Gaussian_Processes.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "dense-anthony",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "    <h1> Tutorial 4 </h1> \n",
    "    <h2> Gaussian Processes </h2>\n",
    "</div>    "
   ]
  },
  {
   "cell_type": "markdown",
   "id": "careful-fashion",
   "metadata": {},
   "source": [
    "## Overview\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "liked-canyon",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "\n",
    "<h1> Gaussian Processes </h1>\n",
    "\n",
    "This tutorial is based on work done by Ollie Pollard, using Gaussian Processes to predict sea-level rise. We follow the steps outlined in this [visual article](https://distill.pub/2019/visual-exploration-gaussian-processes/). Gaussian processes are often used to make predictions on new data by incorporating prior knowledge to fit a function to a data set. For a given set of training points, there are potentially infinitely many functions that fit the data. Gaussian processes offer an elegant solution to this problem by assigning a probability to each of these functions. The mean of this probability distribution then represents the most probable characterization of the data.\n",
    "\n",
    "\n",
    "    \n",
    "## Recommended reading\n",
    "\n",
    "* [Overview of Linear Regression](https://towardsdatascience.com/linear-regression-detailed-view-ea73175f6e86)\n",
    "* [An Intuitive Guide to Gaussian Processes](https://towardsdatascience.com/an-intuitive-guide-to-gaussian-processes-ec2f0b45c71d)\n",
    "    \n",
    "</div>\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "hispanic-recall",
   "metadata": {},
   "source": [
    "<hr>\n",
    "\n",
    "\n",
    "<div style=\"background-color: #e6ccff; padding: 10px;\">\n",
    "    \n",
    "\n",
    "<h1> Machine Learning Theory </h1>\n",
    "    \n",
    "<a href=\"https://distill.pub/2019/visual-exploration-gaussian-processes/\">\n",
    "                <img src=\"animations/post_sample.gif\">\n",
    "</a>\n",
    "    \n",
    "This tutorial is mainly focusing on using Python to explore Gaussian Processes. Please read the full [visual article](https://distill.pub/2019/visual-exploration-gaussian-processes/) for a more in-depth explanation.\n",
    "    \n",
    "The [gaussian distribution](https://en.wikipedia.org/wiki/Normal_distribution) forms the main building block of Gaussian Processes:\n",
    "   $\\displaystyle f(x) ={\\frac {1}{\\sigma {\\sqrt {2\\pi }}}}e^{-{\\frac {1}{2}}\\left({\\frac {x-\\mu }{\\sigma }}\\right)^{2}}$\n",
    " \n",
    "<a href=\"https://distill.pub/2019/visual-exploration-gaussian-processes/\">\n",
    "    <img src=\"animations/gaussian_dist.gif\">\n",
    "</a>\n",
    "\n",
    "For Machine Learning Gaussian Processes we are interested in the [multivariate case](https://en.wikipedia.org/wiki/Multivariate_normal_distribution). The multivariate Gaussian distribution is defined by a mean vector $\\mu$ (the expected value of the distribution) and a covariance matrix $\\Sigma$. $\\Sigma$ models the variance along each dimension and determines how the different random variables are correlated. If $X$ follows a normal distribution:\n",
    "    \n",
    "\\begin{equation} X = \\begin{bmatrix} X_1 \\\\ X_2 \\\\ \\vdots \\\\ X_n \\end{bmatrix} \\sim \\mathcal{N}(\\mu, \\Sigma) \\end{equation}\n",
    " \n",
    "The covariance matrix $\\Sigma$ describes the shape of the distribution. It can be defined using the expectation operator $E$\n",
    "    \n",
    "\\begin{equation}    \n",
    "\\Sigma = \\text{Cov}(X_i, X_j) = E \\left[ (X_i - \\mu_i)(X_j - \\mu_j)^T \\right]\n",
    "\\end{equation}\n",
    " \n",
    "Gaussian distributions have the nice algebraic property of being closed under conditioning and marginalization. Being closed under conditioning and marginalization means that the resulting distributions from these operations are also Gaussian, which makes many problems in statistics and machine learning tractable.\n",
    "Marginalization and conditioning both work on subsets of the original distribution and we will use the following notation:\n",
    "\n",
    "\\begin{equation}\n",
