Pre-Training with Whole Word Masking for Chinese BERT

"""Functions and classes related to optimization (weight updates)."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import re
import tensorflow as tf
from tensorflow.python.framework import ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.training import optimizer
from tensorflow.python.ops import state_ops
from tensorflow.python.ops import resource_variable_ops


def create_optimizer(loss, init_lr, num_train_steps, num_warmup_steps, use_tpu):
    """Creates an optimizer training op."""
    global_step = tf.train.get_or_create_global_step()

    learning_rate = tf.constant(value=init_lr, shape=[], dtype=tf.float32)

    # Implements linear decay of the learning rate.
    learning_rate = tf.train.polynomial_decay(
        learning_rate,
        global_step,
        num_train_steps,
        end_learning_rate=0.0,
        power=1.0,
        cycle=False)

    # Implements linear warmup. I.e., if global_step < num_warmup_steps, the
    # learning rate will be `global_step/num_warmup_steps * init_lr`.
    if num_warmup_steps:
        global_steps_int = tf.cast(global_step, tf.int32)
        warmup_steps_int = tf.constant(num_warmup_steps, dtype=tf.int32)

        global_steps_float = tf.cast(global_steps_int, tf.float32)
        warmup_steps_float = tf.cast(warmup_steps_int, tf.float32)

        warmup_percent_done = global_steps_float / warmup_steps_float
        warmup_learning_rate = init_lr * warmup_percent_done

        is_warmup = tf.cast(global_steps_int < warmup_steps_int, tf.float32)
        learning_rate = (
                (1.0 - is_warmup) * learning_rate + is_warmup * warmup_learning_rate)

    # It is recommended that you use this optimizer for fine tuning, since this
    # is how the model was trained (note that the Adam m/v variables are NOT
    # loaded from init_checkpoint.)
    optimizer = AdamWeightDecayOptimizer(
        learning_rate=learning_rate,
        weight_decay_rate=0.01,
        beta_1=0.9,
        beta_2=0.999,
        epsilon=1e-6,
        exclude_from_weight_decay=["LayerNorm", "layer_norm", "bias"])

    if use_tpu:
        optimizer = tf.contrib.tpu.CrossShardOptimizer(optimizer)

    tvars = tf.trainable_variables()
    grads = tf.gradients(loss, tvars)

    # This is how the model was pre-trained.
    (grads, _) = tf.clip_by_global_norm(grads, clip_norm=1.0)

    train_op = optimizer.apply_gradients(
        zip(grads, tvars), global_step=global_step)

    # Normally the global step update is done inside of `apply_gradients`.
    # However, `AdamWeightDecayOptimizer` doesn't do this. But if you use
    # a different optimizer, you should probably take this line out.
    new_global_step = global_step + 1
    train_op = tf.group(train_op, [global_step.assign(new_global_step)])
    return train_op

class AdamWeightDecayOptimizer(optimizer.Optimizer):
    """A basic Adam optimizer that includes "correct" L2 weight decay."""

    def __init__(self,
                 learning_rate,
                 weight_decay_rate=0.0,
                 beta_1=0.9,
                 beta_2=0.999,
                 epsilon=1e-6,
                 exclude_from_weight_decay=None,
                 name="AdamWeightDecayOptimizer"):
        """Constructs a AdamWeightDecayOptimizer."""
        super(AdamWeightDecayOptimizer, self).__init__(False, name)

        self.learning_rate = learning_rate
        self.weight_decay_rate = weight_decay_rate
        self.beta_1 = beta_1
        self.beta_2 = beta_2
        self.epsilon = epsilon
        self.exclude_from_weight_decay = exclude_from_weight_decay

    def _prepare(self):
        self.learning_rate_t = ops.convert_to_tensor(
            self.learning_rate, name='learning_rate')
        self.weight_decay_rate_t = ops.convert_to_tensor(
            self.weight_decay_rate, name='weight_decay_rate')
        self.beta_1_t = ops.convert_to_tensor(self.beta_1, name='beta_1')
        self.beta_2_t = ops.convert_to_tensor(self.beta_2, name='beta_2')
        self.epsilon_t = ops.convert_to_tensor(self.epsilon, name='epsilon')

    def _create_slots(self, var_list):
        for v in var_list:
            self._zeros_slot(v, 'm', self._name)
            self._zeros_slot(v, 'v', self._name)

    def _apply_dense(self, grad, var):
        learning_rate_t = math_ops.cast(
            self.learning_rate_t, var.dtype.base_dtype)
        beta_1_t = math_ops.cast(self.beta_1_t, var.dtype.base_dtype)
        beta_2_t = math_ops.cast(self.beta_2_t, var.dtype.base_dtype)
        epsilon_t = math_ops.cast(self.epsilon_t, var.dtype.base_dtype)
        weight_decay_rate_t = math_ops.cast(
            self.weight_decay_rate_t, var.dtype.base_dtype)

        m = self.get_slot(var, 'm')
        v = self.get_slot(var, 'v')

