PPO学习

• openai用tf实现的真的看不懂，大佬的世界…
  

PPO的详细细节

1. 奖励模型和策略的价值头将 query 和 response 的连接作为输入

1. 奖励模型和策略的价值头 不 仅仅查看响应。相反，它将 query 和 response 连接在一起，作为 query_response

   def get_rewards_op(self, queries, responses):
           tokens = tf.concat([queries, responses], axis=1)
           return self._build(tokens)
   

2. 举例来说，如果 query = “他在想某事，但他的眼神很难读懂”。 ，和 response = “他看着他的左手，手臂伸在他的前面。” ，那么奖励模型和策略的价值会对query_response = “他在想某事，但他的眼神很难读懂。他看着他的左手，手臂伸在他的前面。” 进行前向传递，并产生形状为 (B, T, 1) 的奖励和价值，其中 B 是 BS (批量大小)，T 是序列长度，而 1 代表奖励头的输出结构的维度为 1

   def respond_op(self, queries, length):
           contexts = self.embed_queries(queries)
           context_length = tf.shape(contexts)[1]
           result = sample.sample_sequence(
               step=self.step_core,
               context=contexts,
               length=length,
               model_hparams=self.model_hparams,
               temperature=self.temperature,
               extra_outputs={'values':tf.float32},
           )
           return dict(
               responses=result['tokens'][:, context_length:],
               logprobs=result['logprobs'],
               values=result['values'],
           )
   

3. T 意味着每个 token 都有与其和前文关联的奖励。例如，眼神 token 将有一个与他在想某事，但他的眼神很难读懂 相对应的奖励。

2.使用特殊的填充 token 来填充和截断输入

1. Openai 为查询 query_length 设置了固定的输入长度; 它使用 pad_token 填充 过短的序列并 截断 过长的序列

def _generator():
    inner_gen = self.generator(mode, seed=seed, shuffle=shuffle, comm=comm)
     for text in inner_gen:
         tokens = encoder.encode(text)
         if start_token is not None:
             try:
                 first_index = tokens.index(start_token)+1
                 if first_index < len(tokens):
                     tokens = tokens[first_index:]
             except:
                 continue

         tokens = tokens[:sequence_length] # token截断

         if end_token is not None:
             try:
                 last_index = len(tokens)-tokens[::-1].index(end_token)
                 tokens = tokens[:last_index]
             except:
                 continue

         if len(tokens) < sequence_length:
             tokens = tokens + [padding_token] * (sequence_length - len(tokens)) # padding token填充

         assert len(tokens) == sequence_length

         yield dict(tokens=tokens)

2. 在填充输入时，OAI 使用了词汇表之外的 token self.padding_token = len(encoder) + 2 # +2 unnecessary, for historical reasons
   • 在 GPT 和 GPT-2 的预训练期间没有使用填充 token; 因此，transformer 的 gpt2 模型与其分词器没有关联的官方填充 token。通常的做法是设置 tokenizer.pad_token = tokenizer.eos_token ，但在这项工作中，我们将区分这两个特殊 token 以匹配 OAI 的原始设置，所以我们将使用tokenizer.add_special_tokens({"pad_token": "[PAD]"})

# 实战案例
import transformers
tokenizer = transformers.AutoTokenizer.from_pretrained("gpt2", padding_side="right")
tokenizer.add_special_tokens({"pad_token": "[PAD]"})
query_length = 5
texts = [
    "usually, he would",
    "she thought about it",
]
tokens = []
for text in texts:
    tokens.append(tokenizer.encode(text)[:query_length])

print("tokens", tokens)
inputs = tokenizer.pad(
    {"input_ids": tokens},
    padding="max_length",
    max_length=query_length,
    return_tensors="pt",
    return_attention_mask=True,
)
print("inputs", inputs)

"""prints are
tokens [[23073, 11, 339, 561], [7091, 1807, 546, 340]]
inputs {'input_ids': tensor([[23073, 11, 339, 561, 50257],
        [ 7091, 1807, 546, 340, 50257]]), 'attention_mask': tensor([[1, 1, 1, 1, 0],
        [1, 1, 1, 1, 0]])}
"""

