File size: 31,776 Bytes
eccd289
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
{
    "cells": [
        {
            "cell_type": "markdown",
            "metadata": {},
            "source": [
                "# Deep Learning Project - Emotion Classification\n",
                "\n",
                "This notebook implements a **multi-label emotion classification system** using state-of-the-art transformer models. The goal is to predict multiple emotions (anger, fear, joy, sadness, surprise) that may be present in a given text.\n",
                "\n",
                "**Key Features:**\n",
                "- **Model**: Microsoft DeBERTa-v3-base (184M parameters)\n",
                "- **Strategy**: 5-Fold Stratified Cross-Validation for robust performance estimation\n",
                "- **Optimization**: Mixed Precision Training, Gradient Clipping, Learning Rate Warmup\n",
                "- **Evaluation**: Macro F1 Score with Per-Label Threshold Tuning\n",
                "- **Ensemble**: Average predictions across all folds for final submission\n",
                "\n",
                "**Problem Type**: Multi-label classification (each text can have 0 or more emotions)\n",
                "\n",
                "---\n",
                "\n",
                "## 1. Imports & Setup\n",
                "\n",
                "We import all necessary libraries for:\n",
                "- **Data handling**: `numpy`, `pandas` for data manipulation\n",
                "- **Deep learning**: `torch` (PyTorch) and `transformers` (Hugging Face) for model training\n",
                "- **Evaluation**: `sklearn` for F1 metrics and stratified k-fold cross-validation\n",
                "- **Optimization**: Mixed precision training with `autocast` and `GradScaler` to speed up training and reduce memory usage\n",
                "- **Memory management**: `gc` for garbage collection to free up GPU memory between folds"
            ]
        },
        {
            "cell_type": "code",
            "execution_count": null,
            "metadata": {},
            "outputs": [],
            "source": [
                "import numpy as np\n",
                "import pandas as pd\n",
                "import torch\n",
                "import torch.nn as nn\n",
                "from sklearn.model_selection import StratifiedKFold\n",
                "from sklearn.metrics import f1_score\n",
                "from transformers import (\n",
                "    AutoTokenizer,\n",
                "    AutoModelForSequenceClassification,\n",
                "    get_linear_schedule_with_warmup,\n",
                "    AutoConfig\n",
                ")\n",
                "from torch.optim import AdamW\n",
                "from torch.cuda.amp import autocast, GradScaler\n",
                "import gc\n",
                "import warnings\n",
                "import os\n",
                "\n",
                "warnings.filterwarnings(\"ignore\")"
            ]
        },
        {
            "cell_type": "markdown",
            "metadata": {},
            "source": [
                "## 2. Configuration\n",
                "\n",
                "Centralized configuration class containing all hyperparameters and paths. This makes it easy to:\n",
                "- Experiment with different settings\n",
                "- Ensure reproducibility\n",
                "- Keep the code organized\n",
                "\n",
                "**Key Hyperparameters:**\n",
                "- `MODEL_NAME`: DeBERTa-v3-base chosen for its strong performance on text classification tasks\n",
                "- `MAX_LEN=128`: Balance between capturing context and computational efficiency\n",
                "- `BATCH_SIZE=16`: Fits in GPU memory while maintaining good gradient estimates\n",
                "- `LR=1.5e-5`: Small learning rate typical for fine-tuning pre-trained transformers\n",
                "- `EPOCHS=4`: Sufficient for fine-tuning without overfitting\n",
                "- `N_FOLDS=5`: Standard choice for cross-validation, balances between training data and validation reliability"
            ]
        },
        {
            "cell_type": "code",
            "execution_count": null,
            "metadata": {},
            "outputs": [],
            "source": [
                "# ========= CONFIG =========\n",
                "class Config:\n",
                "    SEED = 42\n",
                "    LABELS = [\"anger\", \"fear\", \"joy\", \"sadness\", \"surprise\"]\n",
