py/fine_tune_sentiment.py

import torch
from transformers import (
    AutoTokenizer, 
    AutoModelForSequenceClassification,
    TrainingArguments,
    Trainer,
    DataCollatorWithPadding
)
from datasets import load_dataset, DatasetDict
import numpy as np
from sklearn.metrics import accuracy_score, f1_score, precision_recall_fscore_support, classification_report
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
from tqdm import tqdm
import warnings
warnings.filterwarnings('ignore')

class SentimentFineTuner:
    def __init__(self, model_name="5CD-AI/Vietnamese-Sentiment-visobert", dataset_name="uitnlp/vietnamese_students_feedback"):
        self.model_name = model_name
        self.dataset_name = dataset_name
        self.tokenizer = None
        self.model = None
        self.dataset = None
        self.tokenized_datasets = None
        
    def load_model_and_tokenizer(self):
        """Load the pre-trained model and tokenizer"""
        print(f"Loading model: {self.model_name}")
        print(f"Loading tokenizer...")
        
        self.tokenizer = AutoTokenizer.from_pretrained(self.model_name)
        self.model = AutoModelForSequenceClassification.from_pretrained(self.model_name)
        
        print("Model and tokenizer loaded successfully!")
        print(f"Model architecture: {self.model.config.architectures}")
        print(f"Number of labels: {self.model.config.num_labels}")
        
    def load_and_prepare_dataset(self):
        """Load and prepare the dataset"""
        print(f"Loading dataset: {self.dataset_name}")

        try:
            # Try loading the dataset directly
            self.dataset = load_dataset(self.dataset_name)
        except Exception as e:
            print(f"Error loading dataset directly: {e}")
            print("Attempting alternative dataset loading...")

            # Alternative approach: Create a synthetic Vietnamese sentiment dataset
            try:
                # Try to load a different Vietnamese dataset
                self.dataset = load_dataset("linhtranvi/5cdAI-Vietnamese-sentiment")
                print("Loaded alternative Vietnamese sentiment dataset!")
            except Exception as e2:
                print(f"Alternative dataset also failed: {e2}")
                print("Creating a sample Vietnamese sentiment dataset...")
                self.create_sample_dataset()
                return

        print("Dataset loaded successfully!")
        print(f"Dataset info: {self.dataset}")
        
        # Check the dataset structure
        print("\nDataset structure:")
        for split in self.dataset:
            print(f"{split}: {len(self.dataset[split])} samples")
            print(f"Columns: {self.dataset[split].column_names}")
            if len(self.dataset[split]) > 0:
                print(f"Sample data: {self.dataset[split][0]}")
        
        # The dataset should have sentiment labels
        # Let's check the unique sentiment labels
        if 'train' in self.dataset:
            train_df = pd.DataFrame(self.dataset['train'])
            if 'sentiment' in train_df.columns:
                print(f"\nSentiment distribution in training set:")
                print(train_df['sentiment'].value_counts())
            elif 'label' in train_df.columns:
                print(f"\nLabel distribution in training set:")
                print(train_df['label'].value_counts())
        
    def preprocess_function(self, examples):
        """Tokenize the dataset"""
        # Get the text column
        text_column = None
        for col in ['sentence', 'text', 'comment', 'feedback']:
            if col in examples:
                text_column = col
                break
        
        if text_column is None:
            # Use the first string column
            for col in examples:
                if isinstance(examples[col][0], str):
                    text_column = col
                    break
        
        if text_column is None:
            raise ValueError("No text column found in the dataset")
        
        # Get the label column
        label_column = None
        for col in ['sentiment', 'label', 'labels']:
            if col in examples:
                label_column = col
                break
        
        if label_column is None:
            raise ValueError("No label column found in the dataset")
        
        # Tokenize the text (matching original model max length)
        tokenized_inputs = self.tokenizer(
            examples[text_column],
            truncation=True,
            padding=False,
            max_length=256  # Matching original 5CD-AI/Vietnamese-Sentiment-visobert config
        )
        
        # Add labels
        tokenized_inputs['labels'] = examples[label_column]
        
        return tokenized_inputs
    
    def tokenize_datasets(self):
        """Tokenize all datasets"""
        print("Tokenizing datasets...")
        
        self.tokenized_datasets = self.dataset.map(
            self.preprocess_function,
            batched=True,
            remove_columns=self.dataset['train'].column_names
        )
        
        print("Tokenization completed!")
        
    def compute_metrics(self, eval_pred):
        """Compute evaluation metrics"""
        predictions, labels = eval_pred
        predictions = np.argmax(predictions, axis=1)
        
        accuracy = accuracy_score(labels, predictions)
        f1 = f1_score(labels, predictions, average='weighted')
        precision, recall, f1_weighted, _ = precision_recall_fscore_support(labels, predictions, average='weighted')
        
        return {
            'accuracy': accuracy,
            'f1': f1,
            'precision': precision,
            'recall': recall
        }
    
    def setup_trainer(self, output_dir="./sentiment_model", learning_rate=2e-5, batch_size=16, num_epochs=5):
        """Setup the trainer for fine-tuning"""
        
