docs/lroptimize/optimization.html

<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1, minimum-scale=1" />
<meta name="generator" content="pdoc 0.9.2" />
<title>ktrain.lroptimize.optimization API documentation</title>
<meta name="description" content="Functions and classes related to optimization (weight updates)." />
<link rel="preload stylesheet" as="style" href="https://cdnjs.cloudflare.com/ajax/libs/10up-sanitize.css/11.0.1/sanitize.min.css" integrity="sha256-PK9q560IAAa6WVRRh76LtCaI8pjTJ2z11v0miyNNjrs=" crossorigin>
<link rel="preload stylesheet" as="style" href="https://cdnjs.cloudflare.com/ajax/libs/10up-sanitize.css/11.0.1/typography.min.css" integrity="sha256-7l/o7C8jubJiy74VsKTidCy1yBkRtiUGbVkYBylBqUg=" crossorigin>
<link rel="stylesheet preload" as="style" href="https://cdnjs.cloudflare.com/ajax/libs/highlight.js/10.1.1/styles/github.min.css" crossorigin>
<style>:root{--highlight-color:#fe9}.flex{display:flex !important}body{line-height:1.5em}#content{padding:20px}#sidebar{padding:30px;overflow:hidden}#sidebar > *:last-child{margin-bottom:2cm}.http-server-breadcrumbs{font-size:130%;margin:0 0 15px 0}#footer{font-size:.75em;padding:5px 30px;border-top:1px solid #ddd;text-align:right}#footer p{margin:0 0 0 1em;display:inline-block}#footer p:last-child{margin-right:30px}h1,h2,h3,h4,h5{font-weight:300}h1{font-size:2.5em;line-height:1.1em}h2{font-size:1.75em;margin:1em 0 .50em 0}h3{font-size:1.4em;margin:25px 0 10px 0}h4{margin:0;font-size:105%}h1:target,h2:target,h3:target,h4:target,h5:target,h6:target{background:var(--highlight-color);padding:.2em 0}a{color:#058;text-decoration:none;transition:color .3s ease-in-out}a:hover{color:#e82}.title code{font-weight:bold}h2[id^="header-"]{margin-top:2em}.ident{color:#900}pre code{background:#f8f8f8;font-size:.8em;line-height:1.4em}code{background:#f2f2f1;padding:1px 4px;overflow-wrap:break-word}h1 code{background:transparent}pre{background:#f8f8f8;border:0;border-top:1px solid #ccc;border-bottom:1px solid #ccc;margin:1em 0;padding:1ex}#http-server-module-list{display:flex;flex-flow:column}#http-server-module-list div{display:flex}#http-server-module-list dt{min-width:10%}#http-server-module-list p{margin-top:0}.toc ul,#index{list-style-type:none;margin:0;padding:0}#index code{background:transparent}#index h3{border-bottom:1px solid #ddd}#index ul{padding:0}#index h4{margin-top:.6em;font-weight:bold}@media (min-width:200ex){#index .two-column{column-count:2}}@media (min-width:300ex){#index .two-column{column-count:3}}dl{margin-bottom:2em}dl dl:last-child{margin-bottom:4em}dd{margin:0 0 1em 3em}#header-classes + dl > dd{margin-bottom:3em}dd dd{margin-left:2em}dd p{margin:10px 0}.name{background:#eee;font-weight:bold;font-size:.85em;padding:5px 10px;display:inline-block;min-width:40%}.name:hover{background:#e0e0e0}dt:target .name{background:var(--highlight-color)}.name > span:first-child{white-space:nowrap}.name.class > span:nth-child(2){margin-left:.4em}.inherited{color:#999;border-left:5px solid #eee;padding-left:1em}.inheritance em{font-style:normal;font-weight:bold}.desc h2{font-weight:400;font-size:1.25em}.desc h3{font-size:1em}.desc dt code{background:inherit}.source summary,.git-link-div{color:#666;text-align:right;font-weight:400;font-size:.8em;text-transform:uppercase}.source summary > *{white-space:nowrap;cursor:pointer}.git-link{color:inherit;margin-left:1em}.source pre{max-height:500px;overflow:auto;margin:0}.source pre code{font-size:12px;overflow:visible}.hlist{list-style:none}.hlist li{display:inline}.hlist li:after{content:',\2002'}.hlist li:last-child:after{content:none}.hlist .hlist{display:inline;padding-left:1em}img{max-width:100%}td{padding:0 .5em}.admonition{padding:.1em .5em;margin-bottom:1em}.admonition-title{font-weight:bold}.admonition.note,.admonition.info,.admonition.important{background:#aef}.admonition.todo,.admonition.versionadded,.admonition.tip,.admonition.hint{background:#dfd}.admonition.warning,.admonition.versionchanged,.admonition.deprecated{background:#fd4}.admonition.error,.admonition.danger,.admonition.caution{background:lightpink}</style>
<style media="screen and (min-width: 700px)">@media screen and (min-width:700px){#sidebar{width:30%;height:100vh;overflow:auto;position:sticky;top:0}#content{width:70%;max-width:100ch;padding:3em 4em;border-left:1px solid #ddd}pre code{font-size:1em}.item .name{font-size:1em}main{display:flex;flex-direction:row-reverse;justify-content:flex-end}.toc ul ul,#index ul{padding-left:1.5em}.toc > ul > li{margin-top:.5em}}</style>
<style media="print">@media print{#sidebar h1{page-break-before:always}.source{display:none}}@media print{*{background:transparent !important;color:#000 !important;box-shadow:none !important;text-shadow:none !important}a[href]:after{content:" (" attr(href) ")";font-size:90%}a[href][title]:after{content:none}abbr[title]:after{content:" (" attr(title) ")"}.ir a:after,a[href^="javascript:"]:after,a[href^="#"]:after{content:""}pre,blockquote{border:1px solid #999;page-break-inside:avoid}thead{display:table-header-group}tr,img{page-break-inside:avoid}img{max-width:100% !important}@page{margin:0.5cm}p,h2,h3{orphans:3;widows:3}h1,h2,h3,h4,h5,h6{page-break-after:avoid}}</style>
<script defer src="https://cdnjs.cloudflare.com/ajax/libs/highlight.js/10.1.1/highlight.min.js" integrity="sha256-Uv3H6lx7dJmRfRvH8TH6kJD1TSK1aFcwgx+mdg3epi8=" crossorigin></script>
<script>window.addEventListener('DOMContentLoaded', () => hljs.initHighlighting())</script>
</head>
<body>
<main>
<article id="content">
<header>
<h1 class="title">Module <code>ktrain.lroptimize.optimization</code></h1>
</header>
<section id="section-intro">
<p>Functions and classes related to optimization (weight updates).</p>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python"># Copyright 2019 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the &#34;License&#34;);
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an &#34;AS IS&#34; BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
&#34;&#34;&#34;Functions and classes related to optimization (weight updates).&#34;&#34;&#34;