                "    MODEL_NAME = \"microsoft/deberta-v3-base\"\n",
                "    MAX_LEN = 128\n",
                "    BATCH_SIZE = 16\n",
                "    EPOCHS = 4\n",
                "    LR = 1.5e-5\n",
                "    WEIGHT_DECAY = 0.01\n",
                "    WARMUP_RATIO = 0.1\n",
                "    N_FOLDS = 5\n",
                "    TRAIN_CSV = \"/kaggle/input/2025-sep-dl-gen-ai-project/train.csv\"\n",
                "    TEST_CSV = \"/kaggle/input/2025-sep-dl-gen-ai-project/test.csv\"\n",
                "    SUBMISSION_PATH = \"submission.csv\"\n",
                "\n",
                "CONFIG = Config()"
            ]
        },
        {
            "cell_type": "markdown",
            "metadata": {},
            "source": [
                "## 3. Seed & Device Setup\n",
                "\n",
                "**Reproducibility**: Setting seeds ensures that our results can be replicated exactly.\n",
                "We set seeds for:\n",
                "- NumPy random number generation\n",
                "- PyTorch CPU operations\n",
                "- PyTorch GPU operations (all CUDA devices)\n",
                "- Python's built-in hash function\n",
                "\n",
                "**Device Selection**: Automatically detects and uses GPU if available (CUDA), otherwise falls back to CPU.\n",
                "GPU training is significantly faster (~10-50x) than CPU for deep learning models."
            ]
        },
        {
            "cell_type": "code",
            "execution_count": null,
            "metadata": {},
            "outputs": [],
            "source": [
                "def set_seed(seed=CONFIG.SEED):\n",
                "    np.random.seed(seed)\n",
                "    torch.manual_seed(seed)\n",
                "    torch.cuda.manual_seed_all(seed)\n",
                "    os.environ['PYTHONHASHSEED'] = str(seed)\n",
                "\n",
                "set_seed()\n",
                "\n",
                "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
                "print(f\"Using device: {device}\")"
            ]
        },
        {
            "cell_type": "markdown",
            "metadata": {},
            "source": [
                "## 4. Utility Functions\n",
                "\n",
                "Helper functions used throughout the pipeline:\n",
                "\n",
                "### `ensure_text_column(df)`\n",
                "- Standardizes the text column name across different datasets\n",
                "- Searches for common alternatives like 'comment_text', 'sentence', etc.\n",
                "- Raises error if no text column is found\n",
                "\n",
                "### `tune_thresholds(y_true, y_prob)`\n",
                "- **Critical for multi-label classification performance**\n",
                "- Default threshold of 0.5 is often suboptimal\n",
                "- Finds the best threshold per label that maximizes F1 score\n",
                "- Tests 17 different thresholds between 0.1 and 0.9\n",
                "- Can improve F1 score by 2-5% over default threshold\n",
                "\n",
                "### `get_optimizer_params(model, lr, weight_decay)`\n",
                "- Implements **differential weight decay**\n",
                "- Applies weight decay to most parameters (helps prevent overfitting)\n",
                "- No weight decay for bias and LayerNorm parameters (standard practice in transformer fine-tuning)\n",
                "- This technique is recommended in the BERT and DeBERTa papers"
            ]
        },
        {
            "cell_type": "code",
            "execution_count": null,
            "metadata": {},
            "outputs": [],
            "source": [
                "def ensure_text_column(df: pd.DataFrame) -> pd.DataFrame:\n",
                "    if \"text\" in df.columns:\n",
                "        return df\n",
                "    for c in [\"comment_text\", \"sentence\", \"content\", \"review\"]:\n",
                "        if c in df.columns:\n",
                "            return df.rename(columns={c: \"text\"})\n",
                "    raise ValueError(\"No text column found. Add/rename your text column to 'text'.\")\n",
                "\n",
                "def tune_thresholds(y_true: np.ndarray, y_prob: np.ndarray) -> np.ndarray:\n",
                "    th = np.zeros(y_true.shape[1], dtype=np.float32)\n",
                "    for j in range(y_true.shape[1]):\n",
