train_copy_task.py

"""
Training Demo: Teach the Transformer to copy sequences.

The "copy task" is the classic smoke test for sequence-to-sequence models:
  Input:  [BOS, 5, 3, 8, 2, 7, EOS, PAD, PAD]
  Output: [BOS, 5, 3, 8, 2, 7, EOS, PAD, PAD]

If the model learns to copy, it proves that:
  - Encoder correctly represents the source sequence
  - Cross-attention correctly attends to source positions
  - Decoder correctly generates autoregressively
  - Masking (causal + padding) works correctly
  - The training loop (optimizer, LR schedule, label smoothing) works

We use a small model (d_model=64, 2 layers) so it trains in seconds on CPU.
"""

import torch
import torch.nn as nn
from transformer import Transformer, TransformerLRScheduler, greedy_decode

# Special tokens
PAD_IDX = 0
BOS_IDX = 1
EOS_IDX = 2
VOCAB_SIZE = 15  # Small vocabulary: tokens 3..14 are "real" tokens


def generate_copy_batch(batch_size: int, seq_len: int, device: torch.device):
    """
    Generate a batch for the copy task.
    
    Source: [random tokens, EOS, PAD...]
    Target: [BOS, same random tokens, EOS, PAD...]
    """
    lengths = torch.randint(3, seq_len - 1, (batch_size,))
    
    src = torch.full((batch_size, seq_len), PAD_IDX, dtype=torch.long, device=device)
    tgt = torch.full((batch_size, seq_len), PAD_IDX, dtype=torch.long, device=device)
    
    for i in range(batch_size):
        l = lengths[i].item()
        tokens = torch.randint(3, VOCAB_SIZE, (l,))
        
        src[i, :l] = tokens
        src[i, l] = EOS_IDX
        
        tgt[i, 0] = BOS_IDX
        tgt[i, 1:l+1] = tokens
        tgt[i, l+1] = EOS_IDX
    
    return src, tgt


def train():
    device = torch.device('cpu')
    
    config = {
        'src_vocab_size': VOCAB_SIZE,
        'tgt_vocab_size': VOCAB_SIZE,
        'd_model': 64,
        'n_heads': 4,
        'n_layers': 2,
        'd_ff': 256,
        'dropout': 0.0,
        'max_len': 100,
        'pad_idx': PAD_IDX,
        'tie_weights': True,
    }
    
    model = Transformer(**config).to(device)
    total_params = sum(p.numel() for p in model.parameters())
    print(f"Model parameters: {total_params:,}")
    
    optimizer = torch.optim.Adam(
        model.parameters(), lr=1e-3, betas=(0.9, 0.98), eps=1e-9
    )
    criterion = nn.CrossEntropyLoss(ignore_index=PAD_IDX)
    
    batch_size = 32
    seq_len = 10
    n_steps = 3000
    
    print(f"\nTraining copy task for {n_steps} steps...")
    print(f"Batch size: {batch_size}, Seq length: {seq_len}")
    print("-" * 50)
    
    model.train()
    for step in range(1, n_steps + 1):
        src, tgt = generate_copy_batch(batch_size, seq_len, device)
        
        tgt_input = tgt[:, :-1]
        tgt_label = tgt[:, 1:]
        
        logits = model(src, tgt_input)
        loss = criterion(
            logits.reshape(-1, logits.size(-1)),
            tgt_label.reshape(-1),
        )
        
        optimizer.zero_grad()
        loss.backward()
        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
        optimizer.step()
        
        if step % 300 == 0 or step == 1:
            preds = logits.argmax(dim=-1)
            mask = tgt_label != PAD_IDX
            correct = ((preds == tgt_label) & mask).sum().item()
            total = mask.sum().item()
            acc = correct / total * 100
            lr = optimizer.param_groups[0]['lr']
            print(f"Step {step:>5d} | Loss: {loss.item():.4f} | Acc: {acc:.1f}% | LR: {lr:.6f}")
    
    print("\n" + "=" * 50)
    print("EVALUATION: Greedy Decode Examples")
    print("=" * 50)
    
    model.eval()
    n_correct = 0
    n_total = 10
    
    for i in range(n_total):
        src, tgt = generate_copy_batch(1, seq_len, device)
        
        src_tokens = []
        for t in src[0]:
            if t.item() in (PAD_IDX, EOS_IDX):
                break
            src_tokens.append(t.item())
        
        decoded = greedy_decode(model, src, max_len=seq_len + 2, bos_idx=BOS_IDX, eos_idx=EOS_IDX)
        decoded_tokens = []
        for t in decoded[1:]:
            if t.item() == EOS_IDX:
                break
            decoded_tokens.append(t.item())
        
        match = src_tokens == decoded_tokens
        n_correct += match
        
        status = "✅" if match else "❌"
        print(f"  {status} Source:  {src_tokens}")
        print(f"     Output:  {decoded_tokens}")
        if not match:
            print(f"     Expected: {src_tokens}")
        print()
    
    print(f"Copy accuracy: {n_correct}/{n_total} ({n_correct/n_total*100:.0f}%)")
    
    if n_correct >= 8:
        print("\n🎉 The Transformer learned the copy task! Model is working correctly.")
    else:
        print("\n⚠️  Model didn't fully converge. Try more steps or adjust hyperparameters.")


if __name__ == '__main__':
    train()