import re
from typing import Callable, List, Optional, Union

import tensorflow as tf


class WarmUp(tf.keras.optimizers.schedules.LearningRateSchedule):
    &#34;&#34;&#34;
    ```
    Applies a warmup schedule on a given learning rate decay schedule.
    Args:
        initial_learning_rate (:obj:`float`):
            The initial learning rate for the schedule after the warmup (so this will be the learning rate at the end
            of the warmup).
        decay_schedule_fn (:obj:`Callable`):
            The schedule function to apply after the warmup for the rest of training.
        warmup_steps (:obj:`int`):
            The number of steps for the warmup part of training.
        power (:obj:`float`, `optional`, defaults to 1):
            The power to use for the polynomial warmup (defaults is a linear warmup).
        name (:obj:`str`, `optional`):
            Optional name prefix for the returned tensors during the schedule.
    ```
    &#34;&#34;&#34;

    def __init__(
        self,
        initial_learning_rate: float,
        decay_schedule_fn: Callable,
        warmup_steps: int,
        power: float = 1.0,
        name: str = None,
    ):
        super().__init__()
        self.initial_learning_rate = initial_learning_rate
        self.warmup_steps = warmup_steps
        self.power = power
        self.decay_schedule_fn = decay_schedule_fn
        self.name = name

    def __call__(self, step):
        with tf.name_scope(self.name or &#34;WarmUp&#34;) as name:
            # Implements polynomial warmup. i.e., if global_step &lt; warmup_steps, the
            # learning rate will be `global_step/num_warmup_steps * init_lr`.
            global_step_float = tf.cast(step, tf.float32)
            warmup_steps_float = tf.cast(self.warmup_steps, tf.float32)
            warmup_percent_done = global_step_float / warmup_steps_float
            warmup_learning_rate = self.initial_learning_rate * tf.math.pow(
                warmup_percent_done, self.power
            )
            return tf.cond(
                global_step_float &lt; warmup_steps_float,
                lambda: warmup_learning_rate,
                lambda: self.decay_schedule_fn(step - self.warmup_steps),
                name=name,
            )

    def get_config(self):
        return {
            &#34;initial_learning_rate&#34;: self.initial_learning_rate,
            &#34;decay_schedule_fn&#34;: self.decay_schedule_fn,
            &#34;warmup_steps&#34;: self.warmup_steps,
            &#34;power&#34;: self.power,
            &#34;name&#34;: self.name,
        }


def create_optimizer(
    init_lr: float,
    num_train_steps: int,
    num_warmup_steps: int,
    min_lr_ratio: float = 0.0,
    adam_epsilon: float = 1e-8,
    weight_decay_rate: float = 0.0,
    include_in_weight_decay: Optional[List[str]] = None,
):
    &#34;&#34;&#34;
    ```
    Creates an optimizer with a learning rate schedule using a warmup phase followed by a linear decay.
    Args:
        init_lr (:obj:`float`):
            The desired learning rate at the end of the warmup phase.
        num_train_step (:obj:`int`):
            The total number of training steps.
        num_warmup_steps (:obj:`int`):
            The number of warmup steps.
        min_lr_ratio (:obj:`float`, `optional`, defaults to 0):
            The final learning rate at the end of the linear decay will be :obj:`init_lr * min_lr_ratio`.
        adam_epsilon (:obj:`float`, `optional`, defaults to 1e-8):
            The epsilon to use in Adam.
        weight_decay_rate (:obj:`float`, `optional`, defaults to 0):
            The weight decay to use.
        include_in_weight_decay (:obj:`List[str]`, `optional`):
            List of the parameter names (or re patterns) to apply weight decay to. If none is passed, weight decay is
            applied to all parameters except bias and layer norm parameters.
    ```
    &#34;&#34;&#34;
    # Implements linear decay of the learning rate.
    lr_schedule = tf.keras.optimizers.schedules.PolynomialDecay(
        initial_learning_rate=init_lr,
        decay_steps=num_train_steps - num_warmup_steps,
        end_learning_rate=init_lr * min_lr_ratio,
    )
    if num_warmup_steps:
        lr_schedule = WarmUp(
            initial_learning_rate=init_lr,
            decay_schedule_fn=lr_schedule,
            warmup_steps=num_warmup_steps,
        )
    if weight_decay_rate &gt; 0.0:
        optimizer = AdamWeightDecay(
            learning_rate=lr_schedule,
            weight_decay_rate=weight_decay_rate,
            beta_1=0.9,
            beta_2=0.999,
            epsilon=adam_epsilon,
            exclude_from_weight_decay=[&#34;LayerNorm&#34;, &#34;layer_norm&#34;, &#34;bias&#34;],
            include_in_weight_decay=include_in_weight_decay,
        )
    else:
        optimizer = tf.keras.optimizers.Adam(
            learning_rate=lr_schedule, epsilon=adam_epsilon
        )
    # We return the optimizer and the LR scheduler in order to better track the
    # evolution of the LR independently of the optimizer.
    return optimizer, lr_schedule


class AdamWeightDecay(tf.keras.optimizers.Adam):
    &#34;&#34;&#34;
    ```
    Adam enables L2 weight decay and clip_by_global_norm on gradients. Just adding the square of the weights to the
    loss function is *not* the correct way of using L2 regularization/weight decay with Adam, since that will interact
    with the m and v parameters in strange ways as shown in
    `Decoupled Weight Decay Regularization &lt;https://arxiv.org/abs/1711.05101&gt;`__.
    Instead we want ot decay the weights in a manner that doesn&#39;t interact with the m/v parameters. This is equivalent
    to adding the square of the weights to the loss with plain (non-momentum) SGD.
    Args:
        learning_rate (:obj:`Union[float, tf.keras.optimizers.schedules.LearningRateSchedule]`, `optional`, defaults to 1e-3):
            The learning rate to use or a schedule.
        beta_1 (:obj:`float`, `optional`, defaults to 0.9):
            The beta1 parameter in Adam, which is the exponential decay rate for the 1st momentum estimates.
        beta_2 (:obj:`float`, `optional`, defaults to 0.999):
            The beta2 parameter in Adam, which is the exponential decay rate for the 2nd momentum estimates.
        epsilon (:obj:`float`, `optional`, defaults to 1e-7):
            The epsilon paramenter in Adam, which is a small constant for numerical stability.
        amsgrad (:obj:`bool`, `optional`, default to `False`):
            Wheter to apply AMSGrad varient of this algorithm or not, see
            `On the Convergence of Adam and Beyond &lt;https://arxiv.org/abs/1904.09237&gt;`__.
        weight_decay_rate (:obj:`float`, `optional`, defaults to 0):
            The weight decay to apply.
        include_in_weight_decay (:obj:`List[str]`, `optional`):
            List of the parameter names (or re patterns) to apply weight decay to. If none is passed, weight decay is
            applied to all parameters by default (unless they are in :obj:`exclude_from_weight_decay`).
        exclude_from_weight_decay (:obj:`List[str]`, `optional`):
            List of the parameter names (or re patterns) to exclude from applying weight decay to. If a
            :obj:`include_in_weight_decay` is passed, the names in it will supersede this list.
        name (:obj:`str`, `optional`, defaults to &#39;AdamWeightDecay&#39;):
            Optional name for the operations created when applying gradients.
        kwargs:
            Keyward arguments. Allowed to be {``clipnorm``, ``clipvalue``, ``lr``, ``decay``}. ``clipnorm`` is clip
            gradients by norm; ``clipvalue`` is clip gradients by value, ``decay`` is included for backward
            compatibility to allow time inverse decay of learning rate. ``lr`` is included for backward compatibility,
            recommended to use ``learning_rate`` instead.
    ```
    &#34;&#34;&#34;