                "        best_t, best_f1 = 0.5, -1\n",
                "        for t in np.linspace(0.1, 0.9, 17):\n",
                "            f1 = f1_score(y_true[:, j], (y_prob[:, j] >= t).astype(int), zero_division=0)\n",
                "            if f1 > best_f1:\n",
                "                best_f1, best_t = f1, t\n",
                "        th[j] = best_t\n",
                "    return th\n",
                "\n",
                "def get_optimizer_params(model, lr, weight_decay):\n",
                "    param_optimizer = list(model.named_parameters())\n",
                "    no_decay = [\"bias\", \"LayerNorm.bias\", \"LayerNorm.weight\"]\n",
                "    optimizer_parameters = [\n",
                "        {\n",
                "            \"params\": [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],\n",
                "            \"weight_decay\": weight_decay,\n",
                "        },\n",
                "        {\n",
                "            \"params\": [p for n, p in param_optimizer if any(nd in n for nd in no_decay)],\n",
                "            \"weight_decay\": 0.0,\n",
                "        },\n",
                "    ]\n",
                "    return optimizer_parameters"
            ]
        },
        {
            "cell_type": "markdown",
            "metadata": {},
            "source": [
                "## 5. Dataset Class\n",
                "\n",
                "Custom PyTorch Dataset for emotion classification.\n",
                "\n",
                "**Key Features:**\n",
                "- Tokenizes text on-the-fly using the DeBERTa tokenizer\n",
                "- Handles both training data (with labels) and test data (without labels)\n",
                "- Uses `padding='max_length'` to ensure all sequences have the same length (required for batching)\n",
                "- Applies `truncation=True` to handle texts longer than MAX_LEN\n",
                "\n",
                "**Returns:**\n",
                "- `input_ids`: Token IDs representing the text\n",
                "- `attention_mask`: Indicates which tokens are real vs padding\n",
                "- `labels`: Multi-label binary targets (only for training data)\n",
                "\n",
                "**PyTorch DataLoader** will use this dataset to create batches efficiently with multi-processing."
            ]
        },
        {
            "cell_type": "code",
            "execution_count": null,
            "metadata": {},
            "outputs": [],
            "source": [
                "class EmotionDS(torch.utils.data.Dataset):\n",
                "    def __init__(self, df, tokenizer, max_len, is_test=False):\n",
                "        self.texts = df[\"text\"].tolist()\n",
                "        self.is_test = is_test\n",
                "        if not is_test:\n",
                "            self.labels = df[CONFIG.LABELS].values.astype(np.float32)\n",
                "        self.tok = tokenizer\n",
                "        self.max_len = max_len\n",
                "\n",
                "    def __len__(self):\n",
                "        return len(self.texts)\n",
                "\n",
                "    def __getitem__(self, i):\n",
                "        enc = self.tok(\n",
                "            self.texts[i],\n",
                "            truncation=True,\n",
                "            padding=\"max_length\",\n",
                "            max_length=self.max_len,\n",
                "            return_tensors=\"pt\",\n",
                "        )\n",
                "        item = {k: v.squeeze(0) for k, v in enc.items()}\n",
                "        if not self.is_test:\n",
                "            item[\"labels\"] = torch.tensor(self.labels[i])\n",
                "        return item"
            ]
        },
        {
            "cell_type": "markdown",
            "metadata": {},
            "source": [
                "## 6. Training & Validation Helper Functions\n",
                "\n",
                "Core training and validation loops.\n",
                "\n",
                "### `train_one_epoch()`\n",
                "Trains the model for one complete pass through the training data.\n",
                "\n",
                "**Key Techniques:**\n",
                "- **Mixed Precision Training** (`autocast`): Uses float16 where safe, reducing memory and increasing speed by ~2x\n",
                "- **Gradient Scaling** (`GradScaler`): Prevents gradient underflow in mixed precision\n",
