File size: 11,277 Bytes
47c017c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Social Media Sentiment Analysis - End-to-End Training Pipeline
Fine-tunes DeBERTa-v3-base on SST-2 + Tweet Sentiment for 95%+ accuracy
Based on: DeBERTaV3 paper (arxiv:2111.09543) training recipe
Optimized for CPU training with gradient accumulation
"""

import os
import json
import numpy as np
import torch
from datasets import load_dataset, concatenate_datasets
from transformers import (
    AutoTokenizer,
    AutoModelForSequenceClassification,
    TrainingArguments,
    Trainer,
    DataCollatorWithPadding,
    EarlyStoppingCallback,
)
import evaluate
import time

# ── Configuration ──────────────────────────────────────────────────
MODEL_NAME = "microsoft/deberta-v3-base"
HUB_MODEL_ID = "rajvivan/deberta-v3-sentiment-analysis"
MAX_LENGTH = 128  # Shorter for CPU efficiency
NUM_EPOCHS = 3
LEARNING_RATE = 2e-5
TRAIN_BATCH_SIZE = 8  # Small for CPU RAM
EVAL_BATCH_SIZE = 16
GRADIENT_ACCUM = 4  # Effective batch = 32
WARMUP_STEPS = 300
WEIGHT_DECAY = 0.01
SEED = 42

device = 'cuda' if torch.cuda.is_available() else 'cpu'
use_fp16 = torch.cuda.is_available()

print(f"πŸš€ Starting Sentiment Analysis Training")
print(f"   Model: {MODEL_NAME}")
print(f"   Hub ID: {HUB_MODEL_ID}")
print(f"   Device: {device}")
print(f"   Effective batch size: {TRAIN_BATCH_SIZE * GRADIENT_ACCUM}")

start_time = time.time()

# ── 1. Load and Prepare Datasets ──────────────────────────────────
print("\nπŸ“¦ Loading datasets...")

sst2 = load_dataset("stanfordnlp/sst2")
tweets = load_dataset("mteb/tweet_sentiment_extraction")

print(f"   SST-2 train: {len(sst2['train'])} | val: {len(sst2['validation'])}")
print(f"   Tweet train: {len(tweets['train'])} | test: {len(tweets['test'])}")

# ── 2. Tokenizer ──────────────────────────────────────────────────
print("\nπŸ”€ Loading tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

def preprocess_sst2(examples):
    return tokenizer(examples["sentence"], truncation=True, max_length=MAX_LENGTH, padding=False)

def preprocess_tweet_text(text):
    """Twitter-specific preprocessing (TimeLM paper)"""
    if not text:
        return ""
    return " ".join(
        '@user' if t.startswith('@') and len(t) > 1 else ('http' if t.startswith('http') else t)
        for t in text.split(" ")
    )

def preprocess_tweets(examples):
    texts = [preprocess_tweet_text(t) for t in examples["text"]]
    return tokenizer(texts, truncation=True, max_length=MAX_LENGTH, padding=False)

# Tokenize SST-2
print("   Tokenizing SST-2...")
sst2_tok = sst2.map(preprocess_sst2, batched=True, num_proc=2)
sst2_tok = sst2_tok.remove_columns(["idx", "sentence"])

# Prepare tweets: binary (remove neutral, remap labels)
print("   Preparing tweet binary data...")
tweets_train_bin = tweets["train"].filter(lambda x: x["label"] != 1)
tweets_test_bin = tweets["test"].filter(lambda x: x["label"] != 1)

def remap_labels(example):
    example["label"] = 1 if example["label"] == 2 else 0
    return example

tweets_train_bin = tweets_train_bin.map(remap_labels)
tweets_test_bin = tweets_test_bin.map(remap_labels)

tweets_train_tok = tweets_train_bin.map(preprocess_tweets, batched=True, num_proc=2)
tweets_test_tok = tweets_test_bin.map(preprocess_tweets, batched=True, num_proc=2)
tweets_train_tok = tweets_train_tok.remove_columns(["id", "text", "label_text"])
tweets_test_tok = tweets_test_bin.map(preprocess_tweets, batched=True, num_proc=2)
tweets_test_tok = tweets_test_tok.remove_columns(["id", "text", "label_text"])

