train_gpu.py

"""
GPU Training Script for MARS on MovieLens-1M.

Trains MARS (innovative method) and SASRec (baseline) for comparison.
Pushes results to HF Hub.
"""

import os
import sys
import time
import json
import random
import numpy as np
import torch
import torch.nn as nn
from torch.optim import AdamW
from torch.optim.lr_scheduler import CosineAnnealingLR

# ===== Setup =====
def set_seed(seed=42):
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(seed)

set_seed(42)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"Device: {device}")
if torch.cuda.is_available():
    print(f"GPU: {torch.cuda.get_device_name(0)}")
    print(f"GPU Memory: {torch.cuda.get_device_properties(0).total_mem / 1e9:.1f} GB")

# ===== Config =====
CONFIGS = {
    'mars': {
        'embed_dim': 64,
        'max_seq_len': 512,
        'short_term_len': 50,
        'num_memory_tokens': 8,
        'num_tadn_layers': 3,
        'num_attn_layers': 2,
        'num_heads': 2,
        'state_dim': 64,
        'dropout': 0.1,
        'batch_size': 128,
        'lr': 1e-3,
        'weight_decay': 0.01,
        'epochs': 80,
        'num_negatives': 4,
    },
    'sasrec': {
        'embed_dim': 64,
        'max_seq_len': 200,
        'num_heads': 2,
        'num_layers': 2,
        'dropout': 0.1,
        'batch_size': 256,
        'lr': 1e-3,
        'weight_decay': 0.0,
        'epochs': 80,
        'num_negatives': 4,
    },
}

# ===== Import our modules =====
from model import MARS, SASRecBaseline
from data import load_movielens_1m, ReindexedData, create_dataloaders, save_data_config
from evaluate import evaluate_model, print_comparison

# ===== Initialize Trackio =====
try:
    import trackio
    trackio.init(name="MARS-SeqRec-ML1M", project="mars-seqrec")
    use_trackio = True
    print("Trackio initialized successfully")
except Exception as e:
    print(f"Trackio not available: {e}")
    use_trackio = False

# ===== Load MovieLens-1M =====
print("\n" + "="*60)
print("Loading MovieLens-1M Dataset")
print("="*60)

sequences = load_movielens_1m(min_interactions=5)
print(f"Loaded {len(sequences)} user sequences")

# Compute sequence length distribution
seq_lens = [len(v['item_ids']) for v in sequences.values()]
print(f"Sequence length: mean={np.mean(seq_lens):.1f}, median={np.median(seq_lens):.1f}, "
      f"max={np.max(seq_lens)}, p90={np.percentile(seq_lens, 90):.0f}")
print(f"Users with 100+ interactions: {sum(1 for l in seq_lens if l >= 100)}")
print(f"Users with 200+ interactions: {sum(1 for l in seq_lens if l >= 200)}")
print(f"Users with 500+ interactions: {sum(1 for l in seq_lens if l >= 500)}")


def train_model(model_name, config, sequences, device):
    """Train a model with given config and return results."""
    print(f"\n{'='*60}")
    print(f"Training: {model_name.upper()}")
    print(f"{'='*60}")
    
    max_seq_len = config['max_seq_len']
    batch_size = config['batch_size']
    epochs = config['epochs']
    
    # Prepare data
    data = ReindexedData(sequences, max_seq_len=max_seq_len)
    train_loader, val_loader, test_loader = create_dataloaders(
        data, max_seq_len=max_seq_len,
        batch_size=batch_size,
        num_negatives=config['num_negatives'],
        num_workers=4,
    )
    
    # Create model
    if model_name == 'mars':
        model = MARS(
            num_items=data.num_items,
            embed_dim=config['embed_dim'],
            max_seq_len=max_seq_len,
            short_term_len=config['short_term_len'],
            num_memory_tokens=config['num_memory_tokens'],
            num_tadn_layers=config['num_tadn_layers'],
            num_attn_layers=config['num_attn_layers'],
            num_heads=config['num_heads'],
            state_dim=config['state_dim'],
            dropout=config['dropout'],
        )
    else:
        model = SASRecBaseline(
            num_items=data.num_items,
            embed_dim=config['embed_dim'],
            max_seq_len=config['max_seq_len'],
            num_heads=config['num_heads'],
            num_layers=config['num_layers'],
            dropout=config['dropout'],
        )
    
    model = model.to(device)
    num_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
    print(f"Parameters: {num_params:,}")
    
