TensorFlow版本的AdamWeightDecayOptimizer在fine-tune BERT时不起作用，请问您有遇到过吗？

[guotong1988]

"""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！

[ymcui]

没太看明白你说的问题。“不起作用”具体指的是出现什么情况？

[guotong1988]

感谢回复，

就是等于没训练，Loss和Accuracy完全不动，

可能是我的AdamWeightDecayOptimizer抄错了，您要是知道有能用的AdamWeightDecayOptimizer给下就行~~

[ymcui]

用谷歌原版的：https://github.com/google-research/bert/blob/master/optimization.py

[guotong1988]

谢谢！我试试，您发的这个没有必须TPU的限制吧？

[ymcui]

不会的，TPU上用的时候会进入判断：https://github.com/google-research/bert/blob/master/optimization.py#L67 其余代码都是通用的。

[guotong1988]

用了谷歌原版的还是不行， 就是基于谷歌原版把global_step改成自己传进去

self.global_step = tf.Variable(0, name="global_step", trainable=False, dtype=tf.int64)
self.train_op = create_optimizer(self.loss,0.00001,num_train_steps=20000,global_step=self.global_step,num_warmup_steps=320)

这个可以：

self.train_op = tf.train.AdamOptimizer(0.00001).minimize(self.loss, global_step=self.global_step)

[ymcui]

你改成自己定义的global_step的目的是？另外你上面的self.global_step和self.train_op是在哪定义的？

[ymcui]

注意这里：https://github.com/google-research/bert/blob/master/optimization.py#L79

[guotong1988]

感谢回复

self.train_op是create_optimizer返回的

global_step原来有两个所以报错

[guotong1988]

看到了

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.

[guotong1988]

感谢回复，好像不是global_step的问题

[guotong1988]

方便的话，您那试一下其实也挺方便，跪谢了

[ymcui]

抱歉，这个问题与本repo没有什么联系，属于TensorFlow的问题。 建议你阅读一下的tf.train.get_or_create_global_step()的源代码，或许会有收获。 https://github.com/tensorflow/tensorflow/blob/v2.3.0/tensorflow/python/training/training_util.py#L148

[guotong1988]

感谢回复！

[guotong1988]

我再看看