        # Standard Adam update.
        next_m = (
            tf.multiply(beta_1_t, m) +
            tf.multiply(1.0 - beta_1_t, grad))
        next_v = (
            tf.multiply(beta_2_t, v) + tf.multiply(1.0 - beta_2_t,
                                                   tf.square(grad)))

        update = next_m / (tf.sqrt(next_v) + epsilon_t)

        if self._do_use_weight_decay(var.name):
            update += weight_decay_rate_t * var

        update_with_lr = learning_rate_t * update

        next_param = var - update_with_lr

        return control_flow_ops.group(*[var.assign(next_param),
                                        m.assign(next_m),
                                        v.assign(next_v)])

    def _resource_apply_dense(self, grad, var):
        learning_rate_t = math_ops.cast(
            self.learning_rate_t, var.dtype.base_dtype)
        beta_1_t = math_ops.cast(self.beta_1_t, var.dtype.base_dtype)
        beta_2_t = math_ops.cast(self.beta_2_t, var.dtype.base_dtype)
        epsilon_t = math_ops.cast(self.epsilon_t, var.dtype.base_dtype)
        weight_decay_rate_t = math_ops.cast(
            self.weight_decay_rate_t, var.dtype.base_dtype)

        m = self.get_slot(var, 'm')
        v = self.get_slot(var, 'v')

        # Standard Adam update.
        next_m = (
            tf.multiply(beta_1_t, m) +
            tf.multiply(1.0 - beta_1_t, grad))
        next_v = (
            tf.multiply(beta_2_t, v) + tf.multiply(1.0 - beta_2_t,
                                                   tf.square(grad)))

        update = next_m / (tf.sqrt(next_v) + epsilon_t)

        if self._do_use_weight_decay(var.name):
            update += weight_decay_rate_t * var

        update_with_lr = learning_rate_t * update

        next_param = var - update_with_lr

        return control_flow_ops.group(*[var.assign(next_param),
                                        m.assign(next_m),
                                        v.assign(next_v)])

    def _apply_sparse_shared(self, grad, var, indices, scatter_add):
        learning_rate_t = math_ops.cast(
            self.learning_rate_t, var.dtype.base_dtype)
        beta_1_t = math_ops.cast(self.beta_1_t, var.dtype.base_dtype)
        beta_2_t = math_ops.cast(self.beta_2_t, var.dtype.base_dtype)
        epsilon_t = math_ops.cast(self.epsilon_t, var.dtype.base_dtype)
        weight_decay_rate_t = math_ops.cast(
            self.weight_decay_rate_t, var.dtype.base_dtype)

        m = self.get_slot(var, 'm')
        v = self.get_slot(var, 'v')

        m_t = state_ops.assign(m, m * beta_1_t,
                               use_locking=self._use_locking)

        m_scaled_g_values = grad * (1 - beta_1_t)
        with ops.control_dependencies([m_t]):
            m_t = scatter_add(m, indices, m_scaled_g_values)

        v_scaled_g_values = (grad * grad) * (1 - beta_2_t)
        v_t = state_ops.assign(v, v * beta_2_t, use_locking=self._use_locking)
        with ops.control_dependencies([v_t]):
            v_t = scatter_add(v, indices, v_scaled_g_values)

        update = m_t / (math_ops.sqrt(v_t) + epsilon_t)

        if self._do_use_weight_decay(var.name):
            update += weight_decay_rate_t * var

        update_with_lr = learning_rate_t * update

        var_update = state_ops.assign_sub(var,
                                          update_with_lr,
                                          use_locking=self._use_locking)
        return control_flow_ops.group(*[var_update, m_t, v_t])

    def _apply_sparse(self, grad, var):
        return self._apply_sparse_shared(
            grad.values, var, grad.indices,
            lambda x, i, v: state_ops.scatter_add(  # pylint: disable=g-long-lambda
                x, i, v, use_locking=self._use_locking))

    def _resource_scatter_add(self, x, i, v):
        with ops.control_dependencies(
            [resource_variable_ops.resource_scatter_add(
                x.handle, i, v)]):
            return x.value()

    def _resource_apply_sparse(self, grad, var, indices):
        return self._apply_sparse_shared(
            grad, var, indices, self._resource_scatter_add)

    def _do_use_weight_decay(self, param_name):
        """Whether to use L2 weight decay for `param_name`."""
        if not self.weight_decay_rate:
            return False
        if self.exclude_from_weight_decay:
            for r in self.exclude_from_weight_decay:
                if re.search(r, param_name) is not None:
                    return False
        return True