3.生成固定长度响应的响应生成不需要填充

1. 在响应生成期间，OAI 使用 top_k=0, top_p=1.0 并仅在词汇表上做分类样本 ，代码会一直采样，直到生成固定长度的响应。值得注意的是，即使遇到 EOS (序列结束) token ，它也会继续采样。

def body(past, prev, output, logprobs, *extras):
     next_outputs = step(model_hparams, prev[:, tf.newaxis], past=past,
                          past_tokens=output[:, :-1])
      logits = tf.cast(next_outputs['logits'], tf.float32) * beta
      if top_k != 0:
          logits = tf.cond(tf.equal(top_k, 0),
                           lambda: logits,
                           lambda: utils.take_top_k_logits(logits, top_k))
      if top_p != 1.0:
          logits = utils.take_top_p_logits(logits, top_p)
      # 采样
      next_sample = utils.sample_from_logits(logits, dtype=tf.int32) 
      next_logprob = utils.logprobs_from_logits(logits=logits, labels=next_sample)
      
      return [
          tf.concat([past, next_outputs['presents']], axis=-2),
          tf.squeeze(next_sample, axis=[1]),
          tf.concat([output, next_sample], axis=1),
          tf.concat([logprobs, next_logprob], axis=1),
          *[tf.concat([prev, next_outputs[k]], axis=1) for k, prev in zip(extra_outputs, extras)],
      ]

实战案例：

import torch
import transformers
tokenizer = transformers.AutoTokenizer.from_pretrained("gpt2", padding_side="right")
tokenizer.add_special_tokens({"pad_token": "[PAD]"}) # 添加特殊词元
pad_id = tokenizer.pad_token_id
query = torch.tensor([
    [pad_id, pad_id, 23073],
])
response = torch.tensor([
    [11, 339, 561],
])
response_length = 4
temperature = 0.7
pretrained_model = transformers.AutoModelForCausalLM.from_pretrained("gpt2")
pretrained_model.generation_config.eos_token_id = None # disable `pad_token_id` and `eos_token_id` because we just want to
pretrained_model.generation_config.pad_token_id = None # generate tokens without truncation / padding
generation_config = transformers.GenerationConfig(
    max_new_tokens=response_length,
    min_new_tokens=response_length,
    temperature=temperature,
    top_k=0.0,
    top_p=1.0,
    do_sample=True,
)
context_length = query.shape[1]
attention_mask = query != tokenizer.pad_token_id
input_ids = query.clone()
input_ids[~attention_mask] = 0 # set padding tokens to 0
output = pretrained_model.generate(
    input_ids=input_ids,
    attention_mask=attention_mask,
    # position_ids=attention_mask.cumsum(1) - attention_mask.long(), # generation collapsed if this was turned on. TODO: why does generation collapse with this?
    generation_config=generation_config,
    return_dict_in_generate=True,
)
print(output.sequences)

"""
tensor([[ 0, 0, 23073, 16851, 11, 475, 991]])
"""

5.奖励模型和策略训练的学习率退火

1. 奖励模型只训练一个 epcho，以避免过度拟合有限量的人类注释数据 (例如，descriptiveness 任务只有大约 5000 个标签)。在这个单一的 epcho 中，学习率会退火至零

2. 类似于奖励模型训练，策略训练的学习率也会退火至零

def train(self):
        labels = download_labels(
            self.hparams.labels.source,
            label_type=self.label_type,
            question_schemas=self.question_schemas,
            total_labels=self.hparams.labels.num_train,
            comm=self.comm
        )

        self.add_to_buffer(labels)

        if self.hparams.normalize_before:
            target_mean, target_std = self.target_mean_std()
            self.normalize(self.sample_policy_responses, target_mean, target_std)