    # Optimizer & scheduler
    optimizer = AdamW(
        model.parameters(),
        lr=config['lr'],
        weight_decay=config['weight_decay'],
    )
    scheduler = CosineAnnealingLR(optimizer, T_max=epochs, eta_min=config['lr'] * 0.01)
    
    # Training loop
    best_val_hr10 = 0
    best_epoch = 0
    best_state = None
    
    save_dir = f'./checkpoints/{model_name}'
    os.makedirs(save_dir, exist_ok=True)
    
    for epoch in range(1, epochs + 1):
        model.train()
        total_loss = 0
        num_batches = 0
        t0 = time.time()
        
        for batch_idx, batch in enumerate(train_loader):
            batch = {k: v.to(device) for k, v in batch.items()}
            
            optimizer.zero_grad()
            loss = model(batch)
            loss.backward()
            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
            optimizer.step()
            
            total_loss += loss.item()
            num_batches += 1
        
        scheduler.step()
        avg_loss = total_loss / num_batches
        epoch_time = time.time() - t0
        lr = scheduler.get_last_lr()[0]
        
        print(f"Epoch {epoch:3d}/{epochs} | Loss: {avg_loss:.4f} | "
              f"LR: {lr:.6f} | Time: {epoch_time:.1f}s")
        
        if use_trackio:
            trackio.log({
                f"{model_name}/train_loss": avg_loss,
                f"{model_name}/lr": lr,
                f"{model_name}/epoch_time": epoch_time,
                "epoch": epoch,
            })
        
        # Evaluate every 10 epochs or at end
        if epoch % 10 == 0 or epoch == epochs or epoch <= 5:
            metrics = evaluate_model(
                model, val_loader, data.num_items, device,
                ks=[5, 10, 20, 50]
            )
            
            print(f"  Val | HR@10={metrics['HR@10']:.4f} NDCG@10={metrics['NDCG@10']:.4f} "
                  f"MRR@10={metrics['MRR@10']:.4f}")
            
            if use_trackio:
                trackio.log({f"{model_name}/val_{k}": v for k, v in metrics.items() if k != 'eval_time'})
                trackio.log({"epoch": epoch})
            
            hr10 = metrics['HR@10']
            if hr10 > best_val_hr10:
                best_val_hr10 = hr10
                best_epoch = epoch
                best_state = {k: v.cpu().clone() for k, v in model.state_dict().items()}
                print(f"  ✓ New best! HR@10={hr10:.4f}")
    
    # Load best model and do final test evaluation
    model.load_state_dict(best_state)
    model = model.to(device)
    
    test_metrics = evaluate_model(
        model, test_loader, data.num_items, device,
        ks=[5, 10, 20, 50]
    )
    
    print(f"\nFinal Test Results ({model_name.upper()}, best epoch {best_epoch}):")
    for k, v in test_metrics.items():
        if k != 'eval_time':
            print(f"  {k}: {v:.4f}")
    
    if use_trackio:
        trackio.log({f"{model_name}/test_{k}": v for k, v in test_metrics.items() if k != 'eval_time'})
    
    # Save checkpoint
    torch.save({
        'model_state_dict': best_state,
        'config': config,
        'test_metrics': test_metrics,
        'best_epoch': best_epoch,
        'num_items': data.num_items,
    }, os.path.join(save_dir, 'best_model.pt'))
    
    return test_metrics, config, num_params


# ===== Train both models =====
print("\n" + "="*60)
print("PHASE 1: Training SASRec Baseline")
print("="*60)

sasrec_metrics, sasrec_config, sasrec_params = train_model(
    'sasrec', CONFIGS['sasrec'], sequences, device
)

print("\n" + "="*60)
print("PHASE 2: Training MARS (our method)")
print("="*60)

mars_metrics, mars_config, mars_params = train_model(
    'mars', CONFIGS['mars'], sequences, device
)

# ===== Comparison =====
print_comparison(mars_metrics, sasrec_metrics, ks=[5, 10, 20, 50])

# ===== Save final results =====
final_results = {
    'mars': {
        'metrics': mars_metrics,
        'config': mars_config,
        'params': mars_params,
    },
    'sasrec': {
        'metrics': sasrec_metrics,
        'config': sasrec_config,
        'params': sasrec_params,
    },
    'dataset': 'MovieLens-1M',
    'num_users': len(sequences),
}

with open('./checkpoints/final_results.json', 'w') as f:
    json.dump(final_results, f, indent=2, default=str)

print("\n✓ All training complete! Results saved to ./checkpoints/")