@ymcui 多谢多谢多谢！

Chinese-BERT-wwm对于Masked LM 任务（完形填空，ERNIE中提到的那种），表现怎么样？有测试过么？ 另外WWM采取了新的Mask策略，如果我想在此基础上增加一些专业领域词汇做pre-training，该怎么做？BERT原声模型中有相应的代码，但是因为两个模型mask策略不同，应该不能直接用原生BERT模型，能不能开源一下修改后的模型呢？

seqeval的评估值有micro avg和 macro avg，报告里写的是那个值？ precision recall f1-score support

   MISC       0.00      0.00      0.00         1
    PER       1.00      1.00      1.00         1

micro avg 0.50 0.50 0.50 2 macro avg 0.50 0.50 0.50 2

你好，我们的BERT-wwm-ext（tf）模型做ner和情感分类任务，在相同的配置也指定了随机种子的情况下，两次运行结果不同，换了google的BERT或roBERTa都不会出现这个问题。用部分样本训练发现，样本不多（20条）时结果是一样的，但训练样本大以后，结果就不同了。不知道什么原因，谢谢你的解答。

您好，我有几个问题，想咨询一下： 1、中文使用WWM预训练时，会将"公司" tokenize 为 [“公”,"##司"]，mask时，会将"公司"mask为“[[MASK],[MASK]]”，请问此时label是["公","司"]还是["公","##司"]？ 2、如果label使用的是["公","##司"]来训练的话，那么训练的就是“##司”，而不是“司”；那么在后续模型的使用时，使用bert的BertTokenizer，只会将"公司"tokenize为[“公”,"司"]，但是按照WWM的思想，应该要tokenize 为 [“公”,"##司"]。“##司”和“司”二者的embeddings是不一样的。后续模型的使用，是不是也需要分词？ 麻烦帮忙解答一下！！！谢谢

在你们的工作中，比如mask词的时候，一个词为哈利波特，那么在你们的方法中，是不是只要这个词被mask，那一定是[mask][mask][mask][mask]的形式，还是偶尔会出现[mask]利[mask][mask]的形式，不知道你们是如何设置的（不考虑那个mask80%10%10%的那个随机概率），如果是前者，那么这种完全避免局部共现的设置会不会对结果有影响。

您好，请问加载TFBertModel时出现 NotImplementedError: Weights may only be loaded based on topology into Models when loading TensorFlow-formatted weights (got by_name=True to load_weights)报错是什么原因导致的呢？

ValueError: Couldn't find 'checkpoint' file or checkpoints in given directory chinese_roberta_wwm_ext_L-12_H-768_A-12

尝试用 https://github.com/huggingface/pytorch-transformers/blob/master/pytorch_transformers/convert_tf_checkpoint_to_pytorch.py 这个代码convert roberta的checkpoint，报了这个错，请问是为什么呢？

把Bert、和Bert-wwm的模型转成paddle, 基于 https://github.com/PaddlePaddle/LARK/tree/develop/BERT 的代码在DRCD数据集上finetune，配置跟论文一致，评估脚本：https://github.com/ymcui/CMRC2018-DRCD-BERT/blob/master/cmrc2018_evaluate.py 。效果如下(十次平均)：

复现的结果普遍比论文结果高2~3个点，是不是论文实现的阅读理解有问题？

您好，在readme中的快速加载有这样一句：

注意：本目录中的所有模型均使用BertTokenizer以及BertModel加载，请勿使用RobertaTokenizer/RobertaModel！

我从huggingface官网下下来了roberta-chinese-wwm的pytorch_model.bin，tokenizer.json和vocab，再想在代码使用.from_pretrained()方法引用，即不是采用快速加载的情况，请问这样的话是用RobertaModel引用还是BertModel引用呢？我尝试用BertModel引用，但是提示:

You are using a model of type roberta to instantiate a model of type bert. This is not supported for all configurations of models and can yield errors.`

还请解答一下，非常感谢！

您好，非常感谢您这边的工作。 我最近在 transformers 中实现了 WWM 的 Fine-tune，相关 PR 已经被 merge 到 master 分支， 具体可以参见：这里。

具体的实现参考了 #13 中的描述：对于BERT 的分词结果，使用 LTP 的分词结果，对特定位置的索引添加 ##，后面 mask 部分代码就可以直接套用 Google 代码。

有两个细节上的小问题，希望您不吝解答：

1. LTP 的分词结果是否是动态的 ？因为我看是分词结果是有隐层向量输出的，而不是像 Jieba 一样是基于词表。
2. 如果想要 Fine-tune，对数据的清洗应该做到什么程度，或者是否可以提供一下参考的文献？