        # Collect training data for reward model training.  train_indices will include the indices
        # trained on across all ranks, and its size must be a multiple of minibatch_size.
        per_rank_batch_size = utils.exact_div(self.hparams.batch_size, self.num_ranks)

        # Make sure each rank gets the same shuffle so we train on each point exactly once
        train_indices = self.comm.bcast(np.random.permutation(self.hparams.labels.num_train))

        # Train on train_indices
        print(self.rank, "training on", self.hparams.labels.num_train, "in batches of", per_rank_batch_size)
        for start_index in range(0, self.hparams.labels.num_train, self.hparams.batch_size):
            end_index = start_index + self.hparams.batch_size
            all_ranks_indices = train_indices[start_index:end_index]
            our_indices = all_ranks_indices[self.rank::self.num_ranks]
            lr = (1 - start_index / self.hparams.labels.num_train) * self.hparams.lr  # 学习率退火
            self.train_batch(our_indices, lr)

        if self.hparams.normalize_after:
            target_mean, target_std = np.zeros([]), np.ones([])
            self.normalize(self.sample_policy_responses, target_mean, target_std)

Reward模型训练细节

1.奖励模型只输出最后一个 token 的值

1. 在对 query 和 response 的连接进行前向传递后获得的奖励将具有形状 (B, T, 1) ，其中 B 是 BS(批量大小)，T 是序列长度 (始终相同; 在 OAI 的设置中，它是 query_length + response_length = 64 + 24 = 88 ，用于风格任务，参见 launch.py#L9-L11)，1 是奖励头其维度为 1
   • 原始代码库提取最后一个 token 的奖励 ，因此奖励将只具有形状 (B, 1)

 def _build(self, X):
        results = self.model(X=X, padding_token=self.padding_token)
        reward = results['reward'][:, -1] # 取最后一个token
        with tf.variable_scope(f'{self.scope}/reward_norm'):
            self.reward_gain = tf.get_variable('gain', shape=(), initializer=tf.constant_initializer(1))
            self.reward_bias = tf.get_variable('bias', shape=(), initializer=tf.constant_initializer(0))
        reward = self.reward_gain * reward + self.reward_bias
        self._set_initializers()
        return reward

2.奖励头层初始化

$$initial=\mathcal{N}\left(0,1/\left(\sqrt{d_{\text{model }}+1}\right)\right)$$

def fc_layer(x, outshape, *, in_axes=1, scale=None):
    inshape = tuple([int(d) for d in x.shape[-in_axes:]]) if in_axes>0 else ()
    outshape = tuple(outshape)
    if scale is None:
        scale = 1 / np.sqrt(np.prod(inshape) + 1)
    w = tf.get_variable('w', inshape + outshape, initializer=tf.random_normal_initializer(stddev=scale)) # 权重初始化
    b = tf.get_variable('b', outshape, initializer=tf.constant_initializer(0)) # 偏置初始化为0
    # Call the regularizer manually so that it works correctly with GradientTape
    regularizer = tf.contrib.layers.l2_regularizer(scale=1/np.prod(outshape)) #so that initial value of regularizer is 1
    reg_loss = regularizer(w)
    return tensordot(x, w, in_axes) + b, reg_loss

dropped_h = dropout(h, self.hparams.head_pdrop, do_dropout=do_dropout, seed=head_seed, name='drop')
# TODO: refactor this, perhaps move to Policy
res, reg_loss = fc_layer(dropped_h, (), scale=0 if head_name == 'value' else None)