    "\\begin{bmatrix} X \\\\ Y \\end{bmatrix} \\sim \\mathcal{N}(\\mu, \\Sigma) = \\mathcal{N} \\left( \\begin{bmatrix} \\mu_X \\\\ \\mu_Y \\end{bmatrix}, \\begin{bmatrix} \\Sigma_{XX} \\, \\Sigma_{XY} \\\\ \\Sigma_{YX} \\, \\Sigma_{YY} \\end{bmatrix} \\right)   \n",
    "\\end{equation}\n",
    "\n",
    "With $X$ and $Y$ representing subsets of original random variables.   \n",
    "Through marginalization, we can extract partial information from multivariate probability distributions. In particular, given a normal probability distribution $P(X, Y)$ over vectors of random variables $X$ and $Y$, we can determine their marginalized probability distributions in the following way:\n",
    "    \n",
    "\\begin{equation}\n",
    "\\begin{aligned}\n",
    "X \\sim \\mathcal{N}(\\mu_X, \\Sigma_{XX}) \\\\\n",
    "Y \\sim \\mathcal{N}(\\mu_Y, \\Sigma_{YY})\n",
    "\\end{aligned}\n",
    "\\end{equation}    \n",
    "    \n",
    "The interpretation of this equation is that each partition $X$ and $Y$ only depends on its corresponding entries in $\\mu$ and $\\Sigma$. To marginalize a random variable from a Gaussian distribution we can simply drop the variables from $\\mu$ and $\\Sigma$.\n",
    "    \n",
    "\\begin{equation}  \n",
    "p_X(x) = \\int_y p_{X,Y}(x,y)dy = \\int_y p_{X|Y}(x|y) p_Y(y) dy\n",
    "\\end{equation}\n",
    "    \n",
    "The way to interpret this equation is that if we are interested in the probability density of $X=x$, we need to consider all possible outcomes of $Y$ that can jointly lead to the result.\n",
    "\n",
    "Another important operation for Gaussian processes is conditioning. It is used to determine the probability of one variable depending on another variable. Similar to marginalization, this operation is also closed and yields a modified Gaussian distribution. This operation is the cornerstone of Gaussian processes since it allows Bayesian inference. Conditioning is defined by:\n",
    "    \n",
    "\\begin{equation}     \n",
    "\\begin{aligned}\n",
    "X|Y \\sim \\mathcal{N}(\\:\\mu_X + \\Sigma_{XY}\\Sigma_{YY}^{-1}(Y - \\mu_Y),\\: \\Sigma_{XX}-\\Sigma_{XY}\\Sigma_{YY}^{-1}\\Sigma_{YX}\\:) \\\\\n",
    "Y|X \\sim \\mathcal{N}(\\:\\mu_Y + \\Sigma_{YX}\\Sigma_{XX}^{-1}(X - \\mu_X),\\: \\Sigma_{YY}-\\Sigma_{YX}\\Sigma_{XX}^{-1}\\Sigma_{XY}\\:) \\\\\n",
    "\\end{aligned}\n",
    "\\end{equation}\n",
    "\n",
    "The new mean only depends on the conditioned variable, while the covariance matrix is independent of this variable.\n",
    "    \n",
    "</div>    \n",
    " \n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "expired-attribute",
   "metadata": {},
   "source": [
    "  \n",
    "<div style=\"background-color: #cce5ff; padding: 10px;\">\n",
    "\n",
    "<h1> Python </h1>\n",
    "\n",
    "## Tensorflow and GPflow\n",
    "    \n",
    "There are many machine learning python libraries available, [TensorFlow](https://www.tensorflow.org/) a is one such library. Throughout this tutorial, you will see some complex machine learning tasks executed in just a few lines of code by calling [GPflow](https://gpflow.readthedocs.io/en/master/) functions which use TensorFlow. If you have GPUs on the machine you are using, these python libraries will automatically use them and run the code even faster!\n",
    "\n",
    "## Further Reading\n",
    "    \n",
    "* [GPflow example notebooks](https://gpflow.readthedocs.io/en/develop/notebooks_file.html)\n",
    "\n",
    "</div>\n",
    "    \n",
    "<hr>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "seasonal-concept",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ffffcc; padding: 10px;\">\n",
    "    \n",
    "<h1> Requirements </h1>\n",
    "\n",
    "These notebooks should run on a standard laptop with the correct python environment.\n",
    "\n",
    "<h2> Python Packages: </h2>\n",
    "\n",
    "* Python 3.8\n",
    "* tensorflow > 2.1\n",
    "* gpflow 2.1 *(must be installed via pip to get latest version)*\n",
    "* numpy \n",
    "* matplotlib\n",
    "* plotly\n",
    "* scipy\n",
    "* pandas\n",
    "\n",
    "\n",
    "<h2> Data Requirements</h2>\n",
    "    \n",
    "This notebook refers to some data included in the git hub repository in the [data](data) folder.\n",