    def __init__(
        self,
        learning_rate: Union[
            float, tf.keras.optimizers.schedules.LearningRateSchedule
        ] = 0.001,
        beta_1: float = 0.9,
        beta_2: float = 0.999,
        epsilon: float = 1e-7,
        amsgrad: bool = False,
        weight_decay_rate: float = 0.0,
        include_in_weight_decay: Optional[List[str]] = None,
        exclude_from_weight_decay: Optional[List[str]] = None,
        name: str = &#34;AdamWeightDecay&#34;,
        **kwargs
    ):
        super().__init__(
            learning_rate, beta_1, beta_2, epsilon, amsgrad, name, **kwargs
        )
        self.weight_decay_rate = weight_decay_rate
        self._include_in_weight_decay = include_in_weight_decay
        self._exclude_from_weight_decay = exclude_from_weight_decay

    @classmethod
    def from_config(cls, config):
        &#34;&#34;&#34;Creates an optimizer from its config with WarmUp custom object.&#34;&#34;&#34;
        custom_objects = {&#34;WarmUp&#34;: WarmUp}
        return super(AdamWeightDecay, cls).from_config(
            config, custom_objects=custom_objects
        )

    def _prepare_local(self, var_device, var_dtype, apply_state):
        super(AdamWeightDecay, self)._prepare_local(var_device, var_dtype, apply_state)
        apply_state[(var_device, var_dtype)][&#34;weight_decay_rate&#34;] = tf.constant(
            self.weight_decay_rate, name=&#34;adam_weight_decay_rate&#34;
        )

    def _decay_weights_op(self, var, learning_rate, apply_state):
        do_decay = self._do_use_weight_decay(var.name)
        if do_decay:
            return var.assign_sub(
                learning_rate
                * var
                * apply_state[(var.device, var.dtype.base_dtype)][&#34;weight_decay_rate&#34;],
                use_locking=self._use_locking,
            )
        return tf.no_op()

    def apply_gradients(self, grads_and_vars, name=None, **kwargs):
        # 2020-07-05: added **kwargs based on https://github.com/tensorflow/addons/issues/1267
        grads, tvars = list(zip(*grads_and_vars))
        return super(AdamWeightDecay, self).apply_gradients(
            zip(grads, tvars),
            name=name,
        )

    def _get_lr(self, var_device, var_dtype, apply_state):
        &#34;&#34;&#34;Retrieves the learning rate with the given state.&#34;&#34;&#34;
        if apply_state is None:
            return self._decayed_lr_t[var_dtype], {}

        apply_state = apply_state or {}
        coefficients = apply_state.get((var_device, var_dtype))
        if coefficients is None:
            coefficients = self._fallback_apply_state(var_device, var_dtype)
            apply_state[(var_device, var_dtype)] = coefficients

        return coefficients[&#34;lr_t&#34;], dict(apply_state=apply_state)

    def _resource_apply_dense(self, grad, var, apply_state=None):
        lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
        decay = self._decay_weights_op(var, lr_t, apply_state)
        with tf.control_dependencies([decay]):
            return super(AdamWeightDecay, self)._resource_apply_dense(
                grad, var, **kwargs
            )

    def _resource_apply_sparse(self, grad, var, indices, apply_state=None):
        lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
        decay = self._decay_weights_op(var, lr_t, apply_state)
        with tf.control_dependencies([decay]):
            return super(AdamWeightDecay, self)._resource_apply_sparse(
                grad, var, indices, **kwargs
            )

    def get_config(self):
        config = super().get_config()
        config.update({&#34;weight_decay_rate&#34;: self.weight_decay_rate})
        return config

    def _do_use_weight_decay(self, param_name):
        &#34;&#34;&#34;Whether to use L2 weight decay for `param_name`.&#34;&#34;&#34;
        if self.weight_decay_rate == 0:
            return False

        if self._include_in_weight_decay:
            for r in self._include_in_weight_decay:
                if re.search(r, param_name) is not None:
                    return True

        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


# Extracted from https://github.com/OpenNMT/OpenNMT-tf/blob/master/opennmt/optimizers/utils.py
class GradientAccumulator(object):
    &#34;&#34;&#34;Gradient accumulation utility.
    When used with a distribution strategy, the accumulator should be called in a
    replica context. Gradients will be accumulated locally on each replica and
    without synchronization. Users should then call ``.gradients``, scale the
    gradients if required, and pass the result to ``apply_gradients``.
    &#34;&#34;&#34;

    # We use the ON_READ synchronization policy so that no synchronization is
    # performed on assignment. To get the value, we call .value() which returns the
    # value on the current replica without synchronization.

    def __init__(self):
        &#34;&#34;&#34;Initializes the accumulator.&#34;&#34;&#34;
        self._gradients = []
        self._accum_steps = None

    @property
    def step(self):
        &#34;&#34;&#34;Number of accumulated steps.&#34;&#34;&#34;
        if self._accum_steps is None:
            self._accum_steps = tf.Variable(
                tf.constant(0, dtype=tf.int64),
                trainable=False,
                synchronization=tf.VariableSynchronization.ON_READ,
                aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA,
            )

        return self._accum_steps.value()