3.奖励模型的前后归一化

1. 在论文中，Ziegler 等人 (2019) 提到“为了保持训练过程中奖励模型的规模一致，我们将其归一化，使其在 $$(x\sim \mathcal{D},y\sim \rho (\cdot |x))$$ 的情况下，均值为$0$，方差为 $1$”。为了执行归一化过程，代码首先创建了 reward_gain 和 reward_bias ，以便可以通过 reward = reward * reward_gain + reward_bias 来计算奖励值

def _build(self, tokens, do_dropout=False, name=None):
        with tf.variable_scope(self.scope, reuse=self.built, auxiliary_name_scope=not self.built, use_resource=self.use_resource):
            lm_output = self.model(X=tokens, do_dropout=do_dropout, padding_token=self.padding_token)

            reward = lm_output['reward'][:, -1] # 奖励取最后一个token
            with tf.variable_scope('reward_norm'):
                if not self.built:
                    self.reward_gain = tf.get_variable('gain', shape=(), initializer=tf.constant_initializer(1)) # 奖励权重w
                    self.reward_bias = tf.get_variable('bias', shape=(), initializer=tf.constant_initializer(0)) # 奖励偏置b
                    self._reward_gain_p = tf.placeholder(name='gain_p', dtype=tf.float32, shape=())
                    self._reward_bias_p = tf.placeholder(name='bias_p', dtype=tf.float32, shape=())
                    self._set_reward_norm = tf.group(self.reward_gain.assign(self._reward_gain_p),
                                                     self.reward_bias.assign(self._reward_bias_p))
                if reward is not None:
                    reward = self.reward_gain * reward + self.reward_bias  # reward计算
            if not self.built:
                self._set_initializers()
            self.built = True
            return reward

2. 在执行归一化过程时，代码首先设置 reward_gain=1, reward_bias=0 ，然后从目标数据集 中收集采样查询、完成的响应和评估的奖励。接着，它得到评估奖励的 实证均值和标准差，并尝试计算 reward_gain 和 reward_bias 应该是什么。

 def normalize(self, sample_fn, target_means, target_stds):
        if not self.hparams.normalize_samples:
            return

        self.reset_reward_scales() #  reward_gain=1, reward_bias=0 
        query_responses = sample_fn(self.hparams.normalize_samples) # 采样
        means, stds = self.stats(query_responses) # 评估奖励的 实证均值和标准差

        self.set_reward_norms(means, stds, target_means, target_stds) # 归一化
        if self.hparams.debug_normalize:
            query_responses = sample_fn(self.hparams.debug_normalize)
            stats = self.stats(query_responses)
            self.log_stats_after_normalize(stats)

我们用(${\mu }_{\mathcal{D}}$) 来表示实证均值，用( ${\sigma }_{\mathcal{D}}$ ) 表示实证标准差，用$(g)$表示 reward_gain ，用($b$) 表示 reward_bias ，用( ${\mu }_{\mathcal{T}}=0$) 表示 目标均值，用( ${\sigma }_{\mathcal{T}}=1$) 表示 目标标准差。然后我们有以下公式。

$$\begin{array}{rlrl}g\mathcal{N}({\mu }_{\mathcal{D}},{\sigma }_{\mathcal{D}})+b& =\mathcal{N}(g{\mu }_{\mathcal{D}},g{\sigma }_{\mathcal{D}})+b=\mathcal{N}(g{\mu }_{\mathcal{D}}+b,g{\sigma }_{\mathcal{D}})=\mathcal{N}({\mu }_{\mathcal{T}},{\sigma }_{\mathcal{T}})\cdot g& =\frac{{\sigma }_{\mathcal{T}}}{{\sigma }_{\mathcal{D}}}\cdot b& ={\mu }_{\mathcal{T}}-g{\mu }_{\mathcal{D}}\end{array}$$

3. 然后在奖励模型训练的 前 和 后 应用归一化过程

def train(self):
        labels = download_labels(
            self.hparams.labels.source,
            label_type=self.label_type,
            question_schemas=self.question_schemas,
            total_labels=self.hparams.labels.num_train,
            comm=self.comm
        )

        self.add_to_buffer(labels)
		
        if self.hparams.normalize_before: # 训练前进行mean和std归一化
            target_mean, target_std = self.target_mean_std()
            self.normalize(self.sample_policy_responses, target_mean, target_std)

        # Collect training data for reward model training.  train_indices will include the indices
        # trained on across all ranks, and its size must be a multiple of minibatch_size.
        per_rank_batch_size = utils.exact_div(self.hparams.batch_size, self.num_ranks)