    "    \n",
    "</div>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "coral-attempt",
   "metadata": {},
   "source": [
    "\n",
    "**Contents:**\n",
    "\n",
    "1. [Overview of Gaussian Processes](#Overview-of-Gausian-Processes)\n",
    "2. [Sea Level Example](#Sea-Level-Example)\n",
    "1. [Load Data](#Load-Data)\n",
    "2. [Normalise Data](#Normalise-Data)\n",
    "3. [Plot Data](#Plot-Data)\n",
    "4. [Define GP flow model](#Define-GP-flow-model)\n",
    "5. [Optimization](#Optimization)\n",
    "6. [Prediction](#Prediction)\n",
    "7. [Cross Validation](#Cross-Validation)\n",
    "8. [Plot High Stands](#Plot-High-Stands)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "enabling-steam",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #cce5ff; padding: 10px;\">\n",
    "    \n",
    "Load in all required modules and turn off warnings. If you have no [GPU](https://www.analyticsvidhya.com/blog/2020/09/why-gpus-are-more-suited-for-deep-learning/) available, you may see some TensorFlow warnings.\n",
    "\n",
    "</div>\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "inappropriate-magic",
   "metadata": {},
   "outputs": [],
   "source": [
    "# For readability: disable warnings\n",
    "import warnings\n",
    "warnings.filterwarnings('ignore')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "affecting-scanner",
   "metadata": {},
   "outputs": [],
   "source": [
    "# import modules\n",
    "import gpflow\n",
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "import plotly.express as px\n",
    "from gpflow.utilities import print_summary\n",
    "from scipy.spatial.distance import cdist\n",
    "from scipy.stats import multivariate_normal\n",
    "\n",
    "gpflow.config.set_default_summary_fmt(\"notebook\")\n",
    "plt.style.use(\"seaborn-v0_8-whitegrid\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "922ed27c",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #cce5ff; padding: 10px;\">\n",
    "\n",
    "The next cell checks for GPUs. If the result is not what you expect, check your python installation.\n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "affccf11",
   "metadata": {},
   "outputs": [],
   "source": [
    "if tf.config.list_physical_devices('GPU'):\n",
    "    print('GPU found')\n",
    "    tf.config.list_physical_devices('GPU')\n",
    "else:\n",
    "    print('No GPU found')"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "07ce84e8",
   "metadata": {},
   "source": [
    "# Overview of Gaussian Processes\n",
    "\n",
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "\n",
    "Say we want to find an unknown function `f` where $f(x) = x^3 - 9x + e^{(x^3/30)}$ from a random sample of points [`x_samp`, `y_samp`]\n",
    "    \n",
    "**NB function `f(x)` is chosen arbitrarily for this demonstration. You can change `f` to whatever you like and see similar results!**\n",
    "    \n",
    "\n",
    "</div>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "27c14374",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    " \n",
    "After you've run through the code you might want to re-run with a different choice of `f`.\n",
    "    \n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d8440359",
   "metadata": {},
   "outputs": [],
   "source": [
    "# simulated data from some function that we're going to try and find\n",
    "# set x to between -6 and 6\n",
    "x = np.linspace(-6,6,100)\n",
    "# y = f(x) chosen to demonstrate Gaussian Processes\n",
    "y = x**3 - 9*x + np.exp(x**3/30)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "05515d9e",
   "metadata": {},
   "source": [
    "# Generate sample"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "86dadabc",
   "metadata": {},
   "outputs": [],
   "source": [
    "# a random sample of points\n",
    "# whole x range is between -6 and 6 and we'll sample x between -3 and 3 \n",
    "sample_size = 5\n",
    "\n",
    "samp_index = np.random.randint((len(x)-1)/4, high=(len(x)-1)*3/4, size=sample_size)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "75c75cb9",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "\n",