    @property
    def gradients(self):
        &#34;&#34;&#34;The accumulated gradients on the current replica.&#34;&#34;&#34;
        if not self._gradients:
            raise ValueError(
                &#34;The accumulator should be called first to initialize the gradients&#34;
            )
        return list(
            gradient.value() if gradient is not None else gradient
            for gradient in self._gradients
        )

    def __call__(self, gradients):
        &#34;&#34;&#34;Accumulates :obj:`gradients` on the current replica.&#34;&#34;&#34;
        if not self._gradients:
            _ = self.step  # Create the step variable.
            self._gradients.extend(
                [
                    tf.Variable(
                        tf.zeros_like(gradient),
                        trainable=False,
                        synchronization=tf.VariableSynchronization.ON_READ,
                        aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA,
                    )
                    if gradient is not None
                    else gradient
                    for gradient in gradients
                ]
            )
        if len(gradients) != len(self._gradients):
            raise ValueError(
                &#34;Expected %s gradients, but got %d&#34;
                % (len(self._gradients), len(gradients))
            )

        for accum_gradient, gradient in zip(self._gradients, gradients):
            if accum_gradient is not None and gradient is not None:
                accum_gradient.assign_add(gradient)

        self._accum_steps.assign_add(1)

    def reset(self):
        &#34;&#34;&#34;Resets the accumulated gradients on the current replica.&#34;&#34;&#34;
        if not self._gradients:
            return
        self._accum_steps.assign(0)
        for gradient in self._gradients:
            if gradient is not None:
                gradient.assign(tf.zeros_like(gradient))</code></pre>
</details>
</section>
<section>
</section>
<section>
</section>
<section>
<h2 class="section-title" id="header-functions">Functions</h2>
<dl>
<dt id="ktrain.lroptimize.optimization.create_optimizer"><code class="name flex">
<span>def <span class="ident">create_optimizer</span></span>(<span>init_lr: float, num_train_steps: int, num_warmup_steps: int, min_lr_ratio: float = 0.0, adam_epsilon: float = 1e-08, weight_decay_rate: float = 0.0, include_in_weight_decay: Union[List[str], NoneType] = None)</span>
</code></dt>
<dd>
<div class="desc"><pre><code>Creates an optimizer with a learning rate schedule using a warmup phase followed by a linear decay.
Args:
    init_lr (:obj:`float`):
        The desired learning rate at the end of the warmup phase.
    num_train_step (:obj:`int`):
        The total number of training steps.
    num_warmup_steps (:obj:`int`):
        The number of warmup steps.
    min_lr_ratio (:obj:`float`, `optional`, defaults to 0):
        The final learning rate at the end of the linear decay will be :obj:`init_lr * min_lr_ratio`.
    adam_epsilon (:obj:`float`, `optional`, defaults to 1e-8):
        The epsilon to use in Adam.
    weight_decay_rate (:obj:`float`, `optional`, defaults to 0):
        The weight decay to use.
    include_in_weight_decay (:obj:`List[str]`, `optional`):
        List of the parameter names (or re patterns) to apply weight decay to. If none is passed, weight decay is
        applied to all parameters except bias and layer norm parameters.
</code></pre></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">def create_optimizer(
    init_lr: float,
    num_train_steps: int,
    num_warmup_steps: int,
    min_lr_ratio: float = 0.0,
    adam_epsilon: float = 1e-8,
    weight_decay_rate: float = 0.0,
    include_in_weight_decay: Optional[List[str]] = None,
):
    &#34;&#34;&#34;
    ```
    Creates an optimizer with a learning rate schedule using a warmup phase followed by a linear decay.
    Args:
        init_lr (:obj:`float`):
            The desired learning rate at the end of the warmup phase.
        num_train_step (:obj:`int`):
            The total number of training steps.
        num_warmup_steps (:obj:`int`):
            The number of warmup steps.
        min_lr_ratio (:obj:`float`, `optional`, defaults to 0):
            The final learning rate at the end of the linear decay will be :obj:`init_lr * min_lr_ratio`.
        adam_epsilon (:obj:`float`, `optional`, defaults to 1e-8):
            The epsilon to use in Adam.
        weight_decay_rate (:obj:`float`, `optional`, defaults to 0):
            The weight decay to use.
        include_in_weight_decay (:obj:`List[str]`, `optional`):
            List of the parameter names (or re patterns) to apply weight decay to. If none is passed, weight decay is
            applied to all parameters except bias and layer norm parameters.
    ```
    &#34;&#34;&#34;
    # Implements linear decay of the learning rate.
    lr_schedule = tf.keras.optimizers.schedules.PolynomialDecay(
        initial_learning_rate=init_lr,
        decay_steps=num_train_steps - num_warmup_steps,
        end_learning_rate=init_lr * min_lr_ratio,
    )
    if num_warmup_steps:
        lr_schedule = WarmUp(
            initial_learning_rate=init_lr,
            decay_schedule_fn=lr_schedule,
            warmup_steps=num_warmup_steps,
        )
    if weight_decay_rate &gt; 0.0:
        optimizer = AdamWeightDecay(
            learning_rate=lr_schedule,
            weight_decay_rate=weight_decay_rate,
            beta_1=0.9,
            beta_2=0.999,
            epsilon=adam_epsilon,
            exclude_from_weight_decay=[&#34;LayerNorm&#34;, &#34;layer_norm&#34;, &#34;bias&#34;],
            include_in_weight_decay=include_in_weight_decay,
        )
    else:
        optimizer = tf.keras.optimizers.Adam(
            learning_rate=lr_schedule, epsilon=adam_epsilon
        )
    # We return the optimizer and the LR scheduler in order to better track the
    # evolution of the LR independently of the optimizer.
    return optimizer, lr_schedule</code></pre>
</details>
</dd>
</dl>
</section>
<section>
<h2 class="section-title" id="header-classes">Classes</h2>
<dl>
<dt id="ktrain.lroptimize.optimization.AdamWeightDecay"><code class="flex name class">
<span>class <span class="ident">AdamWeightDecay</span></span>
<span>(</span><span>learning_rate: Union[float, tensorflow.python.keras.optimizer_v2.learning_rate_schedule.LearningRateSchedule] = 0.001, beta_1: float = 0.9, beta_2: float = 0.999, epsilon: float = 1e-07, amsgrad: bool = False, weight_decay_rate: float = 0.0, include_in_weight_decay: Union[List[str], NoneType] = None, exclude_from_weight_decay: Union[List[str], NoneType] = None, name: str = 'AdamWeightDecay', **kwargs)</span>
</code></dt>
<dd>
<div class="desc"><pre><code>Adam enables L2 weight decay and clip_by_global_norm on gradients. Just adding the square of the weights to the
loss function is *not* the correct way of using L2 regularization/weight decay with Adam, since that will interact
with the m and v parameters in strange ways as shown in
