฌ ํ‰๊ฐ€
                labels_np = torch.stack(task_preds['labels']).numpy()
                preds_np = torch.stack(task_preds['preds']).numpy()
                # === NaN ๋งˆ์Šคํ‚น ์ถ”๊ฐ€ ===
                valid_mask = (~np.isnan(labels_np)) & (~np.isnan(preds_np))
                labels_np_valid = labels_np[valid_mask]
                preds_np_valid = preds_np[valid_mask]

                if len(preds_np_valid) < 2:  # ๋ชจ๋‘ ๊ฒฐ์ธก ๋˜๋Š” ์ƒ˜ํ”Œ ๋ถ€์กฑ
                    continue
                mse = self.metrics[task]['mse'](labels_np_valid, preds_np_valid)
                mae = self.metrics[task]['mae'](labels_np_valid, preds_np_valid)
                r2 = self.metrics[task]['r2'](labels_np_valid, preds_np_valid)
                # ์ˆ˜์ •: test metric๋งŒ ๋กœ๊น…
                self.log(f'test_{task}_mse', mse, sync_dist=True)
                self.log(f'test_{task}_mae', mae, sync_dist=True)
                self.log(f'test_{task}_r2', r2, sync_dist=True)

        # SMILES์™€ predictions๋ฅผ results์— ์ถ”๊ฐ€
        results['smiles'] = all_smiles
        results['predictions'] = all_predictions

        # ๊ฒฐ๊ณผ๋ฅผ ์บ์‹œ์— ์ €์žฅ (trainer์—์„œ ์ ‘๊ทผ ๊ฐ€๋Šฅํ•˜๋„๋ก)
        self.test_results_cache = results

        # ๋‹จ๊ณ„๊ฐ€ ๋๋‚˜๋ฉด ์ถœ๋ ฅ ๋ชฉ๋ก ์ดˆ๊ธฐํ™”
        self.test_step_outputs.clear()

        # ์ „์ฒด ๊ฒฐ๊ณผ ๋ฐ˜ํ™˜ (์ˆซ์ž ๋ฉ”ํŠธ๋ฆญ๋งŒ ๋กœ๊น…)
        numeric_results = {k: v for k, v in results.items() if isinstance(v, (int, float))}
        self.log_dict(numeric_results, sync_dist=True)
        return results
    
    def configure_optimizers(self):
        """์˜ตํ‹ฐ๋งˆ์ด์ € ์„ค์ •"""
        # ๊ฐ€์ค‘์น˜ ๊ฐ์‡  ์ œ์™ธ ํŒŒ๋ผ๋ฏธํ„ฐ ์„ค์ •
        no_decay = ["bias", "LayerNorm.weight"]
        optimizer_grouped_parameters = [
            {
                "params": [p for n, p in self.named_parameters() if not any(nd in n for nd in no_decay)],
                "weight_decay": self.weight_decay,
            },
            {
                "params": [p for n, p in self.named_parameters() if any(nd in n for nd in no_decay)],
                "weight_decay": 0.0,
            },
        ]
        
        # ์˜ตํ‹ฐ๋งˆ์ด์ € ์„ค์ •
        optimizer = torch.optim.AdamW(optimizer_grouped_parameters, lr=self.learning_rate)
        
        # ์Šค์ผ€์ค„๋Ÿฌ ์„ค์ • (์›Œ๋ฐ์—… ํฌํ•จ)
        # ์ด ์Šคํ… ์ˆ˜๋Š” Trainer๊ฐ€ ์ œ๊ณตํ•˜๋Š” ์ถ”์ •์น˜๋ฅผ ์‚ฌ์šฉ
        if self.trainer is not None:
            total_steps = max(1, int(self.trainer.estimated_stepping_batches))
        else:
            total_steps = 1000

        num_training_steps = total_steps
        scheduler = get_linear_schedule_with_warmup(
            optimizer,
            num_warmup_steps=min(self.warmup_steps, num_training_steps - 1),
            num_training_steps=num_training_steps
        )
        
        return {
            "optimizer": optimizer,
            "lr_scheduler": {
                "scheduler": scheduler,
                "interval": "step",
            },
        }