    "Later, if you wish, you can uncomment the below cell to see the effects of changing the sample size or sample area (below is set to sample the whole range with a sample size of 8).\n",
    "    \n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "bf732d59",
   "metadata": {},
   "outputs": [],
   "source": [
    "#sample_size = 8\n",
    "#samp_index = np.random.randint(0, high=len(x), size=sample_size)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5f2f9385",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #cce5ff; padding: 10px;\">\n",
    " \n",
    "As we've taken a random sample there's a small chance we might have sampled the same point twice so the below code is going to check we have 5 unique sample points.\n",
    "\n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "3526ef0a",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Check no duplicates in sample index\n",
    "\n",
    "if len(np.unique(samp_index)) != sample_size :\n",
    "    print(\"duplicate sample index found please rerun above cell\")\n",
    "    print(\"the sample will automatically be reset to 5 over the whole range\")\n",
    "    sample_size = 5\n",
    "    samp_index =  np.random.randint(0, high=len(x), size=sample_size)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "016c0340",
   "metadata": {},
   "outputs": [],
   "source": [
    "samp_index"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "b48cff78",
   "metadata": {},
   "outputs": [],
   "source": [
    "x_samp = x[samp_index]\n",
    "y_samp = y[samp_index]\n",
    "# Plot Sample points\n",
    "fig, ax = plt.subplots(figsize=(10,5))\n",
    "point_c = \"#f77f00\"\n",
    "ax.scatter(x_samp, y_samp, s=50, c=point_c, zorder=10)\n",
    "ax.set_xlim([-6,6])\n",
    "ax.set_xlabel(\"x\")\n",
    "ax.set_ylabel(\"f(x)\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "68be05c9",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "\n",
    "From these few points, it's not obvious what function `f` could possibly be. So we can use Gaussian Processes to help work out the unknown function from these few points.\n",
    "\n",
    "The covariance matrix $\\Sigma$ is determined by its covariance function $K$, which is often also called the kernel of the Gaussian process. Here we will use the [Radial Basis Function Kernel](https://towardsdatascience.com/radial-basis-function-rbf-kernel-the-go-to-kernel-acf0d22c798a):\n",
    "    \n",
    "\\begin{equation}\n",
    " K(X_1,X_2) = \\sigma^2 \\exp \\left( -\\dfrac{||X_1 - X_2||^2}{2 l^2} \\right)\n",
    "\\end{equation}\n",
    "\n",
    "where $\\sigma^2$ and $l$ are the variance and lengthscale hyperparameters respectively and $||X_1 - X_2|| ^2$ is the squared Euclidean distance between two points.\n",
    "    \n",
    "This covariance function is defined in the function `rbf_kernel`.\n",
    "    \n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9c77c430",
   "metadata": {},
   "outputs": [],
   "source": [
    "# radial basis function kernel\n",
    "def rbf_kernel(x1, x2, var, lscale):\n",
    "    \"\"\"\n",
    "        Compute the Euclidean distance between each row of X1 and X2, or between\n",
    "        each pair of rows of X1 if X2 is None and feed it to the kernel.\n",
    "    \"\"\"\n",
    "    if x2 is None:\n",
    "        d = cdist(x1, x1)\n",
    "    else:\n",
    "        d = cdist(x1, x2)\n",
    "    K = var*np.exp(-np.power(d,2)/(2*lscale**2))\n",
    "    return K"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "82ca3e8d",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "\n",
    "So `K` can be obtained via `rbf_kernel` and $\\mu$ `mu` is often assumed to be zero as a starting point.\n",
    "</div>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e90b5f6d",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #cce5ff; padding: 10px;\">\n",
    "\n",
    "Try adjusting `lscale` to see what happens with the results. E.g., after running through the first time, set `lscale=5` and run through all cells again.\n",
    "\n",
    "Increasing the lengthscale parameter increases the banding, as points further away from each other become more correlated.\n",