`Decoupled Weight Decay Regularization &lt;https://arxiv.org/abs/1711.05101&gt;`__.
Instead we want ot decay the weights in a manner that doesn't interact with the m/v parameters. This is equivalent
to adding the square of the weights to the loss with plain (non-momentum) SGD.
Args:
    learning_rate (:obj:`Union[float, tf.keras.optimizers.schedules.LearningRateSchedule]`, `optional`, defaults to 1e-3):
        The learning rate to use or a schedule.
    beta_1 (:obj:`float`, `optional`, defaults to 0.9):
        The beta1 parameter in Adam, which is the exponential decay rate for the 1st momentum estimates.
    beta_2 (:obj:`float`, `optional`, defaults to 0.999):
        The beta2 parameter in Adam, which is the exponential decay rate for the 2nd momentum estimates.
    epsilon (:obj:`float`, `optional`, defaults to 1e-7):
        The epsilon paramenter in Adam, which is a small constant for numerical stability.
    amsgrad (:obj:`bool`, `optional`, default to `False`):
        Wheter to apply AMSGrad varient of this algorithm or not, see
        `On the Convergence of Adam and Beyond &lt;https://arxiv.org/abs/1904.09237&gt;`__.
    weight_decay_rate (:obj:`float`, `optional`, defaults to 0):
        The weight decay to apply.
    include_in_weight_decay (:obj:`List[str]`, `optional`):
        List of the parameter names (or re patterns) to apply weight decay to. If none is passed, weight decay is
        applied to all parameters by default (unless they are in :obj:`exclude_from_weight_decay`).
    exclude_from_weight_decay (:obj:`List[str]`, `optional`):
        List of the parameter names (or re patterns) to exclude from applying weight decay to. If a
        :obj:`include_in_weight_decay` is passed, the names in it will supersede this list.
    name (:obj:`str`, `optional`, defaults to 'AdamWeightDecay'):
        Optional name for the operations created when applying gradients.
    kwargs:
        Keyward arguments. Allowed to be {``clipnorm``, ``clipvalue``, ``lr``, ``decay``}. ``clipnorm`` is clip
        gradients by norm; ``clipvalue`` is clip gradients by value, ``decay`` is included for backward
        compatibility to allow time inverse decay of learning rate. ``lr`` is included for backward compatibility,
        recommended to use ``learning_rate`` instead.
</code></pre>
<p>Create a new Optimizer.</p>
<p>This must be called by the constructors of subclasses.
Note that Optimizer instances should not bind to a single graph,
and so shouldn't keep Tensors as member variables. Generally
you should be able to use the _set_hyper()/state.get_hyper()
facility instead.</p>
<p>This class in stateful and thread-compatible.</p>
<h2 id="args">Args</h2>
<dl>
<dt><strong><code>name</code></strong></dt>
<dd>A non-empty string.
The name to use for accumulators created
for the optimizer.</dd>
<dt><strong><code>**kwargs</code></strong></dt>
<dd>keyword arguments. Allowed to be {<code>clipnorm</code>, <code>clipvalue</code>, <code>lr</code>,
<code>decay</code>}. <code>clipnorm</code> is clip gradients by norm; <code>clipvalue</code> is clip
gradients by value, <code>decay</code> is included for backward compatibility to
allow time inverse decay of learning rate. <code>lr</code> is included for backward
compatibility, recommended to use <code>learning_rate</code> instead.</dd>
</dl>
<h2 id="raises">Raises</h2>
<dl>
<dt><code>ValueError</code></dt>
<dd>If name is malformed.</dd>
</dl></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">class AdamWeightDecay(tf.keras.optimizers.Adam):
    &#34;&#34;&#34;
    ```
    Adam enables L2 weight decay and clip_by_global_norm on gradients. Just adding the square of the weights to the
    loss function is *not* the correct way of using L2 regularization/weight decay with Adam, since that will interact
    with the m and v parameters in strange ways as shown in
    `Decoupled Weight Decay Regularization &lt;https://arxiv.org/abs/1711.05101&gt;`__.
    Instead we want ot decay the weights in a manner that doesn&#39;t interact with the m/v parameters. This is equivalent
    to adding the square of the weights to the loss with plain (non-momentum) SGD.
    Args:
        learning_rate (:obj:`Union[float, tf.keras.optimizers.schedules.LearningRateSchedule]`, `optional`, defaults to 1e-3):
            The learning rate to use or a schedule.
        beta_1 (:obj:`float`, `optional`, defaults to 0.9):
            The beta1 parameter in Adam, which is the exponential decay rate for the 1st momentum estimates.
        beta_2 (:obj:`float`, `optional`, defaults to 0.999):
            The beta2 parameter in Adam, which is the exponential decay rate for the 2nd momentum estimates.
        epsilon (:obj:`float`, `optional`, defaults to 1e-7):
            The epsilon paramenter in Adam, which is a small constant for numerical stability.
        amsgrad (:obj:`bool`, `optional`, default to `False`):
            Wheter to apply AMSGrad varient of this algorithm or not, see
            `On the Convergence of Adam and Beyond &lt;https://arxiv.org/abs/1904.09237&gt;`__.
        weight_decay_rate (:obj:`float`, `optional`, defaults to 0):
            The weight decay to apply.
        include_in_weight_decay (:obj:`List[str]`, `optional`):
            List of the parameter names (or re patterns) to apply weight decay to. If none is passed, weight decay is
            applied to all parameters by default (unless they are in :obj:`exclude_from_weight_decay`).
        exclude_from_weight_decay (:obj:`List[str]`, `optional`):
            List of the parameter names (or re patterns) to exclude from applying weight decay to. If a
            :obj:`include_in_weight_decay` is passed, the names in it will supersede this list.
        name (:obj:`str`, `optional`, defaults to &#39;AdamWeightDecay&#39;):
            Optional name for the operations created when applying gradients.
        kwargs:
            Keyward arguments. Allowed to be {``clipnorm``, ``clipvalue``, ``lr``, ``decay``}. ``clipnorm`` is clip
            gradients by norm; ``clipvalue`` is clip gradients by value, ``decay`` is included for backward
            compatibility to allow time inverse decay of learning rate. ``lr`` is included for backward compatibility,
            recommended to use ``learning_rate`` instead.
    ```
    &#34;&#34;&#34;

    def __init__(
        self,
        learning_rate: Union[
            float, tf.keras.optimizers.schedules.LearningRateSchedule
        ] = 0.001,
        beta_1: float = 0.9,
        beta_2: float = 0.999,
        epsilon: float = 1e-7,
        amsgrad: bool = False,
        weight_decay_rate: float = 0.0,
        include_in_weight_decay: Optional[List[str]] = None,
        exclude_from_weight_decay: Optional[List[str]] = None,
        name: str = &#34;AdamWeightDecay&#34;,
        **kwargs
    ):
        super().__init__(
            learning_rate, beta_1, beta_2, epsilon, amsgrad, name, **kwargs
        )
        self.weight_decay_rate = weight_decay_rate
        self._include_in_weight_decay = include_in_weight_decay
        self._exclude_from_weight_decay = exclude_from_weight_decay

