model.py

from google.colab import drive
import os
import glob
import torch
import torch.nn as nn
import torch.optim as optim
import pdfplumber
import random
import math
from tqdm import tqdm
from transformers import AutoTokenizer
from torch.utils.data import DataLoader, Dataset, random_split
from torch.cuda.amp import autocast, GradScaler  # Fixed import
from huggingface_hub import login
from torch.utils.tensorboard import SummaryWriter
import logging
from typing import Tuple, List, Dict

# Configuration
class Config:
    # Model
    D_MODEL = 512
    NHEAD = 8
    ENC_LAYERS = 6
    DEC_LAYERS = 6
    DIM_FEEDFORWARD = 2048
    DROPOUT = 0.1
    
    # Training
    BATCH_SIZE = 4
    GRAD_ACCUM_STEPS = 2
    LR = 1e-4
    EPOCHS = 20
    MAX_GRAD_NORM = 1.0
    
    # Data
    INPUT_MAX_LEN = 512
    SUMMARY_MAX_LEN = 128
    CHUNK_SIZE = 512
    
    # Paths
    CHECKPOINT_DIR = "/content/drive/MyDrive/legal_summarization_checkpoints_6"
    LOG_DIR = os.path.join(CHECKPOINT_DIR, "logs")
    
    @classmethod
    def setup_paths(cls):
        os.makedirs(cls.CHECKPOINT_DIR, exist_ok=True)
        os.makedirs(cls.LOG_DIR, exist_ok=True)

# Initialize config
Config.setup_paths()

# Setup logging
logging.basicConfig(
    level=logging.INFO,
    format='%(asctime)s - %(levelname)s - %(message)s',
    handlers=[
        logging.FileHandler(os.path.join(Config.LOG_DIR, 'training.log')),
        logging.StreamHandler()
    ]
)
logger = logging.getLogger(_name_)

# Authenticate Hugging Face
login(token="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")

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Mount Google Drive
drive.mount('/content/drive', force_remount=True)

# Tokenizer
tokenizer = AutoTokenizer.from_pretrained("t5-small")
vocab_size = tokenizer.vocab_size

# TensorBoard
writer = SummaryWriter(Config.LOG_DIR)

def clean_text(text: str) -> str:
    """Basic text cleaning"""
    text = ' '.join(text.split())  # Remove extra whitespace
    return text.strip()

def extract_text_from_pdf(pdf_path: str, chunk_size: int = Config.CHUNK_SIZE) -> List[str]:
    """Extract and chunk text from PDF with error handling"""
    text = ''
    try:
        with pdfplumber.open(pdf_path) as pdf:
            for page in pdf.pages:
                page_text = page.extract_text() or ''
                text += page_text + ' '
    except Exception as e:
        logger.warning(f"Error processing {pdf_path}: {str(e)}")
        return []
    
    text = clean_text(text)
    words = text.split()
    return [' '.join(words[i:i+chunk_size]) for i in range(0, len(words), chunk_size)] if words else []

def load_texts_from_folder(folder_path: str, chunk_size: int = Config.CHUNK_SIZE) -> List[str]:
    """Load and chunk texts from folder with multiple file types"""
    texts = []
    for fname in sorted(os.listdir(folder_path)):
        path = os.path.join(folder_path, fname)
        try:
            if path.endswith('.pdf'):
                chunks = extract_text_from_pdf(path, chunk_size)
                if chunks:
                    texts.extend(chunks)
            else:
                with open(path, 'r', encoding='utf-8', errors='ignore') as f:
                    content = clean_text(f.read())
                    if content:
                        texts.extend([content[i:i+chunk_size] for i in range(0, len(content), chunk_size)])
        except Exception as e:
            logger.warning(f"Error loading {path}: {str(e)}")
            continue
    return texts

class LegalDataset(Dataset):
    def _init_(self, texts: List[str], summaries: List[str], tokenizer: AutoTokenizer, 

                 input_max_len: int = Config.INPUT_MAX_LEN, 

                 summary_max_len: int = Config.SUMMARY_MAX_LEN):
        assert len(texts) == len(summaries), "Texts and summaries must be same length"
        self.texts = texts
        self.summaries = summaries
        self.tokenizer = tokenizer
        self.input_max_len = input_max_len
        self.summary_max_len = summary_max_len

    def _len_(self):
        return len(self.texts)

    def _getitem_(self, idx):
        src = self.texts[idx]
        tgt = self.summaries[idx]
        
        enc = self.tokenizer(
            src, 
            padding='max_length', 
            truncation=True, 
            max_length=self.input_max_len, 
            return_tensors='pt'
        )
        