        # Make sure each rank gets the same shuffle so we train on each point exactly once
        train_indices = self.comm.bcast(np.random.permutation(self.hparams.labels.num_train))

        # Train on train_indices
        print(self.rank, "training on", self.hparams.labels.num_train, "in batches of", per_rank_batch_size)
        for start_index in range(0, self.hparams.labels.num_train, self.hparams.batch_size):
            end_index = start_index + self.hparams.batch_size
            all_ranks_indices = train_indices[start_index:end_index]
            our_indices = all_ranks_indices[self.rank::self.num_ranks]
            lr = (1 - start_index / self.hparams.labels.num_train) * self.hparams.lr
            self.train_batch(our_indices, lr)

        if self.hparams.normalize_after: #训练后进行mean和std归一化
            target_mean, target_std = np.zeros([]), np.ones([])
            self.normalize(self.sample_policy_responses, target_mean, target_std)

4. 归一化目的生成的响应 ( $y\sim \rho (\cdot |x)$) 来自预训练的语言模型 ($\rho$ )。模型 ($\rho$) 被固定为参考，并且在奖励学习中不会更新

reward_trainer = RewardModelTrainer(
            reward_model=reward_model,
            policy=ref_policy, # reward模型来更新p
            query_sampler=query_sampler,
            hparams=hparams,
            comm=comm,
        )

策略训练细节

1. 通过采样温度来缩放 logits

1. 在计算响应的对数概率时，模型首先输出响应中 token 的 logits，然后用采样温度除以这些 logits ,即 logits /= self.temperature
2. 在一个非正式的测试中，我们发现如果不进行此缩放，KL 散度会比预期更快地上升，性能会下降。

2. 价值头层的初始化

1. 价值头的权重是根据 ($\mathcal{N}(0,0)$) 进行初始化的
2. 价值头的p偏置为0

3. 选择以句号开始和结束的查询文本

数据预处理的细节:

1. 尝试仅在 start_text=“.” 之后选择文本
2. 尝试在 end_text=“.” 之前选择文本
3. 然后填充文本

 def _generator():
            inner_gen = self.generator(mode, seed=seed, shuffle=shuffle, comm=comm)
            for text in inner_gen:
                tokens = encoder.encode(text)
                if start_token is not None: # start_text="." 
                    try:
                        first_index = tokens.index(start_token)+1
                        if first_index < len(tokens):
                            tokens = tokens[first_index:]
                    except:
                        continue

                tokens = tokens[:sequence_length]

                if end_token is not None: #  end_text="."
                    try:
                        last_index = len(tokens)-tokens[::-1].index(end_token)
                        tokens = tokens[:last_index]
                    except:
                        continue

                if len(tokens) < sequence_length: # 填充文本
                    tokens = tokens + [padding_token] * (sequence_length - len(tokens))

                assert len(tokens) == sequence_length

                yield dict(tokens=tokens)

禁用 dropout

• 策略训练中不使用 dropout

def step_core(self, model_hparams, tokens, past=None, past_tokens=None, do_dropout=False, name=None):
        with tf.name_scope(name, 'step'):
            with tf.variable_scope(
                    self.scope,
                    reuse=self.built,
                    auxiliary_name_scope=not self.built,
                    use_resource=self.use_resource):
                lm_output = self.model(X=tokens, past=past, past_tokens=past_tokens,
                                       do_dropout=do_dropout, padding_token=self.padding_token)

                # need to slice logits since we don't want to generate special tokens
                logits = lm_output['lm_logits'][:,:,:self.model_hparams.n_vocab]
                presents = lm_output['present']
                value = lm_output['value']
                if not self.built:
                    self._set_initializers()
                self.built = True
                return {
                    'logits': logits,
                    'values': value,
                    'presents': presents,
                }