        # Data collator
        data_collator = DataCollatorWithPadding(tokenizer=self.tokenizer)
        
        # Training arguments (matching original 5CD-AI/Vietnamese-Sentiment-visobert config)
        training_args = TrainingArguments(
            output_dir=output_dir,
            learning_rate=learning_rate,
            per_device_train_batch_size=batch_size,
            per_device_eval_batch_size=batch_size,
            num_train_epochs=num_epochs,
            weight_decay=0.01,
            eval_strategy="epoch",
            save_strategy="epoch",
            load_best_model_at_end=True,
            metric_for_best_model="f1",
            greater_is_better=True,
            push_to_hub=False,
            logging_dir=f"{output_dir}/logs",
            logging_steps=10,
            save_total_limit=2,
            seed=42,
            # Original model specific parameters
            gradient_accumulation_steps=1,
            optim="adamw_torch",  # AdamW with default betas=(0.9, 0.999), epsilon=1e-08
        )
        
        # Initialize trainer
        self.trainer = Trainer(
            model=self.model,
            args=training_args,
            train_dataset=self.tokenized_datasets["train"],
            eval_dataset=self.tokenized_datasets["test"] if "test" in self.tokenized_datasets else self.tokenized_datasets["validation"],
            tokenizer=self.tokenizer,
            data_collator=data_collator,
            compute_metrics=self.compute_metrics
        )
        
        print("Trainer setup completed!")
        
    def train_model(self):
        """Train the model"""
        print("Starting training...")
        
        # Train the model
        train_result = self.trainer.train()
        
        print("Training completed!")
        print(f"Training loss: {train_result.training_loss}")
        
        # Save the model
        self.trainer.save_model()
        self.tokenizer.save_pretrained(self.trainer.args.output_dir)
        
        print(f"Model saved to: {self.trainer.args.output_dir}")
        
        return train_result
    
    def evaluate_model(self):
        """Evaluate the model"""
        print("Evaluating model...")
        
        # Evaluate on test set
        eval_results = self.trainer.evaluate()
        
        print("Evaluation results:")
        for key, value in eval_results.items():
            print(f"{key}: {value:.4f}")
        
        # Get predictions for detailed analysis
        predictions = self.trainer.predict(self.tokenized_datasets["test"] if "test" in self.tokenized_datasets else self.tokenized_datasets["validation"])
        
        y_pred = np.argmax(predictions.predictions, axis=1)
        y_true = predictions.label_ids
        
        # Print classification report
        print("\nClassification Report:")
        print(classification_report(y_true, y_pred))
        
        return eval_results, y_pred, y_true
    
    def plot_training_history(self):
        """Plot training history"""
        if hasattr(self.trainer, 'state') and hasattr(self.trainer.state, 'log_history'):
            logs = self.trainer.state.log_history
            
            # Extract training and validation metrics
            train_loss = [log['train_loss'] for log in logs if 'train_loss' in log]
            eval_loss = [log['eval_loss'] for log in logs if 'eval_loss' in log]
            eval_f1 = [log['eval_f1'] for log in logs if 'eval_f1' in log]
            
            # Create plots
            fig, axes = plt.subplots(1, 3, figsize=(15, 5))
            
            # Training loss
            axes[0].plot(train_loss, label='Training Loss')
            axes[0].set_title('Training Loss')
            axes[0].set_xlabel('Steps')
            axes[0].set_ylabel('Loss')
            axes[0].legend()
            
            # Evaluation loss
            axes[1].plot(eval_loss, label='Evaluation Loss')
            axes[1].set_title('Evaluation Loss')
            axes[1].set_xlabel('Epoch')
            axes[1].set_ylabel('Loss')
            axes[1].legend()
            
            # Evaluation F1
            axes[2].plot(eval_f1, label='Evaluation F1')
            axes[2].set_title('Evaluation F1 Score')
            axes[2].set_xlabel('Epoch')
            axes[2].set_ylabel('F1 Score')
            axes[2].legend()
            
            plt.tight_layout()
            plt.savefig('training_history.png', dpi=300, bbox_inches='tight')
            plt.show()
            print("Training history plots saved as 'training_history.png'")
    
    def plot_confusion_matrix(self, y_true, y_pred):
        """Plot confusion matrix"""
        from sklearn.metrics import confusion_matrix
        
        cm = confusion_matrix(y_true, y_pred)
        
        plt.figure(figsize=(8, 6))
        sns.heatmap(cm, annot=True, fmt='d', cmap='Blues')
        plt.title('Confusion Matrix')
        plt.xlabel('Predicted')
        plt.ylabel('Actual')
        plt.savefig('confusion_matrix.png', dpi=300, bbox_inches='tight')
        plt.show()
        print("Confusion matrix saved as 'confusion_matrix.png'")

    def create_sample_dataset(self):
        """Create a sample Vietnamese sentiment dataset for demonstration"""
        print("Creating sample Vietnamese sentiment dataset...")

