scripts/examples/better_sampling.py

#!/usr/bin/env python3
"""
Better Sampling for BitTransformerLM
"""

import sys
import torch
import torch.nn.functional as F

sys.path.append('/data')
sys.path.append('/data/BitTransformerLM')

from bit_transformer import BitTransformerLM, text_to_bits, bits_to_text

def load_model():
    model = BitTransformerLM(
        d_model=512, nhead=16, num_layers=8, dim_feedforward=1024,
        max_seq_len=512, reversible=True, use_checkpoint=False,
        use_autocast=False, use_act=True, act_threshold=0.9,
        lambda_K=0.05, lambda_C=0.05, lambda_S=0.05
    )
    
    checkpoint = torch.load('/data/BitTransformerLM/checkpoints/checkpoint_best.pt', map_location='cpu')
    model.load_state_dict(checkpoint['model_state_dict'])
    model.eval()
    
    return model

def smart_generate(model, prompt, max_chars=5):
    """Generate with better sampling strategies."""
    print(f"\n🎯 Smart generating from: '{prompt}'")
    
    input_bits = text_to_bits(prompt)
    generated_bits = input_bits.copy()
    
    with torch.no_grad():
        for char_idx in range(max_chars):
            # Generate 9 bits for one character (8 data + 1 parity)
            char_bits = []
            
            for bit_idx in range(9):
                # Context (keep reasonable length)
                context = generated_bits + char_bits
                context = context[-300:] if len(context) > 300 else context
                context_tensor = torch.tensor(context, dtype=torch.long).unsqueeze(0)
                
                logits, telemetry = model(context_tensor)
                next_bit_logits = logits[0, -1, :]
                
                # Different strategies based on bit position
                if bit_idx < 8:  # Data bits
                    # Use higher temperature for more variety
                    temperature = 0.8
                    next_bit_logits = next_bit_logits / temperature
                    
                    # Top-k sampling
                    k = 2  # Only 2 options anyway (0 or 1)
                    top_k_logits, top_k_indices = torch.topk(next_bit_logits, k)
                    probs = F.softmax(top_k_logits, dim=-1)
                    selected_idx = torch.multinomial(probs, 1).item()
                    next_bit = top_k_indices[selected_idx].item()
                else:  # Parity bit
                    # Calculate correct parity
                    data_bits = char_bits[:8]
                    expected_parity = sum(data_bits) % 2
                    next_bit = expected_parity
                
                char_bits.append(next_bit)
            
            # Add completed character
            generated_bits.extend(char_bits)
            
            # Try to decode the new character
            try:
                new_char_bits = char_bits
                # Convert to bytes (remove parity)
                data_bits = new_char_bits[:8]
                byte_val = sum(bit * (2**(7-i)) for i, bit in enumerate(data_bits))
                
                if 32 <= byte_val <= 126:  # Printable ASCII
                    char = chr(byte_val)
                    print(f"  Char {char_idx+1}: '{char}' (byte={byte_val})")
                    
                    # Early stopping for sentence enders
                    if char in '.!?\n':
                        break
                else:
                    print(f"  Char {char_idx+1}: Non-printable (byte={byte_val})")
                    
            except Exception as e:
                print(f"  Char {char_idx+1}: Decode error: {e}")
    
    # Final decode attempt
    generated_only = generated_bits[len(input_bits):]
    try:
        final_text = bits_to_text(generated_only)
        print(f"✨ Result: '{prompt}' + '{final_text}'")
        return final_text
    except Exception as e:
        print(f"❌ Final decode failed: {e}")
        
        # Manual decode of complete characters
        manual_result = ""
        for i in range(0, len(generated_only), 9):
            if i + 8 < len(generated_only):
                char_bits = generated_only[i:i+8]  # Just data bits
                byte_val = sum(bit * (2**(7-j)) for j, bit in enumerate(char_bits))
                if 32 <= byte_val <= 126:
                    manual_result += chr(byte_val)
                else:
                    manual_result += '?'
        
        print(f"🔧 Manual decode: '{prompt}' + '{manual_result}'")
        return manual_result

def main():
    print("🚀 SMART BITRANSFORMERLM GENERATION")
    print("=" * 40)
    
    model = load_model()
    print("✅ Model loaded!")
    
    # Test different prompt styles
    prompts = [
        "Hello",
        "Hi",
        "A",
        "The cat",
        "I am",
        "Yes",
        "No"
    ]
    
    for prompt in prompts:
        result = smart_generate(model, prompt, max_chars=4)

if __name__ == "__main__":
    main()