    @classmethod
    def from_config(cls, config):
        &#34;&#34;&#34;Creates an optimizer from its config with WarmUp custom object.&#34;&#34;&#34;
        custom_objects = {&#34;WarmUp&#34;: WarmUp}
        return super(AdamWeightDecay, cls).from_config(
            config, custom_objects=custom_objects
        )

    def _prepare_local(self, var_device, var_dtype, apply_state):
        super(AdamWeightDecay, self)._prepare_local(var_device, var_dtype, apply_state)
        apply_state[(var_device, var_dtype)][&#34;weight_decay_rate&#34;] = tf.constant(
            self.weight_decay_rate, name=&#34;adam_weight_decay_rate&#34;
        )

    def _decay_weights_op(self, var, learning_rate, apply_state):
        do_decay = self._do_use_weight_decay(var.name)
        if do_decay:
            return var.assign_sub(
                learning_rate
                * var
                * apply_state[(var.device, var.dtype.base_dtype)][&#34;weight_decay_rate&#34;],
                use_locking=self._use_locking,
            )
        return tf.no_op()

    def apply_gradients(self, grads_and_vars, name=None, **kwargs):
        # 2020-07-05: added **kwargs based on https://github.com/tensorflow/addons/issues/1267
        grads, tvars = list(zip(*grads_and_vars))
        return super(AdamWeightDecay, self).apply_gradients(
            zip(grads, tvars),
            name=name,
        )

    def _get_lr(self, var_device, var_dtype, apply_state):
        &#34;&#34;&#34;Retrieves the learning rate with the given state.&#34;&#34;&#34;
        if apply_state is None:
            return self._decayed_lr_t[var_dtype], {}

        apply_state = apply_state or {}
        coefficients = apply_state.get((var_device, var_dtype))
        if coefficients is None:
            coefficients = self._fallback_apply_state(var_device, var_dtype)
            apply_state[(var_device, var_dtype)] = coefficients

        return coefficients[&#34;lr_t&#34;], dict(apply_state=apply_state)

    def _resource_apply_dense(self, grad, var, apply_state=None):
        lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
        decay = self._decay_weights_op(var, lr_t, apply_state)
        with tf.control_dependencies([decay]):
            return super(AdamWeightDecay, self)._resource_apply_dense(
                grad, var, **kwargs
            )

    def _resource_apply_sparse(self, grad, var, indices, apply_state=None):
        lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
        decay = self._decay_weights_op(var, lr_t, apply_state)
        with tf.control_dependencies([decay]):
            return super(AdamWeightDecay, self)._resource_apply_sparse(
                grad, var, indices, **kwargs
            )

    def get_config(self):
        config = super().get_config()
        config.update({&#34;weight_decay_rate&#34;: self.weight_decay_rate})
        return config

    def _do_use_weight_decay(self, param_name):
        &#34;&#34;&#34;Whether to use L2 weight decay for `param_name`.&#34;&#34;&#34;
        if self.weight_decay_rate == 0:
            return False

        if self._include_in_weight_decay:
            for r in self._include_in_weight_decay:
                if re.search(r, param_name) is not None:
                    return True

        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</code></pre>
</details>
<h3>Ancestors</h3>
<ul class="hlist">
<li>tensorflow.python.keras.optimizer_v2.adam.Adam</li>
<li>tensorflow.python.keras.optimizer_v2.optimizer_v2.OptimizerV2</li>
<li>tensorflow.python.training.tracking.base.Trackable</li>
</ul>
<h3>Static methods</h3>
<dl>
<dt id="ktrain.lroptimize.optimization.AdamWeightDecay.from_config"><code class="name flex">
<span>def <span class="ident">from_config</span></span>(<span>config)</span>
</code></dt>
<dd>
<div class="desc"><p>Creates an optimizer from its config with WarmUp custom object.</p></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">@classmethod
def from_config(cls, config):
    &#34;&#34;&#34;Creates an optimizer from its config with WarmUp custom object.&#34;&#34;&#34;
    custom_objects = {&#34;WarmUp&#34;: WarmUp}
    return super(AdamWeightDecay, cls).from_config(
        config, custom_objects=custom_objects
    )</code></pre>
</details>
</dd>
</dl>
<h3>Methods</h3>
<dl>
<dt id="ktrain.lroptimize.optimization.AdamWeightDecay.apply_gradients"><code class="name flex">
<span>def <span class="ident">apply_gradients</span></span>(<span>self, grads_and_vars, name=None, **kwargs)</span>
</code></dt>
<dd>
<div class="desc"><p>Apply gradients to variables.</p>
<p>This is the second part of <code>minimize()</code>. It returns an <code>Operation</code> that
applies gradients.</p>
<p>The method sums gradients from all replicas in the presence of
<code>tf.distribute.Strategy</code> by default. You can aggregate gradients yourself by
passing <code>experimental_aggregate_gradients=False</code>.</p>
<p>Example:</p>
<pre><code class="language-python">grads = tape.gradient(loss, vars)
grads = tf.distribute.get_replica_context().all_reduce('sum', grads)
# Processing aggregated gradients.
optimizer.apply_gradients(zip(grads, vars),
    experimental_aggregate_gradients=False)