                "- **Gradient Clipping** (max_norm=1.0): Prevents exploding gradients, stabilizes training\n",
                "- **Memory Efficient**: Uses `zero_grad(set_to_none=True)` and `non_blocking=True` for async GPU transfers\n",
                "\n",
                "### `validate()`\n",
                "Evaluates the model on validation data without updating weights.\n",
                "\n",
                "**Features:**\n",
                "- Runs in `model.eval()` mode (disables dropout, fixes batch normalization)\n",
                "- Uses `torch.no_grad()` to save memory (no gradient computation)\n",
                "- Applies sigmoid to convert logits to probabilities [0, 1]\n",
                "- Returns predictions and targets for metric calculation"
            ]
        },
        {
            "cell_type": "code",
            "execution_count": null,
            "metadata": {},
            "outputs": [],
            "source": [
                "def train_one_epoch(model, loader, optimizer, scheduler, scaler, criterion):\n",
                "    model.train()\n",
                "    losses = []\n",
                "    for batch in loader:\n",
                "        batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}\n",
                "        optimizer.zero_grad(set_to_none=True)\n",
                "        with autocast(enabled=True):\n",
                "            out = model(input_ids=batch[\"input_ids\"], attention_mask=batch[\"attention_mask\"])\n",
                "            loss = criterion(out.logits, batch[\"labels\"])\n",
                "        \n",
                "        scaler.scale(loss).backward()\n",
                "        scaler.unscale_(optimizer)\n",
                "        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)\n",
                "        scaler.step(optimizer)\n",
                "        scaler.update()\n",
                "        scheduler.step()\n",
                "        losses.append(loss.item())\n",
                "    return np.mean(losses)\n",
                "\n",
                "def validate(model, loader, criterion):\n",
                "    model.eval()\n",
                "    losses = []\n",
                "    preds = []\n",
                "    targs = []\n",
                "    with torch.no_grad():\n",
                "        for batch in loader:\n",
                "            batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}\n",
                "            with autocast(enabled=True):\n",
                "                out = model(input_ids=batch[\"input_ids\"], attention_mask=batch[\"attention_mask\"])\n",
                "                loss = criterion(out.logits, batch[\"labels\"])\n",
                "            losses.append(loss.item())\n",
                "            preds.append(torch.sigmoid(out.logits).float().cpu().numpy())\n",
                "            targs.append(batch[\"labels\"].cpu().numpy())\n",
                "    \n",
                "    return np.mean(losses), np.vstack(preds), np.vstack(targs)"
            ]
        },
        {
            "cell_type": "markdown",
            "metadata": {},
            "source": [
                "## 7. Main K-Fold Training Loop\n",
                "\n",
                "The heart of our training pipeline - implements **5-Fold Stratified Cross-Validation**.\n",
                "\n",
                "### Why K-Fold Cross-Validation?\n",
                "- More reliable performance estimates than a single train/val split\n",
                "- Every sample is used for validation exactly once\n",
                "- Out-of-fold predictions can be used for threshold optimization\n",
                "- Reduces variance in model performance\n",
                "\n",
                "### Why Stratified?\n",
                "- Maintains label distribution in each fold\n",
                "- Important for imbalanced datasets\n",
                "- For multi-label, we concatenate all labels into a string for stratification\n",
                "\n",
                "### Training Process per Fold:\n",
                "1. **Split data**: 80% training, 20% validation\n",
                "2. **Initialize model**: Fresh DeBERTa-v3-base with random classification head\n",
                "3. **Setup optimizer**: AdamW with differential weight decay\n",
                "4. **Setup scheduler**: Linear warmup (10% of steps) then linear decay to 0\n",