# Cast label types to match SST-2 ClassLabel
from datasets import ClassLabel, Features, Value
target_features = sst2_tok["train"].features
tweets_train_tok = tweets_train_tok.cast(target_features)
tweets_test_tok = tweets_test_tok.cast(target_features)

# Combine for training
combined_train = concatenate_datasets([sst2_tok["train"], tweets_train_tok])
print(f"   Combined train: {len(combined_train)} samples")
print(f"   SST-2 val: {len(sst2_tok['validation'])} samples")
print(f"   Tweet test: {len(tweets_test_tok)} samples")

# ── 3. Model ──────────────────────────────────────────────────────
print("\n🧠 Loading model...")
id2label = {0: "NEGATIVE", 1: "POSITIVE"}
label2id = {"NEGATIVE": 0, "POSITIVE": 1}

model = AutoModelForSequenceClassification.from_pretrained(
    MODEL_NAME, num_labels=2, id2label=id2label, label2id=label2id,
)
total_params = model.num_parameters()
trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f"   Total params: {total_params:,}")
print(f"   Trainable: {trainable:,}")

# ── 4. Metrics ────────────────────────────────────────────────────
accuracy_metric = evaluate.load("accuracy")
f1_metric = evaluate.load("f1")
precision_metric = evaluate.load("precision")
recall_metric = evaluate.load("recall")

def compute_metrics(eval_pred):
    logits, labels = eval_pred
    preds = np.argmax(logits, axis=-1)
    return {
        "accuracy": accuracy_metric.compute(predictions=preds, references=labels)["accuracy"],
        "f1": f1_metric.compute(predictions=preds, references=labels, average="weighted")["f1"],
        "precision": precision_metric.compute(predictions=preds, references=labels, average="weighted")["precision"],
        "recall": recall_metric.compute(predictions=preds, references=labels, average="weighted")["recall"],
    }

def preprocess_logits_for_metrics(logits, labels):
    if isinstance(logits, tuple):
        logits = logits[0]
    return logits.argmax(dim=-1)

# ── 5. Training ───────────────────────────────────────────────────
print("\nβš™οΈ Configuring training...")
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)

training_args = TrainingArguments(
    output_dir="./deberta-sentiment-output",
    hub_model_id=HUB_MODEL_ID,
    push_to_hub=True,
    num_train_epochs=NUM_EPOCHS,
    per_device_train_batch_size=TRAIN_BATCH_SIZE,
    per_device_eval_batch_size=EVAL_BATCH_SIZE,
    gradient_accumulation_steps=GRADIENT_ACCUM,
    learning_rate=LEARNING_RATE,
    weight_decay=WEIGHT_DECAY,
    warmup_steps=WARMUP_STEPS,
    lr_scheduler_type="linear",
    max_grad_norm=1.0,
    eval_strategy="epoch",
    save_strategy="epoch",
    logging_strategy="steps",
    logging_steps=100,
    logging_first_step=True,
    disable_tqdm=True,
    load_best_model_at_end=True,
    metric_for_best_model="accuracy",
    greater_is_better=True,
    bf16=False,
    fp16=use_fp16,
    seed=SEED,
    report_to="none",
    save_total_limit=2,
    dataloader_num_workers=2,
    gradient_checkpointing=True,  # Save memory
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=combined_train,
    eval_dataset=sst2_tok["validation"],
    processing_class=tokenizer,
    data_collator=data_collator,
    compute_metrics=compute_metrics,
    preprocess_logits_for_metrics=preprocess_logits_for_metrics,
    callbacks=[EarlyStoppingCallback(early_stopping_patience=2)],
)

