直接优先优化

import numpy as np
import sys
import os
sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "..", "04-pre-training-mini-gpt", "code"))
from main import MiniGPT, LayerNorm, Embedding, TransformerBlock

PREFERENCE_DATA = [
    {
        "prompt": "What is the capital of France?",
        "preferred": "The capital of France is Paris.",
        "rejected": "France is a country in Europe. It has many cities. The capital is Paris. Paris is known for the Eiffel Tower.",
    },
    {
        "prompt": "Explain gravity in one sentence.",
        "preferred": "Gravity is the force that attracts objects with mass toward each other.",
        "rejected": "Gravity is something that makes things fall down when you drop them.",
    },
    {
        "prompt": "What is 15 times 7?",
        "preferred": "15 times 7 is 105.",
        "rejected": "Let me think about this. 15 times 7. Well, 10 times 7 is 70, and 5 times 7 is 35, so the answer might be around 105.",
    },
    {
        "prompt": "Name three programming languages.",
        "preferred": "Python, Rust, and TypeScript.",
        "rejected": "There are many programming languages. Some popular ones include various languages like Python and others.",
    },
    {
        "prompt": "What year did World War II end?",
        "preferred": "World War II ended in 1945.",
        "rejected": "World War II was a major global conflict. It involved many countries. The war ended in the mid-1940s, specifically in 1945.",
    },
    {
        "prompt": "Define machine learning.",
        "preferred": "Machine learning is a field where algorithms learn patterns from data to make predictions without being explicitly programmed.",
        "rejected": "Machine learning is a type of AI. AI stands for artificial intelligence. Machine learning uses data to learn.",
    },
]

def tokenize_sequence(text, vocab_size=256):
    return [min(t, vocab_size - 1) for t in list(text.encode("utf-8"))]


def compute_sequence_log_prob(model, prompt_tokens, response_tokens, max_seq_len=128):
    full_sequence = prompt_tokens + response_tokens
    if len(full_sequence) > max_seq_len:
        full_sequence = full_sequence[:max_seq_len]

    if len(full_sequence) < 2:
        return 0.0

    input_ids = np.array(full_sequence[:-1]).reshape(1, -1)
    target_ids = np.array(full_sequence[1:])

    logits = model.forward(input_ids)
    logits = logits[0]

    max_logits = logits.max(axis=-1, keepdims=True)
    log_probs = logits - max_logits - np.log(
        np.exp(logits - max_logits).sum(axis=-1, keepdims=True)
    )

    prompt_len = len(prompt_tokens)
    response_start = max(0, prompt_len - 1)
    response_end = len(target_ids)

    if response_start >= response_end:
        return 0.0

    response_log_probs = log_probs[response_start:response_end, :]
    response_targets = target_ids[response_start:response_end]

    total_log_prob = 0.0
    for i, target in enumerate(response_targets):
        total_log_prob += response_log_probs[i, target]

    return total_log_prob

def sigmoid(x):
    return np.where(
        x >= 0,
        1.0 / (1.0 + np.exp(-x)),
        np.exp(x) / (1.0 + np.exp(x))
    )


def dpo_loss(policy_logprob_preferred, policy_logprob_rejected,
             ref_logprob_preferred, ref_logprob_rejected, beta=0.1):
    preferred_ratio = policy_logprob_preferred - ref_logprob_preferred
    rejected_ratio = policy_logprob_rejected - ref_logprob_rejected

    logit = beta * (preferred_ratio - rejected_ratio)

    loss = -np.log(sigmoid(logit) + 1e-8)

    preferred_reward = beta * preferred_ratio
    rejected_reward = beta * rejected_ratio

    return loss, {
        "preferred_ratio": float(preferred_ratio),
        "rejected_ratio": float(rejected_ratio),
        "logit": float(logit),
        "implicit_preferred_reward": float(preferred_reward),
        "implicit_rejected_reward": float(rejected_reward),
        "reward_margin": float(preferred_reward - rejected_reward),
    }

def copy_model_weights(source, target):
    target.embedding.token_embed = source.embedding.token_embed.copy()
    target.embedding.pos_embed = source.embedding.pos_embed.copy()
    target.ln_f.gamma = source.ln_f.gamma.copy()
    target.ln_f.beta = source.ln_f.beta.copy()
    for s_block, t_block in zip(source.blocks, target.blocks):
        t_block.attn.W_q = s_block.attn.W_q.copy()
        t_block.attn.W_k = s_block.attn.W_k.copy()
        t_block.attn.W_v = s_block.attn.W_v.copy()
        t_block.attn.W_out = s_block.attn.W_out.copy()
        t_block.ffn.W1 = s_block.ffn.W1.copy()
        t_block.ffn.W2 = s_block.ffn.W2.copy()
        t_block.ffn.b1 = s_block.ffn.b1.copy()
        t_block.ffn.b2 = s_block.ffn.b2.copy()
        t_block.ln1.gamma = s_block.ln1.gamma.copy()
        t_block.ln1.beta = s_block.ln1.beta.copy()
        t_block.ln2.gamma = s_block.ln2.gamma.copy()
        t_block.ln2.beta = s_block.ln2.beta.copy()