</code></pre>
<h2 id="args">Args</h2>
<dl>
<dt><strong><code>grads_and_vars</code></strong></dt>
<dd>List of (gradient, variable) pairs.</dd>
<dt><strong><code>name</code></strong></dt>
<dd>Optional name for the returned operation. Default to the name passed
to the <code>Optimizer</code> constructor.</dd>
<dt><strong><code>experimental_aggregate_gradients</code></strong></dt>
<dd>Whether to sum gradients from different
replicas in the presense of <code>tf.distribute.Strategy</code>. If False, it's
user responsibility to aggregate the gradients. Default to True.</dd>
</dl>
<h2 id="returns">Returns</h2>
<p>An <code>Operation</code> that applies the specified gradients. The <code>iterations</code>
will be automatically increased by 1.</p>
<h2 id="raises">Raises</h2>
<dl>
<dt><code>TypeError</code></dt>
<dd>If <code>grads_and_vars</code> is malformed.</dd>
<dt><code>ValueError</code></dt>
<dd>If none of the variables have gradients.</dd>
</dl></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">def apply_gradients(self, grads_and_vars, name=None, **kwargs):
    # 2020-07-05: added **kwargs based on https://github.com/tensorflow/addons/issues/1267
    grads, tvars = list(zip(*grads_and_vars))
    return super(AdamWeightDecay, self).apply_gradients(
        zip(grads, tvars),
        name=name,
    )</code></pre>
</details>
</dd>
<dt id="ktrain.lroptimize.optimization.AdamWeightDecay.get_config"><code class="name flex">
<span>def <span class="ident">get_config</span></span>(<span>self)</span>
</code></dt>
<dd>
<div class="desc"><p>Returns the config of the optimizer.</p>
<p>An optimizer config is a Python dictionary (serializable)
containing the configuration of an optimizer.
The same optimizer can be reinstantiated later
(without any saved state) from this configuration.</p>
<h2 id="returns">Returns</h2>
<p>Python dictionary.</p></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">def get_config(self):
    config = super().get_config()
    config.update({&#34;weight_decay_rate&#34;: self.weight_decay_rate})
    return config</code></pre>
</details>
</dd>
</dl>
</dd>
<dt id="ktrain.lroptimize.optimization.GradientAccumulator"><code class="flex name class">
<span>class <span class="ident">GradientAccumulator</span></span>
</code></dt>
<dd>
<div class="desc"><p>Gradient accumulation utility.
When used with a distribution strategy, the accumulator should be called in a
replica context. Gradients will be accumulated locally on each replica and
without synchronization. Users should then call <code>.gradients</code>, scale the
gradients if required, and pass the result to <code>apply_gradients</code>.</p>
<p>Initializes the accumulator.</p></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">class GradientAccumulator(object):
    &#34;&#34;&#34;Gradient accumulation utility.
    When used with a distribution strategy, the accumulator should be called in a
    replica context. Gradients will be accumulated locally on each replica and
    without synchronization. Users should then call ``.gradients``, scale the
    gradients if required, and pass the result to ``apply_gradients``.
    &#34;&#34;&#34;

    # We use the ON_READ synchronization policy so that no synchronization is
    # performed on assignment. To get the value, we call .value() which returns the
    # value on the current replica without synchronization.

    def __init__(self):
        &#34;&#34;&#34;Initializes the accumulator.&#34;&#34;&#34;
        self._gradients = []
        self._accum_steps = None

    @property
    def step(self):
        &#34;&#34;&#34;Number of accumulated steps.&#34;&#34;&#34;
        if self._accum_steps is None:
            self._accum_steps = tf.Variable(
                tf.constant(0, dtype=tf.int64),
                trainable=False,
                synchronization=tf.VariableSynchronization.ON_READ,
                aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA,
            )

        return self._accum_steps.value()

    @property
    def gradients(self):
        &#34;&#34;&#34;The accumulated gradients on the current replica.&#34;&#34;&#34;
        if not self._gradients:
            raise ValueError(
                &#34;The accumulator should be called first to initialize the gradients&#34;
            )
        return list(
            gradient.value() if gradient is not None else gradient
            for gradient in self._gradients
        )

    def __call__(self, gradients):
        &#34;&#34;&#34;Accumulates :obj:`gradients` on the current replica.&#34;&#34;&#34;
        if not self._gradients:
            _ = self.step  # Create the step variable.
            self._gradients.extend(
                [
                    tf.Variable(
                        tf.zeros_like(gradient),
                        trainable=False,
                        synchronization=tf.VariableSynchronization.ON_READ,
                        aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA,
                    )
                    if gradient is not None
                    else gradient
                    for gradient in gradients
                ]
            )
        if len(gradients) != len(self._gradients):
            raise ValueError(
                &#34;Expected %s gradients, but got %d&#34;
                % (len(self._gradients), len(gradients))
            )

        for accum_gradient, gradient in zip(self._gradients, gradients):
            if accum_gradient is not None and gradient is not None:
                accum_gradient.assign_add(gradient)

        self._accum_steps.assign_add(1)

    def reset(self):
        &#34;&#34;&#34;Resets the accumulated gradients on the current replica.&#34;&#34;&#34;
        if not self._gradients:
            return
        self._accum_steps.assign(0)
        for gradient in self._gradients:
            if gradient is not None:
                gradient.assign(tf.zeros_like(gradient))</code></pre>
</details>
<h3>Instance variables</h3>
<dl>
<dt id="ktrain.lroptimize.optimization.GradientAccumulator.gradients"><code class="name">var <span class="ident">gradients</span></code></dt>
<dd>
<div class="desc"><p>The accumulated gradients on the current replica.</p></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">@property
def gradients(self):
    &#34;&#34;&#34;The accumulated gradients on the current replica.&#34;&#34;&#34;
    if not self._gradients:
        raise ValueError(
            &#34;The accumulator should be called first to initialize the gradients&#34;
        )
    return list(
        gradient.value() if gradient is not None else gradient
        for gradient in self._gradients
    )</code></pre>
</details>
</dd>
<dt id="ktrain.lroptimize.optimization.GradientAccumulator.step"><code class="name">var <span class="ident">step</span></code></dt>
<dd>
<div class="desc"><p>Number of accumulated steps.</p></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">@property
def step(self):
    &#34;&#34;&#34;Number of accumulated steps.&#34;&#34;&#34;
    if self._accum_steps is None:
        self._accum_steps = tf.Variable(
            tf.constant(0, dtype=tf.int64),
            trainable=False,
            synchronization=tf.VariableSynchronization.ON_READ,
            aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA,
        )

    return self._accum_steps.value()</code></pre>
</details>
</dd>
</dl>
<h3>Methods</h3>
<dl>
<dt id="ktrain.lroptimize.optimization.GradientAccumulator.reset"><code class="name flex">
<span>def <span class="ident">reset</span></span>(<span>self)</span>
</code></dt>
<dd>
<div class="desc"><p>Resets the accumulated gradients on the current replica.</p></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">def reset(self):
    &#34;&#34;&#34;Resets the accumulated gradients on the current replica.&#34;&#34;&#34;
    if not self._gradients:
        return
    self._accum_steps.assign(0)
    for gradient in self._gradients:
        if gradient is not None:
            gradient.assign(tf.zeros_like(gradient))</code></pre>
</details>
</dd>
</dl>
</dd>
<dt id="ktrain.lroptimize.optimization.WarmUp"><code class="flex name class">
<span>class <span class="ident">WarmUp</span></span>
<span>(</span><span>initial_learning_rate: float, decay_schedule_fn: Callable, warmup_steps: int, power: float = 1.0, name: str = None)</span>
</code></dt>
<dd>
<div class="desc"><pre><code>Applies a warmup schedule on a given learning rate decay schedule.
Args:
    initial_learning_rate (:obj:`float`):
        The initial learning rate for the schedule after the warmup (so this will be the learning rate at the end
        of the warmup).
    decay_schedule_fn (:obj:`Callable`):
        The schedule function to apply after the warmup for the rest of training.
    warmup_steps (:obj:`int`):
        The number of steps for the warmup part of training.
    power (:obj:`float`, `optional`, defaults to 1):
        The power to use for the polynomial warmup (defaults is a linear warmup).
    name (:obj:`str`, `optional`):
        Optional name prefix for the returned tensors during the schedule.
</code></pre></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">class WarmUp(tf.keras.optimizers.schedules.LearningRateSchedule):
    &#34;&#34;&#34;
    ```
    Applies a warmup schedule on a given learning rate decay schedule.
    Args:
        initial_learning_rate (:obj:`float`):
            The initial learning rate for the schedule after the warmup (so this will be the learning rate at the end
            of the warmup).
        decay_schedule_fn (:obj:`Callable`):
            The schedule function to apply after the warmup for the rest of training.
        warmup_steps (:obj:`int`):
            The number of steps for the warmup part of training.
        power (:obj:`float`, `optional`, defaults to 1):
            The power to use for the polynomial warmup (defaults is a linear warmup).
        name (:obj:`str`, `optional`):
            Optional name prefix for the returned tensors during the schedule.
    ```
    &#34;&#34;&#34;