拒绝采样

• Ziegler 等人 (2019) 建议: “我们使用拒绝采样来确保在第 16 和 24 个 token 之间有一个句号，然后在那个句号处截断 (这是‘句子结束’的粗略近似。我们选择它是因为它很容易集成到 RL 循环中，即使是粗略的近似也足以使人类评估任务变得稍微容易一些)。
• 在 RL 微调期间，我们对没有这样的句号的延续给予固定奖励 -1。

如何实现？

1. token 截断：我们想要在第一个出现在响应的 truncate_after 位置之后的 truncate_token 处截断，将截断 token 后的所有 token 替换为填充 token

2. 在截断响应上运行奖励模型: 在 token 截断过程将响应截断后，代码然后在 截断的响应 上运行奖励模型。

3. 拒绝采样: 如果在第 16 和 24 个 token 之间没有句号，那么将响应的分数替换为固定的低值

def make_score_fn(hparams, score_model):
    padding_token = score_model.padding_token

    postprocess_fn = lm_tasks.postprocess_fn_from_hparams(hparams, padding_token)
    #decorate requires a named function, postprocess_fn can be anonymous
    @utils.graph_function(responses=Schema(tf.int32, (None, None)))
    def postprocess(responses):
        return postprocess_fn(responses)

    filter_fn = lm_tasks.filter_fn_from_hparams(hparams)
    @utils.graph_function(
        responses=Schema(tf.int32, (None, None)),
        rewards=Schema(tf.float32, (None,)))
    def penalize(responses, rewards):
        valid = filter_fn(responses)
        return tf.where(valid, rewards, hparams.penalty_reward_value * tf.ones_like(rewards))

    @utils.graph_function(
        queries=Schema(tf.int32, (None, None)),
        responses=Schema(tf.int32, (None, None))
    )
    def unpenalized_score_fn(queries, responses):
        return score_model.score_fn(queries, responses)
	
	# 打分函数
    def score_fn(queries, responses): 
        responses = postprocess(responses)
        score = penalize(responses, unpenalized_score_fn(queries, responses))
        return score, responses, dict(score=score)
    score_fn.stat_schemas = dict(score=Schema(tf.float32, (None,)))
    return score_fn

折现因子 (discount factor) = 1

• 折现因子 (\gamma) 设置为 1 ，这意味着未来的奖励与即时奖励具有相同的权重。

PPO 中的批次和小批次

1. 额外添加了 micro_batch_size 来帮助处理梯度累积的情况

import numpy as np
batch_size = 8
nminibatches = 2
gradient_accumulation_steps = 2
mini_batch_size = batch_size // nminibatches
micro_batch_size = mini_batch_size // gradient_accumulation_steps
data = np.arange(batch_size).astype(np.float32)
print("data:", data)
print("batch_size:", batch_size)
print("mini_batch_size:", mini_batch_size)
print("micro_batch_size:", micro_batch_size)
for epoch in range(4):
    batch_inds = np.random.permutation(batch_size)
    print("epoch:", epoch, "batch_inds:", batch_inds)
    for mini_batch_start in range(0, batch_size, mini_batch_size):
        mini_batch_end = mini_batch_start + mini_batch_size
        mini_batch_inds = batch_inds[mini_batch_start:mini_batch_end]
        
        # `optimizer.zero_grad()` set optimizer to zero for gradient accumulation
        for micro_batch_start in range(0, mini_batch_size, micro_batch_size):
            micro_batch_end = micro_batch_start + micro_batch_size
            micro_batch_inds = mini_batch_inds[micro_batch_start:micro_batch_end]
            print("____⏩ a forward pass on", data[micro_batch_inds])
        # `optimizer.step()`
        print("⏪ a backward pass on", data[mini_batch_inds])