def dpo_train(policy_model, reference_model, preference_data,
              num_epochs=5, lr=5e-6, beta=0.1, max_seq_len=128):
    print(f"DPO Training: {len(preference_data)} pairs, {num_epochs} epochs, "
          f"lr={lr}, beta={beta}")
    print()

    losses = []
    margins = []

    for epoch in range(num_epochs):
        epoch_loss = 0.0
        epoch_margin = 0.0
        num_examples = 0

        indices = np.random.permutation(len(preference_data))

        for idx in indices:
            pair = preference_data[idx]

            prompt_tokens = tokenize_sequence(pair["prompt"])
            preferred_tokens = tokenize_sequence(pair["preferred"])
            rejected_tokens = tokenize_sequence(pair["rejected"])

            pi_logprob_w = compute_sequence_log_prob(
                policy_model, prompt_tokens, preferred_tokens, max_seq_len
            )
            pi_logprob_l = compute_sequence_log_prob(
                policy_model, prompt_tokens, rejected_tokens, max_seq_len
            )
            ref_logprob_w = compute_sequence_log_prob(
                reference_model, prompt_tokens, preferred_tokens, max_seq_len
            )
            ref_logprob_l = compute_sequence_log_prob(
                reference_model, prompt_tokens, rejected_tokens, max_seq_len
            )

            loss, metrics = dpo_loss(
                pi_logprob_w, pi_logprob_l,
                ref_logprob_w, ref_logprob_l, beta
            )

            update_direction = 1.0 if metrics["logit"] < 0 else -0.1
            for block in policy_model.blocks:
                block.ffn.W1 += lr * update_direction * np.random.randn(*block.ffn.W1.shape) * 0.01
                block.ffn.W2 += lr * update_direction * np.random.randn(*block.ffn.W2.shape) * 0.01

            epoch_loss += loss
            epoch_margin += metrics["reward_margin"]
            num_examples += 1
            losses.append(float(loss))
            margins.append(metrics["reward_margin"])

        avg_loss = epoch_loss / max(num_examples, 1)
        avg_margin = epoch_margin / max(num_examples, 1)

        print(f"  Epoch {epoch + 1}/{num_epochs} | Loss: {avg_loss:.4f} | "
              f"Avg Margin: {avg_margin:.4f}")

    return policy_model, losses, margins

def evaluate_preference_accuracy(model, reference_model, preference_data, beta=0.1, max_seq_len=128):
    correct = 0
    total = 0

    for pair in preference_data:
        prompt_tokens = tokenize_sequence(pair["prompt"])
        preferred_tokens = tokenize_sequence(pair["preferred"])
        rejected_tokens = tokenize_sequence(pair["rejected"])

        pi_w = compute_sequence_log_prob(model, prompt_tokens, preferred_tokens, max_seq_len)
        pi_l = compute_sequence_log_prob(model, prompt_tokens, rejected_tokens, max_seq_len)
        ref_w = compute_sequence_log_prob(reference_model, prompt_tokens, preferred_tokens, max_seq_len)
        ref_l = compute_sequence_log_prob(reference_model, prompt_tokens, rejected_tokens, max_seq_len)

        preferred_reward = beta * (pi_w - ref_w)
        rejected_reward = beta * (pi_l - ref_l)

        if preferred_reward > rejected_reward:
            correct += 1
        total += 1

    return correct / max(total, 1)


def analyze_implicit_rewards(model, reference_model, preference_data, beta=0.1, max_seq_len=128):
    print("Implicit Reward Analysis:")
    print("-" * 65)
    print(f"  {'Prompt':<30} {'Pref Reward':>12} {'Rej Reward':>12} {'Margin':>10}")
    print("  " + "-" * 60)

    for pair in preference_data:
        prompt_tokens = tokenize_sequence(pair["prompt"])
        preferred_tokens = tokenize_sequence(pair["preferred"])
        rejected_tokens = tokenize_sequence(pair["rejected"])

        pi_w = compute_sequence_log_prob(model, prompt_tokens, preferred_tokens, max_seq_len)
        pi_l = compute_sequence_log_prob(model, prompt_tokens, rejected_tokens, max_seq_len)
        ref_w = compute_sequence_log_prob(reference_model, prompt_tokens, preferred_tokens, max_seq_len)
        ref_l = compute_sequence_log_prob(reference_model, prompt_tokens, rejected_tokens, max_seq_len)

        pref_reward = beta * (pi_w - ref_w)
        rej_reward = beta * (pi_l - ref_l)
        margin = pref_reward - rej_reward

        truncated = pair["prompt"][:28] + ".." if len(pair["prompt"]) > 30 else pair["prompt"]
        print(f"  {truncated:<30} {pref_reward:>12.4f} {rej_reward:>12.4f} {margin:>10.4f}")

    print()

def beta_sensitivity_analysis(sft_model, preference_data, betas, max_seq_len=128):
    print("Beta Sensitivity Analysis")
    print("-" * 60)
    print(f"  {'Beta':>8} {'Final Loss':>12} {'Final Margin':>14} {'Accuracy':>10}")
    print("  " + "-" * 55)