                "5. **Train for 4 epochs**: Track training and validation metrics\n",
                "6. **Save best model**: Based on validation F1 score\n",
                "7. **Store OOF predictions**: For threshold tuning\n",
                "8. **Clean up memory**: Delete model and optimizer, run garbage collection\n",
                "\n",
                "### Output:\n",
                "- 5 trained models (one per fold) saved as `model_fold_{0-4}.pth`\n",
                "- Out-of-fold predictions for the entire training set\n",
                "- Cross-validated performance metrics"
            ]
        },
        {
            "cell_type": "code",
            "execution_count": null,
            "metadata": {},
            "outputs": [],
            "source": [
                "def run_training():\n",
                "    if not os.path.exists(CONFIG.TRAIN_CSV):\n",
                "        print(\"Train CSV not found. Please check the path.\")\n",
                "        return None, None\n",
                "\n",
                "    df = pd.read_csv(CONFIG.TRAIN_CSV)\n",
                "    df = ensure_text_column(df)\n",
                "    \n",
                "    # Create Stratified Folds\n",
                "    skf = StratifiedKFold(n_splits=CONFIG.N_FOLDS, shuffle=True, random_state=CONFIG.SEED)\n",
                "    y_str = df[CONFIG.LABELS].astype(str).agg(\"\".join, axis=1)\n",
                "    \n",
                "    oof_preds = np.zeros((len(df), len(CONFIG.LABELS)))\n",
                "    \n",
                "    tokenizer = AutoTokenizer.from_pretrained(CONFIG.MODEL_NAME)\n",
                "    \n",
                "    for fold, (train_idx, val_idx) in enumerate(skf.split(df, y_str)):\n",
                "        print(f\"\\n{'='*20} FOLD {fold+1}/{CONFIG.N_FOLDS} {'='*20}\")\n",
                "        \n",
                "        df_tr = df.iloc[train_idx].reset_index(drop=True)\n",
                "        df_va = df.iloc[val_idx].reset_index(drop=True)\n",
                "        \n",
                "        ds_tr = EmotionDS(df_tr, tokenizer, CONFIG.MAX_LEN)\n",
                "        ds_va = EmotionDS(df_va, tokenizer, CONFIG.MAX_LEN)\n",
                "        \n",
                "        dl_tr = torch.utils.data.DataLoader(ds_tr, batch_size=CONFIG.BATCH_SIZE, shuffle=True, num_workers=2, pin_memory=True)\n",
                "        dl_va = torch.utils.data.DataLoader(ds_va, batch_size=CONFIG.BATCH_SIZE, shuffle=False, num_workers=2, pin_memory=True)\n",
                "        \n",
                "        model = AutoModelForSequenceClassification.from_pretrained(\n",
                "            CONFIG.MODEL_NAME, \n",
                "            num_labels=len(CONFIG.LABELS),\n",
                "            problem_type=\"multi_label_classification\"\n",
                "        )\n",
                "        model.to(device)\n",
                "        \n",
                "        optimizer_params = get_optimizer_params(model, CONFIG.LR, CONFIG.WEIGHT_DECAY)\n",
                "        optimizer = AdamW(optimizer_params, lr=CONFIG.LR)\n",
                "        \n",
                "        total_steps = len(dl_tr) * CONFIG.EPOCHS\n",
                "        scheduler = get_linear_schedule_with_warmup(\n",
                "            optimizer, \n",
                "            num_warmup_steps=int(total_steps * CONFIG.WARMUP_RATIO), \n",
                "            num_training_steps=total_steps\n",
                "        )\n",
                "        \n",
                "        criterion = nn.BCEWithLogitsLoss()\n",
                "        scaler = GradScaler(enabled=True)\n",
                "        \n",
                "        best_f1 = 0\n",
                "        best_state = None\n",
                "        \n",
                "        for ep in range(CONFIG.EPOCHS):\n",
                "            train_loss = train_one_epoch(model, dl_tr, optimizer, scheduler, scaler, criterion)\n",
                "            val_loss, val_preds, val_targs = validate(model, dl_va, criterion)\n",
                "            \n",
                "            val_f1 = f1_score(val_targs, (val_preds >= 0.5).astype(int), average=\"macro\", zero_division=0)\n",
                "            \n",
                "            print(f\"Ep {ep+1}: TrLoss={train_loss:.4f} | VaLoss={val_loss:.4f} | VaF1={val_f1:.4f}\")\n",
                "            \n",
                "            if val_f1 > best_f1:\n",