    def __init__(
        self,
        initial_learning_rate: float,
        decay_schedule_fn: Callable,
        warmup_steps: int,
        power: float = 1.0,
        name: str = None,
    ):
        super().__init__()
        self.initial_learning_rate = initial_learning_rate
        self.warmup_steps = warmup_steps
        self.power = power
        self.decay_schedule_fn = decay_schedule_fn
        self.name = name

    def __call__(self, step):
        with tf.name_scope(self.name or &#34;WarmUp&#34;) as name:
            # Implements polynomial warmup. i.e., if global_step &lt; warmup_steps, the
            # learning rate will be `global_step/num_warmup_steps * init_lr`.
            global_step_float = tf.cast(step, tf.float32)
            warmup_steps_float = tf.cast(self.warmup_steps, tf.float32)
            warmup_percent_done = global_step_float / warmup_steps_float
            warmup_learning_rate = self.initial_learning_rate * tf.math.pow(
                warmup_percent_done, self.power
            )
            return tf.cond(
                global_step_float &lt; warmup_steps_float,
                lambda: warmup_learning_rate,
                lambda: self.decay_schedule_fn(step - self.warmup_steps),
                name=name,
            )

    def get_config(self):
        return {
            &#34;initial_learning_rate&#34;: self.initial_learning_rate,
            &#34;decay_schedule_fn&#34;: self.decay_schedule_fn,
            &#34;warmup_steps&#34;: self.warmup_steps,
            &#34;power&#34;: self.power,
            &#34;name&#34;: self.name,
        }</code></pre>
</details>
<h3>Ancestors</h3>
<ul class="hlist">
<li>tensorflow.python.keras.optimizer_v2.learning_rate_schedule.LearningRateSchedule</li>
</ul>
<h3>Methods</h3>
<dl>
<dt id="ktrain.lroptimize.optimization.WarmUp.get_config"><code class="name flex">
<span>def <span class="ident">get_config</span></span>(<span>self)</span>
</code></dt>
<dd>
<div class="desc"></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">def get_config(self):
    return {
        &#34;initial_learning_rate&#34;: self.initial_learning_rate,
        &#34;decay_schedule_fn&#34;: self.decay_schedule_fn,
        &#34;warmup_steps&#34;: self.warmup_steps,
        &#34;power&#34;: self.power,
        &#34;name&#34;: self.name,
    }</code></pre>
</details>
</dd>
</dl>
</dd>
</dl>
</section>
</article>
<nav id="sidebar">
<h1>Index</h1>
<div class="toc">
<ul></ul>
</div>
<ul id="index">
<li><h3>Super-module</h3>
<ul>
<li><code><a title="ktrain.lroptimize" href="index.html">ktrain.lroptimize</a></code></li>
</ul>
</li>
<li><h3><a href="#header-functions">Functions</a></h3>
<ul class="">
<li><code><a title="ktrain.lroptimize.optimization.create_optimizer" href="#ktrain.lroptimize.optimization.create_optimizer">create_optimizer</a></code></li>
</ul>
</li>
<li><h3><a href="#header-classes">Classes</a></h3>
<ul>
<li>
<h4><code><a title="ktrain.lroptimize.optimization.AdamWeightDecay" href="#ktrain.lroptimize.optimization.AdamWeightDecay">AdamWeightDecay</a></code></h4>
<ul class="">
<li><code><a title="ktrain.lroptimize.optimization.AdamWeightDecay.apply_gradients" href="#ktrain.lroptimize.optimization.AdamWeightDecay.apply_gradients">apply_gradients</a></code></li>
<li><code><a title="ktrain.lroptimize.optimization.AdamWeightDecay.from_config" href="#ktrain.lroptimize.optimization.AdamWeightDecay.from_config">from_config</a></code></li>
<li><code><a title="ktrain.lroptimize.optimization.AdamWeightDecay.get_config" href="#ktrain.lroptimize.optimization.AdamWeightDecay.get_config">get_config</a></code></li>
</ul>
</li>
<li>
<h4><code><a title="ktrain.lroptimize.optimization.GradientAccumulator" href="#ktrain.lroptimize.optimization.GradientAccumulator">GradientAccumulator</a></code></h4>
<ul class="">
<li><code><a title="ktrain.lroptimize.optimization.GradientAccumulator.gradients" href="#ktrain.lroptimize.optimization.GradientAccumulator.gradients">gradients</a></code></li>
<li><code><a title="ktrain.lroptimize.optimization.GradientAccumulator.reset" href="#ktrain.lroptimize.optimization.GradientAccumulator.reset">reset</a></code></li>
<li><code><a title="ktrain.lroptimize.optimization.GradientAccumulator.step" href="#ktrain.lroptimize.optimization.GradientAccumulator.step">step</a></code></li>
</ul>
</li>
<li>
<h4><code><a title="ktrain.lroptimize.optimization.WarmUp" href="#ktrain.lroptimize.optimization.WarmUp">WarmUp</a></code></h4>
<ul class="">
<li><code><a title="ktrain.lroptimize.optimization.WarmUp.get_config" href="#ktrain.lroptimize.optimization.WarmUp.get_config">get_config</a></code></li>
</ul>
</li>
</ul>
</li>
</ul>
</nav>
</main>
<footer id="footer">
<p>Generated by <a href="https://pdoc3.github.io/pdoc"><cite>pdoc</cite> 0.9.2</a>.</p>
</footer>
</body>
</html>