        # Sample Vietnamese texts with sentiment labels
        sample_data = {
            "text": [
                # Positive samples
                "Giảng viên dạy rất hay và tâm huyết, tôi học được nhiều kiến thức bổ ích.",
                "Môn học này rất thú vị và practical, giúp tôi áp dụng được vào thực tế.",
                "Thầy cô rất tận tình và hỗ trợ sinh viên, không khí lớp học rất tích cực.",
                "Nội dung môn học sâu sắc, cách truyền đạt dễ hiểu, tôi rất hài lòng.",
                "Phương pháp giảng dạy mới mẻ, hấp dẫn, khiến tôi say mê học tập.",

                # Negative samples
                "Môn học quá khó và nhàm chán, không có gì để học cả.",
                "Giảng viên dạy không rõ ràng, tốc độ quá nhanh, không theo kịp.",
                "Thời lượng quá ít nhưng nội dung nhiều, không thể học hết.",
                "Thầy cô ít quan tâm đến sinh viên, không giải thích khi có thắc mắc.",
                "Đồ án quá nặng, yêu cầu không rõ ràng, deadline quá gấp.",

                # Neutral samples
                "Môn học ổn định, không có gì đặc biệt để nhận xét.",
                "Nội dung cơ bản, phù hợp với chương trình đề ra.",
                "Lớp học bình thường, giảng viên dạy đúng theo giáo trình.",
                "Đủ kiến thức cơ bản, không quá khó cũng không quá dễ.",
                "Môn học như các môn khác, không có gì nổi bật.",

                # Additional samples
                "Tôi rất thích cách thầy cô tổ chức hoạt động nhóm, rất hiệu quả.",
                "Phòng học quá nóng, thiết bị cũ, ảnh hưởng đến việc học.",
                "Tài liệu học tập đầy đủ, có cả online và offline.",
                "Bài tập nhiều nhưng không quá khó, giúp củng cố kiến thức.",
                "Lịch học ổn, không trùng với môn học quan trọng khác."
            ],
            "label": [
                # Labels: 0 = Negative, 1 = Neutral, 2 = Positive
                2, 2, 2, 2, 2,  # Positive (5 samples)
                0, 0, 0, 0, 0,  # Negative (5 samples)
                1, 1, 1, 1, 1,  # Neutral (5 samples)
                2, 0, 1, 1, 1   # Additional mixed (5 samples)
            ]
        }

        from datasets import Dataset

        # Create dataset
        full_dataset = Dataset.from_dict(sample_data)

        # Split dataset
        train_test_split = full_dataset.train_test_split(test_size=0.2, seed=42)
        train_val_split = train_test_split["train"].train_test_split(test_size=0.25, seed=42)

        self.dataset = DatasetDict({
            "train": train_val_split["train"],
            "validation": train_val_split["test"],
            "test": train_test_split["test"]
        })

        print(f"Created sample dataset with {len(self.dataset['train'])} training, {len(self.dataset['validation'])} validation, and {len(self.dataset['test'])} test samples")

        # Print distribution
        train_df = pd.DataFrame(self.dataset['train'])
        print("\nSentiment distribution in training set:")
        label_counts = train_df['label'].value_counts().sort_index()
        for label, count in label_counts.items():
            sentiment_name = ["Negative", "Neutral", "Positive"][label]
            print(f"  {sentiment_name} (label {label}): {count} samples")

    def run_fine_tuning(self, output_dir="./fine_tuned_sentiment_model", learning_rate=2e-5, batch_size=16, num_epochs=5):
        """Run the complete fine-tuning pipeline"""
        print("=" * 60)
        print("VIETNAMESE SENTIMENT ANALYSIS FINE-TUNING")
        print("=" * 60)
        
        # Load model and tokenizer
        self.load_model_and_tokenizer()
        
        # Load and prepare dataset
        self.load_and_prepare_dataset()
        
        # Tokenize datasets
        self.tokenize_datasets()
        
        # Setup trainer
        self.setup_trainer(output_dir, learning_rate, batch_size, num_epochs)
        
        # Train model
        train_result = self.train_model()
        
        # Evaluate model
        eval_results, y_pred, y_true = self.evaluate_model()
        
        # Plot results
        self.plot_training_history()
        self.plot_confusion_matrix(y_true, y_pred)
        
        print("=" * 60)
        print("FINE-TUNING COMPLETED SUCCESSFULLY!")
        print("=" * 60)
        print(f"Model saved to: {output_dir}")
        print(f"Final evaluation F1: {eval_results['eval_f1']:.4f}")
        print(f"Final evaluation accuracy: {eval_results['eval_accuracy']:.4f}")
        
        return train_result, eval_results

def main():
    """Main function to run the fine-tuning"""
    # Initialize the fine-tuner
    fine_tuner = SentimentFineTuner()
    
    # Run fine-tuning (matching original model configuration)
    train_result, eval_results = fine_tuner.run_fine_tuning(
        output_dir="./vietnamese_sentiment_finetuned",
        learning_rate=2e-5,  # Same as original model
        batch_size=16,        # Recommended batch size
        num_epochs=5          # Same as original model
    )
    
    print("Fine-tuning completed successfully!")

if __name__ == "__main__":
    main()