# data: [0. 1. 2. 3. 4. 5. 6. 7.]
# batch_size: 8
# mini_batch_size: 4
# micro_batch_size: 2
# epoch: 0 batch_inds: [6 4 0 7 3 5 1 2]
# ____⏩ a forward pass on [6. 4.]
# ____⏩ a forward pass on [0. 7.]
# ⏪ a backward pass on [6. 4. 0. 7.]
# ____⏩ a forward pass on [3. 5.]
# ____⏩ a forward pass on [1. 2.]
# ⏪ a backward pass on [3. 5. 1. 2.]
# epoch: 1 batch_inds: [6 7 3 2 0 4 5 1]
# ____⏩ a forward pass on [6. 7.]
# ____⏩ a forward pass on [3. 2.]
# ⏪ a backward pass on [6. 7. 3. 2.]
# ____⏩ a forward pass on [0. 4.]
# ____⏩ a forward pass on [5. 1.]
# ⏪ a backward pass on [0. 4. 5. 1.]
# epoch: 2 batch_inds: [1 4 5 6 0 7 3 2]
# ____⏩ a forward pass on [1. 4.]
# ____⏩ a forward pass on [5. 6.]
# ⏪ a backward pass on [1. 4. 5. 6.]
# ____⏩ a forward pass on [0. 7.]
# ____⏩ a forward pass on [3. 2.]
# ⏪ a backward pass on [0. 7. 3. 2.]
# epoch: 3 batch_inds: [7 2 4 1 3 0 6 5]
# ____⏩ a forward pass on [7. 2.]
# ____⏩ a forward pass on [4. 1.]
# ⏪ a backward pass on [7. 2. 4. 1.]
# ____⏩ a forward pass on [3. 0.]
# ____⏩ a forward pass on [6. 5.]
# ⏪ a backward pass on [3. 0. 6. 5.]

基于每个标记的 KL 惩罚

1. 代码为奖励添加了每个标记的 KL 惩罚，以阻止策略与原始策略差异过大。

def compute_rewards(scores, logprobs, ref_logprobs):
          kl = logprobs - ref_logprobs 
          non_score_reward = -self.kl_ctl.value * kl
          rewards = non_score_reward.copy()
          rewards[:, -1] += scores
          return rewards, non_score_reward, self.kl_ctl.value
           
 self.compute_rewards = compute_rewards

以 “usually, he would” 为例，它被标记化为 [23073, 11, 339, 561] 。假设我们使用 [23073] 作为查询，[11, 339, 561] 作为响应。然后在默认的 gpt2 参数下，响应标记将具有参考策略的对数概率 logprobs=[-3.3213, -4.9980, -3.8690] 。

• 在第一个 PPO 更新时期和小批次更新时，激活策略将具有相同的对数概率new_logprobs=[-3.3213, -4.9980, -3.8690] 。因此，每个标记的 KL 惩罚将为 kl = new_logprobs - logprobs = [0., 0., 0.] 。
• 但是，在第一个梯度反向传播后，我们可能会得到 new_logprob=[3.3213, -4.9980, -3.8690] ，因此每个标记的 KL 惩罚变为 kl = new_logprobs - logprobs = [-0.3315, -0.0426, 0.6351] 。
• 随后，non_score_reward = beta * kl ，其中 beta 是 KL 惩罚系数 (\beta)，它被添加到从奖励模型获得的 score 中，以创建用于训练的 rewards 。
• score 仅在每个回合 ( episode ) 结束时给出，可能类似于 [0.4] ，然后我们有 rewards = [beta * -0.3315, beta * -0.0426, beta * 0.6351 + 0.4] 。

每个小批次的奖励和优势白化，可选择均值平移

1. 名为 whiten 的函数，如下所示，基本上通过减去其均值然后除以其标准差来对 values 进行归一化。可选地，whiten 可以通过 shift_mean=True 将白化后的 values 平移到均值。

def whiten(values, shift_mean=True):
    mean, var = torch.mean(values), torch.var(values, unbiased=False)
    whitened = (values - mean)* torch.rsqrt(var + 1e-8)
    if not shift_mean:
        whitened += mean
    return whitened

