the-jashthakkar/CodeMode

1import torch2import torch.nn as nn3from torch.optim import AdamW4from torch.utils.data import DataLoader5from tqdm import tqdm6import os7import logging8from .model import CodeEmbedder9 10# Setup Logger11logging.basicConfig(level=logging.INFO)12logger = logging.getLogger(__name__)13 14class CodeTrainer:15    def __init__(16        self,17        model: CodeEmbedder,18        train_loader: DataLoader,19        val_loader: DataLoader = None,20        epochs: int = 3,21        learning_rate: float = 2e-5,22        accumulation_steps: int = 1,23        mixed_precision: bool = True,24        output_dir: str = "./output",25        device: str = "cuda" if torch.cuda.is_available() else "cpu"26    ):27        self.model = model.to(device)28        self.train_loader = train_loader29        self.val_loader = val_loader30        self.epochs = epochs31        self.lr = learning_rate32        self.accumulation_steps = accumulation_steps33        self.mixed_precision = mixed_precision34        self.output_dir = output_dir35        self.device = device36        37        # Optimizer38        self.optimizer = AdamW(self.model.parameters(), lr=self.lr)39        40        # Scheduler (Optional: constant for now, can transform to Linear later)41        # self.scheduler = ...42        43        # Mixed Precision Scaler44        self.scaler = torch.cuda.amp.GradScaler(enabled=self.mixed_precision)45        46        # Loss Function: Triplet Margin Loss (Standard for Sentence Embeddings)47        # Tries to maximize distance between Anchor-Negative and minimize Anchor-Positive48        self.criterion = nn.TripletMarginLoss(margin=1.0, p=2)49 50    def train_step(self, batch):51        """52        Runs one training step. Returns loss.53        """54        # Unpack the Triplet Batch55        # We assume the Dataset returns keys: 'anchor_input_ids', 'anchor_attention_mask', etc.56        57        # Helper to move dict to device58        to_device = lambda x: x.to(self.device)59        60        # Autocast for Mixed Precision61        with torch.cuda.amp.autocast(enabled=self.mixed_precision):62            # 1. Forward Pass for all 3 components63            anchor_emb = self.model(to_device(batch['anchor_input_ids']), to_device(batch['anchor_attention_mask']))64            positive_emb = self.model(to_device(batch['positive_input_ids']), to_device(batch['positive_attention_mask']))65            negative_emb = self.model(to_device(batch['negative_input_ids']), to_device(batch['negative_attention_mask']))66            67            # 2. Compute Triplet Loss68            loss = self.criterion(anchor_emb, positive_emb, negative_emb)69            70        return loss71 72    def train(self):73        logger.info(f"Starting training on {self.device}...")74        logger.info(f"Batch Size: {self.train_loader.batch_size}, Accumulation Steps: {self.accumulation_steps}")75        logger.info(f"Effective Batch Size: {self.train_loader.batch_size * self.accumulation_steps}")76        77        self.model.train()78        79        for epoch in range(self.epochs):80            total_loss = 081            self.optimizer.zero_grad()82            83            progress_bar = tqdm(self.train_loader, desc=f"Epoch {epoch+1}/{self.epochs}")84            85            for step, batch in enumerate(progress_bar):86                87                # Forward + Loss Calculation88                loss = self.train_step(batch)89                90                # Gradient Accumulation: Normalize loss91                loss = loss / self.accumulation_steps92                93                # Backward Pass (Scaled)94                self.scaler.scale(loss).backward()95                96                if (step + 1) % self.accumulation_steps == 0:97                    # Update Weights98                    self.scaler.step(self.optimizer)99                    self.scaler.update()100                    self.optimizer.zero_grad()101                102                total_loss += loss.item() * self.accumulation_steps103                progress_bar.set_postfix({'loss': total_loss / (step + 1)})104            105            # Save Checkpoint106            self.save_model(epoch+1)107            108    def save_model(self, epoch):109        save_path = os.path.join(self.output_dir, f"checkpoint-{epoch}")110        os.makedirs(save_path, exist_ok=True)111        112        logger.info(f"Saving model to {save_path}...")113        114        # Save explicitly as safetensors via transformers API115        self.model.encoder.save_pretrained(save_path, safe_serialization=True)116        self.model.config.save_pretrained(save_path)117        # Note: We save the 'encoder' which is the AutoModel, 118        # so it can be loaded easily by others.119