scripts/core/training/train.py

import argparse
import os
import torch
from torch.utils.data import DataLoader, Dataset
from transformers import AutoTokenizer

from scripts.core.training.model import CodeEmbedder
from scripts.core.training.trainer import CodeTrainer

import json

# Real Dataset class for Triplet Training
class RealCodeDataset(Dataset):
    def __init__(self, jsonl_path, tokenizer, max_length=512):
        self.tokenizer = tokenizer
        self.max_length = max_length
        self.data = []
        
        print(f"Loading data from {jsonl_path}...")
        with open(jsonl_path, 'r', encoding='utf-8') as f:
            for line in f:
                if line.strip():
                    self.data.append(json.loads(line))
        print(f"Loaded {len(self.data)} triplets.")

    def __len__(self):
        return len(self.data)

    def __getitem__(self, idx):
        item = self.data[idx]
        
        # Helper to tokenize
        def tokenize_text(text):
            return self.tokenizer(
                text,
                return_tensors='pt',
                padding='max_length',
                truncation=True,
                max_length=self.max_length
            )
        
        # Tokenize all three parts
        anchor = tokenize_text(item['anchor'])
        positive = tokenize_text(item['positive'])
        negative = tokenize_text(item['negative'])
        
        # Return a flat dict with prefixed keys
        return {
            'anchor_input_ids': anchor['input_ids'].squeeze(0),
            'anchor_attention_mask': anchor['attention_mask'].squeeze(0),
            'positive_input_ids': positive['input_ids'].squeeze(0),
            'positive_attention_mask': positive['attention_mask'].squeeze(0),
            'negative_input_ids': negative['input_ids'].squeeze(0),
            'negative_attention_mask': negative['attention_mask'].squeeze(0)
        }

# Dummy Dataset class for MVP testing without the robust data pipeline availability
class DummyCodeDataset(Dataset):
    def __init__(self, tokenizer, size=100):
        self.tokenizer = tokenizer
        self.size = size
        # Generate dummy triplet structure
        self.data = [{"anchor": "def hello(): return 'world'", "positive": "def hi(): return 'earth'", "negative": "class Foo: pass"}] * size

    def __len__(self):
        return self.size

    def __getitem__(self, idx):
        item = self.data[idx]
        
        # Helper to tokenize
        def tokenize_text(text):
            return self.tokenizer(
                text,
                return_tensors='pt',
                padding='max_length',
                truncation=True,
                max_length=128
            )
            
        anchor = tokenize_text(item['anchor'])
        positive = tokenize_text(item['positive'])
        negative = tokenize_text(item['negative'])

        return {
            'anchor_input_ids': anchor['input_ids'].squeeze(0),
            'anchor_attention_mask': anchor['attention_mask'].squeeze(0),
            'positive_input_ids': positive['input_ids'].squeeze(0),
            'positive_attention_mask': positive['attention_mask'].squeeze(0),
            'negative_input_ids': negative['input_ids'].squeeze(0),
            'negative_attention_mask': negative['attention_mask'].squeeze(0)
        }

def main():
    parser = argparse.ArgumentParser(description="Train CodeMode Embeddings")
    
    parser.add_argument("--model_name", type=str, default="microsoft/codebert-base", help="Hub model name")
    parser.add_argument("--data_path", type=str, required=False, help="Path to parsed chunks.jsonl")
    parser.add_argument("--output_dir", type=str, default="./output", help="Where to save checkpoints")
    parser.add_argument("--epochs", type=int, default=3)
    parser.add_argument("--batch_size", type=int, default=8)
    parser.add_argument("--accumulation_steps", type=int, default=4, help="Gradient Accumulation Steps")
    parser.add_argument("--lr", type=float, default=2e-5)
    parser.add_argument("--dry_run", action="store_true", help="Run with dummy data for 1 epoch")

    args = parser.parse_args()
    
    print(f"Initializing Training Pipeline...")
    print(f"   Model: {args.model_name}")
    print(f"   Output: {args.output_dir}")
    print(f"   Device: {'cuda' if torch.cuda.is_available() else 'cpu'}")

    # 1. Initialize Tokenizer
    tokenizer = AutoTokenizer.from_pretrained(args.model_name)

    # 2. Load Dataset (Real or Dummy)
    if args.data_path and os.path.exists(args.data_path):
        train_dataset = RealCodeDataset(args.data_path, tokenizer)
    else:
        print("No data path provided or file missing. Using DUMMY data for verification.")
        train_dataset = DummyCodeDataset(tokenizer, size=100)

    train_loader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True)

    # 3. Initialize Model
    model = CodeEmbedder(model_name_or_path=args.model_name)

    # 4. Initialize Trainer
    trainer = CodeTrainer(
        model=model,
        train_loader=train_loader,
        epochs=args.epochs,
        learning_rate=args.lr,
        accumulation_steps=args.accumulation_steps,
        mixed_precision=True, # Hardcoded True for the "Zero-Cost" philosophy
        output_dir=args.output_dir
    )

    # 5. Connect and Train
    trainer.train()
    
    print("Training Complete.")

if __name__ == "__main__":
    main()