# ── 6. Train ──────────────────────────────────────────────────────
print("\nπŸ‹οΈ Starting training...")
print(f"   Steps per epoch: ~{len(combined_train) // (TRAIN_BATCH_SIZE * GRADIENT_ACCUM)}")
train_result = trainer.train()
train_time = time.time() - start_time

# ── 7. Evaluate ───────────────────────────────────────────────────
print("\nπŸ“Š Evaluating on SST-2 validation...")
sst2_results = trainer.evaluate(sst2_tok["validation"])
print(f"   Accuracy:  {sst2_results['eval_accuracy']:.4f}")
print(f"   F1:        {sst2_results['eval_f1']:.4f}")
print(f"   Precision: {sst2_results['eval_precision']:.4f}")
print(f"   Recall:    {sst2_results['eval_recall']:.4f}")

print("\nπŸ“Š Evaluating on Tweet Sentiment test set...")
tweet_results = trainer.evaluate(tweets_test_tok)
print(f"   Accuracy:  {tweet_results['eval_accuracy']:.4f}")
print(f"   F1:        {tweet_results['eval_f1']:.4f}")
print(f"   Precision: {tweet_results['eval_precision']:.4f}")
print(f"   Recall:    {tweet_results['eval_recall']:.4f}")

# ── 8. Save results ──────────────────────────────────────────────
os.makedirs("./deberta-sentiment-output", exist_ok=True)
results = {
    "model": MODEL_NAME,
    "hub_model_id": HUB_MODEL_ID,
    "total_parameters": total_params,
    "training_config": {
        "learning_rate": LEARNING_RATE,
        "effective_batch_size": TRAIN_BATCH_SIZE * GRADIENT_ACCUM,
        "per_device_batch_size": TRAIN_BATCH_SIZE,
        "gradient_accumulation_steps": GRADIENT_ACCUM,
        "epochs": NUM_EPOCHS,
        "warmup_steps": WARMUP_STEPS,
        "weight_decay": WEIGHT_DECAY,
        "max_length": MAX_LENGTH,
        "seed": SEED,
        "gradient_checkpointing": True,
    },
    "datasets": {
        "sst2_train": len(sst2_tok["train"]),
        "tweet_train_binary": len(tweets_train_tok),
        "combined_train": len(combined_train),
        "sst2_val": len(sst2_tok["validation"]),
        "tweet_test_binary": len(tweets_test_tok),
    },
    "sst2_eval_results": {k.replace("eval_", ""): v for k, v in sst2_results.items()},
    "tweet_eval_results": {k.replace("eval_", ""): v for k, v in tweet_results.items()},
    "training_loss": train_result.training_loss,
    "training_time_seconds": train_time,
}

with open("./deberta-sentiment-output/results.json", "w") as f:
    json.dump(results, f, indent=2)
print(f"\nπŸ’Ύ Results saved")

# ── 9. Push to Hub ────────────────────────────────────────────────
print("\nπŸš€ Pushing model to Hub...")
trainer.push_to_hub(
    commit_message="DeBERTa-v3-base sentiment analysis - SST-2 + Tweet binary",
    tags=["sentiment-analysis", "deberta-v3", "social-media", "text-classification"],
)
print(f"βœ… Model published: https://huggingface.co/{HUB_MODEL_ID}")

# ── Summary ───────────────────────────────────────────────────────
print("\n" + "="*60)
print("TRAINING COMPLETE - FINAL RESULTS")
print("="*60)
print(f"Training time:   {train_time/60:.1f} minutes")
print(f"Training loss:   {train_result.training_loss:.4f}")
print(f"SST-2 Accuracy:  {sst2_results['eval_accuracy']:.4f}")
print(f"SST-2 F1:        {sst2_results['eval_f1']:.4f}")
print(f"Tweet Accuracy:  {tweet_results['eval_accuracy']:.4f}")
print(f"Tweet F1:        {tweet_results['eval_f1']:.4f}")
print("="*60)