        dec = self.tokenizer(
            tgt, 
            padding='max_length', 
            truncation=True, 
            max_length=self.summary_max_len, 
            return_tensors='pt'
        )
        
        return {
            'input_ids': enc.input_ids.squeeze(),
            'attention_mask': enc.attention_mask.squeeze(),
            'labels': dec.input_ids.squeeze()
        }

class PositionalEncoding(nn.Module):
    def _init_(self, d_model: int, dropout: float = 0.1, max_len: int = 1024):
        super()._init_()
        self.dropout = nn.Dropout(dropout)
        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len).unsqueeze(1).float()
        div = torch.exp(torch.arange(0, d_model, 2).float() * -(math.log(10000.0) / d_model))
        pe[:, 0::2] = torch.sin(position * div)
        pe[:, 1::2] = torch.cos(position * div)
        self.register_buffer('pe', pe.unsqueeze(0))

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = x + self.pe[:, :x.size(1)]
        return self.dropout(x)

class CustomTransformer(nn.Module):
    def _init_(self, vocab_size: int, d_model: int = Config.D_MODEL, nhead: int = Config.NHEAD, 

                 enc_layers: int = Config.ENC_LAYERS, dec_layers: int = Config.DEC_LAYERS, 

                 dim_feedforward: int = Config.DIM_FEEDFORWARD, dropout: float = Config.DROPOUT):
        super()._init_()
        self.embed = nn.Embedding(vocab_size, d_model)
        self.pos_enc = PositionalEncoding(d_model, dropout)
        self.transformer = nn.Transformer(
            d_model=d_model,
            nhead=nhead,
            num_encoder_layers=enc_layers,
            num_decoder_layers=dec_layers,
            dim_feedforward=dim_feedforward,
            dropout=dropout,
            batch_first=True
        )
        self.fc = nn.Linear(d_model, vocab_size)
        
        # Initialize weights
        self._init_weights()

    def _init_weights(self):
        for p in self.parameters():
            if p.dim() > 1:
                nn.init.xavier_uniform_(p)

    def forward(self, src_ids: torch.Tensor, tgt_ids: torch.Tensor, 

                src_key_padding_mask: torch.Tensor = None, 

                tgt_key_padding_mask: torch.Tensor = None) -> torch.Tensor:
        
        # Create causal mask for decoder
        tgt_mask = nn.Transformer.generate_square_subsequent_mask(tgt_ids.size(1)).to(tgt_ids.device)
        
        src = self.pos_enc(self.embed(src_ids))
        tgt = self.pos_enc(self.embed(tgt_ids))
        
        out = self.transformer(
            src, tgt,
            tgt_mask=tgt_mask,
            src_key_padding_mask=src_key_padding_mask,
            tgt_key_padding_mask=tgt_key_padding_mask,
            memory_key_padding_mask=src_key_padding_mask
        )
        return self.fc(out)

def create_masks(input_ids: torch.Tensor, decoder_input: torch.Tensor, pad_token_id: int) -> Tuple[torch.Tensor, torch.Tensor]:
    """Create padding masks for transformer"""
    src_pad_mask = (input_ids == pad_token_id)
    tgt_pad_mask = (decoder_input == pad_token_id)
    return src_pad_mask, tgt_pad_mask

def train_model(model: nn.Module, train_loader: DataLoader, val_loader: DataLoader, 

                optimizer: optim.Optimizer, criterion: nn.Module, device: torch.device, 

                epochs: int = Config.EPOCHS, grad_accum_steps: int = Config.GRAD_ACCUM_STEPS):
    
    model.to(device)
    scaler = GradScaler()
    scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', patience=2, factor=0.5)
    best_val_loss = float('inf')
    early_stop_counter = 0
    
    for epoch in range(1, epochs + 1):
        model.train()
        train_loss = 0
        progress_bar = tqdm(train_loader, desc=f"Epoch {epoch}/{epochs}")
        
        for step, batch in enumerate(progress_bar, 1):
            input_ids = batch['input_ids'].to(device)
            attn_mask = batch['attention_mask'].to(device)
            labels = batch['labels'].to(device)
            
            # Prepare decoder input with <pad> as start
            decoder_input = torch.cat([
                torch.full((labels.size(0), 1), tokenizer.pad_token_id, dtype=torch.long, device=device), 
                labels[:, :-1]
            ], dim=1)
            