                "                best_f1 = val_f1\n",
                "                best_state = model.state_dict()\n",
                "                \n",
                "        torch.save(best_state, f\"model_fold_{fold}.pth\")\n",
                "        \n",
                "        model.load_state_dict(best_state)\n",
                "        _, val_preds, _ = validate(model, dl_va, criterion)\n",
                "        oof_preds[val_idx] = val_preds\n",
                "        \n",
                "        del model, optimizer, scaler, scheduler\n",
                "        torch.cuda.empty_cache()\n",
                "        gc.collect()\n",
                "        \n",
                "    return oof_preds, df[CONFIG.LABELS].values\n",
                "\n",
                "if os.path.exists(CONFIG.TRAIN_CSV):\n",
                "    oof_preds, y_true = run_training()\n",
                "else:\n",
                "    print(\"Skipping training as data is not found (likely in a dry-run environment).\")"
            ]
        },
        {
            "cell_type": "markdown",
            "metadata": {},
            "source": [
                "## 8. Threshold Optimization\n",
                "\n",
                "**Why optimize thresholds?**\n",
                "\n",
                "In multi-label classification, we convert probabilities to binary predictions using a threshold:\n",
                "- `prediction = 1 if probability >= threshold else 0`\n",
                "- The default threshold of 0.5 is often suboptimal\n",
                "- Different emotion labels may have different optimal thresholds\n",
                "\n",
                "**Example:**\n",
                "- 'Joy' might be common → optimal threshold could be 0.4\n",
                "- 'Surprise' might be rare → optimal threshold could be 0.6\n",
                "\n",
                "### Process:\n",
                "1. Use out-of-fold predictions (already trained models, no data leakage)\n",
                "2. For each label independently, test thresholds from 0.1 to 0.9\n",
                "3. Select threshold that maximizes F1 score for that label\n",
                "4. Apply optimized thresholds to get final binary predictions\n",
                "\n",
                "**Expected Improvement**: 2-5% increase in Macro F1 score\n",
                "\n",
                "This is a standard technique in Kaggle competitions and production systems."
            ]
        },
        {
            "cell_type": "code",
            "execution_count": null,
            "metadata": {},
            "outputs": [],
            "source": [
                "if os.path.exists(CONFIG.TRAIN_CSV):\n",
                "    best_thresholds = tune_thresholds(y_true, oof_preds)\n",
                "    oof_tuned = (oof_preds >= best_thresholds).astype(int)\n",
                "    final_f1 = f1_score(y_true, oof_tuned, average=\"macro\", zero_division=0)\n",
                "    print(f\"\\nFinal CV Macro F1: {final_f1:.4f}\")\n",
                "    print(f\"Best Thresholds: {best_thresholds}\")\n",
                "else:\n",
                "    best_thresholds = np.array([0.5] * len(CONFIG.LABELS))"
            ]
        },
        {
            "cell_type": "markdown",
            "metadata": {},
            "source": [
                "## 9. Inference & Submission\n",
                "\n",
                "Final prediction pipeline for test data.\n",
                "\n",
                "### Ensemble Strategy:\n",
                "We use **model averaging** across all 5 folds:\n",
                "1. Load each trained fold model\n",
                "2. Make predictions on test set\n",
                "3. Average the probabilities across all folds\n",
                "4. Apply optimized thresholds to get binary predictions\n",
                "\n",
                "### Why ensemble?\n",
                "- Reduces variance and overfitting\n",
                "- More robust predictions\n",
                "- Often improves score by 1-3%\n",
                "- Each fold sees different training data, captures different patterns\n",
                "\n",
                "### Process:\n",
                "1. Load test data and tokenize\n",
                "2. For each fold:\n",
                "   - Load saved model weights\n",
                "   - Generate predictions (probabilities)\n",
                "   - Clean up memory\n",