# ===== Push to Hub =====
try:
    from huggingface_hub import HfApi, upload_folder
    hub_model_id = os.environ.get('HF_HUB_MODEL_ID', 'CyberDancer/MARS-SeqRec')
    api = HfApi()
    api.create_repo(hub_model_id, exist_ok=True)
    
    # Create README
    readme = f"""# MARS: Multi-scale Adaptive Recurrence with State compression

An innovative method for **super long sequence modeling** in sequential recommendation.

## Key Innovations

1. **Temporal-Aware Delta Network (TADN)** — O(n) linear complexity with explicit temporal decay gating
2. **Compressive Memory Tokens** — Fixed-size learnable memory as information bottleneck  
3. **Dual-Branch Architecture** — Long-term (TADN) + Short-term (Self-Attention) with adaptive fusion
4. **Multi-Scale Temporal Encoding** — Captures daily/weekly/seasonal patterns

## Architecture

```
Input: Full user interaction sequence + timestamps
    |
    v
[Item Embedding + Multi-Scale Temporal Encoding]
    |
    +---- Long-term Branch (TADN layers, O(n) complexity)
    |         |
    |     [Compressive Memory] → fixed-size memory tokens
    |         |
    +---- Short-term Branch (Causal Self-Attention, recent K items)
    |
    v
[Adaptive Fusion Gate (per-user learned)]
    |
    v
[Prediction Head] → next item scores
```

## Results on MovieLens-1M

| Model | Params | Max Seq | HR@10 | NDCG@10 | MRR@10 |
|-------|--------|---------|-------|---------|--------|
| SASRec (baseline) | {sasrec_params:,} | {sasrec_config['max_seq_len']} | {sasrec_metrics.get('HR@10', 0):.4f} | {sasrec_metrics.get('NDCG@10', 0):.4f} | {sasrec_metrics.get('MRR@10', 0):.4f} |
| **MARS (ours)** | {mars_params:,} | {mars_config['max_seq_len']} | {mars_metrics.get('HR@10', 0):.4f} | {mars_metrics.get('NDCG@10', 0):.4f} | {mars_metrics.get('MRR@10', 0):.4f} |

## Method Details

### Temporal-Aware Delta Network (TADN)
The TADN layer maintains a state matrix S that is updated with a delta rule incorporating temporal decay:

```
S_t = S_{{t-1}} * (1 - g_t ⊙ β_t ⊗ k_t) + β_t ⊗ v_t ⊗ k_t
g_t = α * σ(W_g * [h_t; Δh_t]) * τ_t + (1-α) * g_static
τ_t = exp(-(t_current - t_behavior) / T)
```

This gives O(n) complexity for processing arbitrarily long sequences, while explicitly modeling temporal decay patterns.

### Compressive Memory
Cross-attention memory queries compress the full encoded history into a fixed number of tokens, acting as an information bottleneck that denoises the sequence.

### Adaptive Fusion
A learned gate balances long-term (TADN) and short-term (attention) signals:
```
output = σ(gate(long, short, memory)) * long + (1 - σ) * short
```

## References

Based on ideas from:
- HyTRec (2602.18283) — Hybrid Temporal-Aware Dual-Branch Attention
- Rec2PM (2602.11605) — Recurrent Preference Memory  
- SIGMA (2408.11451) — Bidirectional Selective Gated Mamba
- HSTU (2402.17152) — Generative Recommenders
- SASRec (1808.09781) — Self-Attentive Sequential Recommendation
"""
    
    with open('./checkpoints/README.md', 'w') as f:
        f.write(readme)
    
    # Copy model files
    import shutil
    for fname in ['model.py', 'data.py', 'evaluate.py', 'train.py', 'train_gpu.py']:
        if os.path.exists(fname):
            shutil.copy(fname, f'./checkpoints/{fname}')
    
    upload_folder(
        folder_path='./checkpoints',
        repo_id=hub_model_id,
        commit_message="MARS: Multi-scale Adaptive Recurrence with State compression"
    )
    print(f"\n✓ Pushed to https://huggingface.co/{hub_model_id}")

except Exception as e:
    print(f"Hub push failed: {e}")

print("\nDone!")