            # Create masks
            src_pad_mask, tgt_pad_mask = create_masks(input_ids, decoder_input, tokenizer.pad_token_id)
            
            with autocast():
                outputs = model(
                    input_ids, 
                    decoder_input,
                    src_key_padding_mask=src_pad_mask,
                    tgt_key_padding_mask=tgt_pad_mask
                )
                loss = criterion(outputs.view(-1, outputs.size(-1)), labels.view(-1))
                loss = loss / grad_accum_steps
            
            scaler.scale(loss).backward()
            
            if step % grad_accum_steps == 0:
                scaler.unscale_(optimizer)
                nn.utils.clip_grad_norm_(model.parameters(), Config.MAX_GRAD_NORM)
                scaler.step(optimizer)
                scaler.update()
                optimizer.zero_grad()
            
            train_loss += loss.item() * grad_accum_steps
            progress_bar.set_postfix({'train_loss': f"{loss.item():.4f}"})
        
        avg_train_loss = train_loss / len(train_loader)
        writer.add_scalar('Loss/train', avg_train_loss, epoch)
        logger.info(f"Epoch {epoch} Train Loss: {avg_train_loss:.4f}")
        
        # Validation
        model.eval()
        val_loss = 0
        with torch.no_grad():
            for batch in tqdm(val_loader, desc="Validating"):
                input_ids = batch['input_ids'].to(device)
                labels = batch['labels'].to(device)
                decoder_input = torch.cat([
                    torch.full((labels.size(0), 1), tokenizer.pad_token_id, dtype=torch.long, device=device), 
                    labels[:, :-1]
                ], dim=1)
                
                src_pad_mask, tgt_pad_mask = create_masks(input_ids, decoder_input, tokenizer.pad_token_id)
                
                with autocast():
                    outputs = model(
                        input_ids, 
                        decoder_input,
                        src_key_padding_mask=src_pad_mask,
                        tgt_key_padding_mask=tgt_pad_mask
                    )
                    loss = criterion(outputs.view(-1, outputs.size(-1)), labels.view(-1))
                    val_loss += loss.item()
        
        avg_val_loss = val_loss / len(val_loader)
        writer.add_scalar('Loss/val', avg_val_loss, epoch)
        logger.info(f"Epoch {epoch} Val Loss: {avg_val_loss:.4f}")
        
        # Learning rate scheduling
        scheduler.step(avg_val_loss)
        
        # Early stopping & checkpointing
        if avg_val_loss < best_val_loss:
            best_val_loss = avg_val_loss
            early_stop_counter = 0
            # Save best model
            ckpt_path = os.path.join(Config.CHECKPOINT_DIR, f"transformer_best.pt")
            torch.save(model.state_dict(), ckpt_path)
            logger.info(f"New best model saved with val loss: {best_val_loss:.4f}")
        else:
            early_stop_counter += 1
            if early_stop_counter >= 3:
                logger.info("Early stopping triggered")
                break
        
        # Save regular checkpoint
        ckpt_path = os.path.join(Config.CHECKPOINT_DIR, f"transformer_epoch_{epoch}.pt")
        torch.save(model.state_dict(), ckpt_path)
        
        # Keep only latest 2 checkpoints
        manage_checkpoints()

def manage_checkpoints():
    """Keep only the 2 most recent checkpoints"""
    files = sorted(glob.glob(os.path.join(Config.CHECKPOINT_DIR, "transformer_epoch_*.pt")), key=os.path.getctime)
    if len(files) > 2:
        for old in files[:-2]:
            os.remove(old)
            logger.info(f"Removed old checkpoint: {old}")

if _name_ == "_main_":
    try:
        logger.info("Starting training process")
        
        # Load data
        logger.info("Loading texts and summaries")
        texts = load_texts_from_folder("/content/drive/MyDrive/dataset/IN-Abs/train-data/judgement")
        sums = load_texts_from_folder("/content/drive/MyDrive/dataset/IN-Abs/train-data/summary")
        
        if not texts or not sums:
            raise ValueError("No data loaded - check your input paths and files")
        
        logger.info(f"Loaded {len(texts)} text chunks and {len(sums)} summary chunks")
        
        # Create dataset
        full_ds = LegalDataset(texts, sums, tokenizer)
        
        # Train/val split
        val_size = int(0.1 * len(full_ds))
        train_size = len(full_ds) - val_size
        train_ds, val_ds = random_split(full_ds, [train_size, val_size])
        
        train_loader = DataLoader(train_ds, batch_size=Config.BATCH_SIZE, shuffle=True)
        val_loader = DataLoader(val_ds, batch_size=Config.BATCH_SIZE)
        
        # Initialize model
        model = CustomTransformer(vocab_size)
        optimizer = optim.Adam(model.parameters(), lr=Config.LR)
        criterion = nn.CrossEntropyLoss(ignore_index=tokenizer.pad_token_id)
        
        # Train
        train_model(model, train_loader, val_loader, optimizer, criterion, device)
        
        logger.info("Training completed successfully")
        
    except Exception as e:
        logger.error(f"Training failed: {str(e)}", exc_info=True)
        raise