                "3. Average all fold predictions\n",
                "4. Apply optimized thresholds\n",
                "5. Create submission file with format: `id, anger, fear, joy, sadness, surprise`\n",
                "\n",
                "### Output:\n",
                "- `submission.csv` ready for Kaggle upload\n",
                "- Binary predictions (0 or 1) for each emotion per text"
            ]
        },
        {
            "cell_type": "code",
            "execution_count": null,
            "metadata": {},
            "outputs": [],
            "source": [
                "def predict_test(thresholds):\n",
                "    if not os.path.exists(CONFIG.TEST_CSV):\n",
                "        print(\"Test CSV not found.\")\n",
                "        return\n",
                "\n",
                "    df_test = pd.read_csv(CONFIG.TEST_CSV)\n",
                "    df_test = ensure_text_column(df_test)\n",
                "    \n",
                "    tokenizer = AutoTokenizer.from_pretrained(CONFIG.MODEL_NAME)\n",
                "    ds_test = EmotionDS(df_test, tokenizer, CONFIG.MAX_LEN, is_test=True)\n",
                "    dl_test = torch.utils.data.DataLoader(ds_test, batch_size=CONFIG.BATCH_SIZE, shuffle=False, num_workers=2)\n",
                "    \n",
                "    fold_preds = []\n",
                "    \n",
                "    for fold in range(CONFIG.N_FOLDS):\n",
                "        model_path = f\"model_fold_{fold}.pth\"\n",
                "        if not os.path.exists(model_path):\n",
                "            print(f\"Model for fold {fold} not found, skipping.\")\n",
                "            continue\n",
                "            \n",
                "        print(f\"Predicting Fold {fold+1}...\")\n",
                "        model = AutoModelForSequenceClassification.from_pretrained(\n",
                "            CONFIG.MODEL_NAME, \n",
                "            num_labels=len(CONFIG.LABELS),\n",
                "            problem_type=\"multi_label_classification\"\n",
                "        )\n",
                "        model.load_state_dict(torch.load(model_path))\n",
                "        model.to(device)\n",
                "        model.eval()\n",
                "        \n",
                "        preds = []\n",
                "        with torch.no_grad():\n",
                "            for batch in dl_test:\n",
                "                batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}\n",
                "                with autocast(enabled=True):\n",
                "                    out = model(input_ids=batch[\"input_ids\"], attention_mask=batch[\"attention_mask\"])\n",
                "                preds.append(torch.sigmoid(out.logits).float().cpu().numpy())\n",
                "        \n",
                "        fold_preds.append(np.vstack(preds))\n",
                "        del model\n",
                "        torch.cuda.empty_cache()\n",
                "        gc.collect()\n",
                "        \n",
                "    if not fold_preds:\n",
                "        print(\"No predictions made.\")\n",
                "        return\n",
                "\n",
                "    avg_preds = np.mean(fold_preds, axis=0)\n",
                "    final_preds = (avg_preds >= thresholds).astype(int)\n",
                "    \n",
                "    sub = pd.DataFrame(columns=[\"id\"] + CONFIG.LABELS)\n",
                "    sub[\"id\"] = df_test[\"id\"] if \"id\" in df_test.columns else np.arange(len(df_test))\n",
                "    sub[CONFIG.LABELS] = final_preds\n",
                "    sub.to_csv(CONFIG.SUBMISSION_PATH, index=False)\n",
                "    print(f\"Submission saved to {CONFIG.SUBMISSION_PATH}\")\n",
                "    print(sub.head())\n",
                "\n",
                "predict_test(best_thresholds)"
            ]
        }
    ],
    "metadata": {
        "kernelspec": {
            "display_name": ".venv",
            "language": "python",
            "name": "python3"
        },
        "language_info": {
            "codemirror_mode": {
                "name": "ipython",
                "version": 3
            },
            "file_extension": ".py",
            "mimetype": "text/x-python",
            "name": "python",
            "nbconvert_exporter": "python",
            "pygments_lexer": "ipython3",
            "version": "3.13.7"
        }
    },
    "nbformat": 4,
    "nbformat_minor": 5
}