{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 06. 结合多保真优化"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### 多保真优化简介"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"在上一个教程中，我们学习了如何实现一个简单的AutoML系统。但这个AutoML系统的核心：贝叶斯优化算法往往需要大量对采样的评价才能获得比较好的结果。\n",
"\n",
"然而，在自动机器学习(Automatic Machine Learning, AutoML)任务中评价往往通过 k 折交叉验证获得，在大数据集的机器学习任务上，`获得一个评价的时间代价巨大`。这也影响了优化算法在自动机器学习问题上的效果。所以一些`减少评价代价`的方法被提出来，其中**多保真度优化**(Multi-Fidelity Optimization)[[1]](#refer-anchor-1)就是其中的一种。而**多臂老虎机算法**(Multi-armed Bandit Algorithm, MBA)[[2]](#refer-anchor-2)是多保真度算法的一种。在此基础上，有两种主流的`bandit-based`优化策略：\n",
"\n",
"- Successive Halving (SH) [[3]](#refer-anchor-3)\n",
"- Hyperband (HB) [[4]](#refer-anchor-4)\n",
"\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"首先我们介绍连续减半(Successive Halving ,SH)。在连续减半策略中， 我们将`评价代价`参数化为一个变量`budget`，即预算。根据BOHB论文[[5]](#refer-anchor-5)的阐述，我们可以根据不同的场景定义不同的budget，举例如下：\n",
"\n",
"1. 迭代算法的迭代数(如：神经网络的epoch、随机森林，GBDT的树的个数)\n",
"2. 机器学习算法所使用的样本数\n",
"3. 贝叶斯神经网络[[6]](#refer-anchor-6)中MCMC链的长度\n",
"4. 深度强化学习中的尝试数\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"举例说明，我们定义$budget_{max}=1$, $budget_{min}=\\frac{1}{8}$, $\\eta=2$ (`eta` = 2) 。在这里`budget`的语义表示使用$100\\times budget$%的样本。\n",
"\n",
"1. 首先我们从配置空间(或称为超参空间)**随机采样**8个配置，实例化为8个机器学习模型。\n",
"2. 然后用$\\frac{1}{8}$的训练样本训练这8个模型并在验证集得到相应的损失值。\n",
"3. 保留这8个模型中loss最低的前4个模型，其余的舍弃。\n",
"4. 依次类推，最后仅保留一个模型，并且其`budget=1`(可以用全部的样本进行训练)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
""
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"上图描述了例子中的迭代过程(图片来自[[1]](#refer-anchor-1)) 。我们可以用`ultraopt.multi_fidelity`中的`SuccessiveHalvingIterGenerator`来实例化这一过程："
]
},
{
"cell_type": "code",
"execution_count": 1,
"metadata": {},
"outputs": [],
"source": [
"from ultraopt.multi_fidelity import SuccessiveHalvingIterGenerator, HyperBandIterGenerator"
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {},
"outputs": [
{
"data": {
"text/html": [
"\n",
"\n",
"
\n",
" \n",
" \n",
" | \n",
" iter 0 | \n",
"
\n",
" \n",
" | \n",
" stage 0 | \n",
" stage 1 | \n",
" stage 2 | \n",
" stage 3 | \n",
"
\n",
" \n",
" \n",
" \n",
" | num_config | \n",
" 8 | \n",
" 4 | \n",
" 2 | \n",
" 1 | \n",
"
\n",
" \n",
" | budget | \n",
" 1/8 | \n",
" 1/4 | \n",
" 1/2 | \n",
" 1 | \n",
"
\n",
" \n",
"
\n",
"
\n",
"\n",
"
\n",
" \n",
" \n",
" | \n",
" iter 0 | \n",
" iter 1 | \n",
" iter 2 | \n",
" iter 3 | \n",
"
\n",
" \n",
" | \n",
" stage 0 | \n",
" stage 1 | \n",
" stage 2 | \n",
" stage 3 | \n",
" stage 0 | \n",
" stage 1 | \n",
" stage 2 | \n",
" stage 0 | \n",
" stage 1 | \n",
" stage 0 | \n",
"
\n",
" \n",
" \n",
" \n",
" | num_config | \n",
" 8 | \n",
" 4 | \n",
" 2 | \n",
" 1 | \n",
" 4 | \n",
" 2 | \n",
" 1 | \n",
" 4 | \n",
" 2 | \n",
" 4 | \n",
"
\n",
" \n",
" | budget | \n",
" 1/8 | \n",
" 1/4 | \n",
" 1/2 | \n",
" 1 | \n",
" 1/4 | \n",
" 1/2 | \n",
" 1 | \n",
" 1/2 | \n",
" 1 | \n",
" 1 | \n",
"
\n",
" \n",
"
\n",
"
\n",
"\n",
"
\n",
" \n",
" \n",
" | \n",
" iter 0 | \n",
" iter 1 | \n",
" iter 2 | \n",
"
\n",
" \n",
" | \n",
" stage 0 | \n",
" stage 1 | \n",
" stage 2 | \n",
" stage 0 | \n",
" stage 1 | \n",
" stage 0 | \n",
"
\n",
" \n",
" \n",
" \n",
" | num_config | \n",
" 4 | \n",
" 2 | \n",
" 1 | \n",
" 2 | \n",
" 1 | \n",
" 3 | \n",
"
\n",
" \n",
" | budget | \n",
" 25 | \n",
" 50 | \n",
" 100 | \n",
" 50 | \n",
" 100 | \n",
" 100 | \n",
"
\n",
" \n",
"
\n",
"