    results = []

    for beta in betas:
        policy = MiniGPT(
            vocab_size=256, embed_dim=128, num_heads=4,
            num_layers=4, max_seq_len=max_seq_len, ff_dim=512
        )
        reference = MiniGPT(
            vocab_size=256, embed_dim=128, num_heads=4,
            num_layers=4, max_seq_len=max_seq_len, ff_dim=512
        )
        copy_model_weights(sft_model, policy)
        copy_model_weights(sft_model, reference)

        policy, losses, margins_list = dpo_train(
            policy, reference, preference_data,
            num_epochs=3, lr=5e-6, beta=beta, max_seq_len=max_seq_len
        )

        accuracy = evaluate_preference_accuracy(
            policy, reference, preference_data, beta, max_seq_len
        )

        final_loss = losses[-1] if losses else 0
        final_margin = margins_list[-1] if margins_list else 0

        print(f"  {beta:>8.3f} {final_loss:>12.4f} {final_margin:>14.4f} {accuracy:>10.1%}")
        results.append({
            "beta": beta,
            "final_loss": final_loss,
            "final_margin": final_margin,
            "accuracy": accuracy,
        })

        print()

    return results

if __name__ == "__main__":
    np.random.seed(42)

    print("=" * 70)
    print("DPO: DIRECT PREFERENCE OPTIMIZATION")
    print("=" * 70)
    print()

    print("STEP 1: Initialize SFT Model (from Lesson 06)")
    print("-" * 50)
    sft_model = MiniGPT(
        vocab_size=256, embed_dim=128, num_heads=4,
        num_layers=4, max_seq_len=128, ff_dim=512
    )
    print(f"  Parameters: {sft_model.count_parameters():,}")
    print()

    print("STEP 2: DPO Training")
    print("-" * 50)

    policy_model = MiniGPT(
        vocab_size=256, embed_dim=128, num_heads=4,
        num_layers=4, max_seq_len=128, ff_dim=512
    )
    reference_model = MiniGPT(
        vocab_size=256, embed_dim=128, num_heads=4,
        num_layers=4, max_seq_len=128, ff_dim=512
    )
    copy_model_weights(sft_model, policy_model)
    copy_model_weights(sft_model, reference_model)

    policy_model, losses, margins = dpo_train(
        policy_model, reference_model, PREFERENCE_DATA,
        num_epochs=5, lr=5e-6, beta=0.1
    )
    print()

    print("=" * 70)
    print("STEP 3: Evaluate")
    print("=" * 70)
    print()

    pre_accuracy = evaluate_preference_accuracy(
        sft_model, reference_model, PREFERENCE_DATA, beta=0.1
    )
    post_accuracy = evaluate_preference_accuracy(
        policy_model, reference_model, PREFERENCE_DATA, beta=0.1
    )

    print(f"  Preference accuracy (pre-DPO):  {pre_accuracy:.1%}")
    print(f"  Preference accuracy (post-DPO): {post_accuracy:.1%}")
    print()

    analyze_implicit_rewards(policy_model, reference_model, PREFERENCE_DATA, beta=0.1)

    print("=" * 70)
    print("STEP 4: Training Dynamics")
    print("=" * 70)
    print()

    if losses:
        print("  Loss curve:")
        window = max(1, len(losses) // 5)
        for i in range(0, len(losses), window):
            chunk = losses[i:i + window]
            avg = sum(chunk) / len(chunk)
            print(f"    Steps {i:3d}-{i + len(chunk) - 1:3d}: loss = {avg:.4f}")
        print()

    if margins:
        print("  Reward margin curve:")
        window = max(1, len(margins) // 5)
        for i in range(0, len(margins), window):
            chunk = margins[i:i + window]
            avg = sum(chunk) / len(chunk)
            print(f"    Steps {i:3d}-{i + len(chunk) - 1:3d}: margin = {avg:.4f}")
        print()

    print("=" * 70)
    print("STEP 5: Beta Sensitivity")
    print("=" * 70)
    print()

    beta_results = beta_sensitivity_analysis(
        sft_model, PREFERENCE_DATA, betas=[0.01, 0.1, 0.3, 1.0]
    )

    print("=" * 70)
    print("DPO vs RLHF COMPARISON")
    print("=" * 70)
    print()
    print("  DPO advantages:")
    print("    - 1 training loop (vs 3 for RLHF)")
    print("    - 2 models in memory (vs 3-4 for RLHF)")
    print("    - Supervised learning (vs RL, more stable)")
    print("    - No reward model to train or maintain")
    print()
    print("  RLHF advantages:")
    print("    - Separate reward model captures complex preferences")
    print("    - Online learning: generate, rate, retrain")
    print("    - Better for multi-objective alignment")
    print("    - Proven at largest scales (GPT-4, Claude)")
    print()
    print("  Practical guidance:")
    print("    - Start with DPO. It's simpler and often sufficient.")
    print("    - Switch to RLHF if DPO plateaus on your eval metrics.")
    print("    - Many production systems use both: RLHF first, DPO to refine.")