2. 在每个小批次中，OAI 使用 whiten(rewards, shift_mean=False) 对奖励进行白化，不对均值进行平移处理 ，并使用平移后的均值对优势进行白化 whiten(advantages)

def loss(self, rollouts):
        values = rollouts['values']
        old_logprob = rollouts['logprobs']
        rewards = rollouts['rewards']
        with tf.name_scope('ppo_loss'):
            if self.hparams.ppo.whiten_rewards:
                rewards = utils.whiten(rewards, shift_mean=False) # 奖励白化

            lastgaelam = 0
            advantages_reversed = []
            gen_length = self.hparams.task.response_length
            for t in reversed(range(gen_length)):
                nextvalues = values[:, t + 1] if t < gen_length - 1 else 0.0
                delta = rewards[:, t] + self.hparams.ppo.gamma * nextvalues - values[:, t]
                lastgaelam = delta + self.hparams.ppo.gamma * self.hparams.ppo.lam * lastgaelam
                advantages_reversed.append(lastgaelam)
            advantages = tf.stack(advantages_reversed[::-1], axis=1)
            returns = advantages + values

            advantages = utils.whiten(advantages)
            advantages = tf.stop_gradient(advantages)  # Shouldn't do anything, but better not to think about it

            outputs = self.policy.analyze_responses_op(rollouts['queries'], rollouts['responses'])
			vpred = outputs['values']
            vpredclipped = tf.clip_by_value(vpred, values - self.hparams.ppo.cliprange_value, values + self.hparams.ppo.cliprange_value)
            vf_losses1 = tf.square(vpred - returns)
            vf_losses2 = tf.square(vpredclipped - returns)
            vf_loss = .5 * tf.reduce_mean(tf.maximum(vf_losses1, vf_losses2))
            vf_clipfrac = tf.reduce_mean(tf.cast(tf.greater(vf_losses2, vf_losses1), tf.float32))

            logprob = outputs['logprobs']
            ratio = tf.exp(logprob - old_logprob)
            pg_losses = -advantages * ratio
            pg_losses2 = -advantages * tf.clip_by_value(ratio, 1.0 - self.hparams.ppo.cliprange, 1.0 + self.hparams.ppo.cliprange)
            pg_loss = tf.reduce_mean(tf.maximum(pg_losses, pg_losses2))
            pg_clipfrac = tf.reduce_mean(tf.cast(tf.greater(pg_losses2, pg_losses), tf.float32))

            loss = pg_loss + self.hparams.ppo.vf_coef * vf_loss

            entropy = tf.reduce_mean(outputs['entropies'])
            approxkl = .5 * tf.reduce_mean(tf.square(logprob - old_logprob))

            return_mean, return_var = tf.nn.moments(returns, axes=list(range(returns.shape.ndims)))
            value_mean, value_var = tf.nn.moments(values, axes=list(range(values.shape.ndims)))

            stats = dict(
                loss=dict(policy=pg_loss, value=vf_loss, total=loss),
                policy=dict(entropy=entropy, approxkl=approxkl, clipfrac=pg_clipfrac),
                returns=dict(mean=return_mean, var=return_var),
                val=dict(vpred=tf.reduce_mean(vpred), error=tf.reduce_mean((vpred - returns) ** 2),
                         clipfrac=vf_clipfrac, mean=value_mean, var=value_var)
            )
            return loss, utils.flatten_dict(stats, sep='/')

裁剪值函数

vpred = outputs['values']
vpredclipped = tf.clip_by_value(vpred, values - self.hparams.ppo.cliprange_value, values + self.hparams.ppo.cliprange_value)
vf_losses1 = tf.square(vpred - returns)
vf_losses2 = tf.square(vpredclipped - returns)
vf_loss = .5 * tf.reduce_mean(tf.maximum(vf_losses1, vf_losses2))
vf_clipfrac = tf.reduce_mean(tf.cast(tf.greater(vf_losses2, vf_losses1), tf.float32))

• 代码：https://github.com/vwxyzjn/lm-human-preference-details
• blog：https://huggingface.co/blog/zh/the_n_implementation_details_of_rlhf_with_ppo