    "\n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9e2bb5b3",
   "metadata": {},
   "outputs": [],
   "source": [
    "lscale = 1.0\n",
    "K_rbf = rbf_kernel(x.reshape(-1,1), None, 1.0, lscale)\n",
    "mu = np.zeros(x.shape[0])\n",
    "f_rbf = np.random.multivariate_normal(mu, K_rbf, 100)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ff646d12",
   "metadata": {},
   "outputs": [],
   "source": [
    "rv = multivariate_normal(np.zeros(2), rbf_kernel(np.arange(0,2).reshape(-1,1), None, 1.0, 2.0))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "64f1ad96",
   "metadata": {},
   "outputs": [],
   "source": [
    "#Create grid and multivariate normal\n",
    "a = np.linspace(-3,3,500)\n",
    "b = np.linspace(-3,3,500)\n",
    "X, Y = np.meshgrid(a,b)\n",
    "pos = np.empty(X.shape + (2,))\n",
    "pos[:, :, 0] = X\n",
    "pos[:, :, 1] = Y"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "98be54da",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "\n",
    "Stochastic processes, such as Gaussian processes, are essentially a set of random variables. In addition, each of these random variables has a corresponding index $i$. We will use this index to refer to the $i$-th dimension of our nnn-dimensional multivariate distributions.\n",
    "    \n",
    "Below, we have a two-dimensional normal distribution. Each dimension $y_i$ is assigned an index $\\displaystyle{i \\in {1,2}}$.  This representation allows us to understand the connection between the covariance and the resulting values: the underlying Gaussian distribution has a positive covariance between $y_1$ and $y_2$ -  this means that $y_2$ will increase as $y_1$ gets larger and vice versa. \n",
    "    \n",
    "</div>\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "23b0eb7b",
   "metadata": {},
   "outputs": [],
   "source": [
    "grid = plt.GridSpec(1, 2, wspace=0.2, hspace=0.3)\n",
    "fig = plt.figure(figsize=(10, 5))\n",
    "ax1 = fig.add_subplot(grid[0, 0])\n",
    "ax2 = fig.add_subplot(grid[0, 1])\n",
    "ax1.contourf(X, Y, rv.pdf(pos), cmap=\"Reds\", levels=300)\n",
    "for index in range(10):\n",
    "    rand_y1, rand_y2 = np.random.multivariate_normal(np.zeros(2), rbf_kernel(np.arange(0,2).reshape(-1,1), None, 1.0, 2.0), 1)[0]\n",
    "    ax1.scatter(rand_y1, rand_y2)\n",
    "    ax1.set_xlabel(\"$y_1$\")\n",
    "    ax1.set_ylabel(\"$y_2$\")\n",
    "    ax2.plot(np.arange(1,3), [rand_y1, rand_y2], '-o')\n",
    "    ax2.set_xlim([0,3])\n",
    "    ax2.set_ylim([-3,3])\n",
    "    ax2.set_xlabel(\"index\")\n",
    "    ax2.set_ylabel(\"$y$\")\n",
    "    ax2.set_xticks([1,2])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "bf7147ba",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "    \n",
    "The stronger the correlation, the more horizontal the lines will be.\n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "75487067",
   "metadata": {},
   "outputs": [],
   "source": [
    "# rbf_kernel(X1,X2,var,lscale)\n",
    "rbf_kernel(np.arange(0,2).reshape(-1,1), None, 1.0, lscale)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "af39fff9",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "\n",
    "## Prior distribution\n",
    "\n",
    "\n",
    "The following figure shows samples of potential functions from our prior distribution (the case where we have not yet observed any training data) that were created using the RBF kernel.\n",
    "    \n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ff4bf6c5",
   "metadata": {},
   "outputs": [],
   "source": [
    "var_index = np.linspace(0, 20,100)\n",
    "fig, ax = plt.subplots(figsize=(10,5))\n",
    "for index in range(10):\n",
    "    ax.plot(var_index, np.random.multivariate_normal(np.zeros(var_index.shape[0]), rbf_kernel(var_index.reshape(-1,1), None, 1.0, 5.0), 1)[0])\n",
    "    ax.set_xlim(0,20)\n",
    "    ax.set_ylim(-3,3)\n",
    "    ax.set_xlabel(\"$x$\")\n",
    "    ax.set_ylabel(\"$y$\")\n",
    "    ax.set_xticks([])\n",
    "ax.set_title('GP samples from RBF where lscale ='+str(lscale))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "198512d1",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Covariance matrix from the RBF kernel\n",
    "# rbf_kernel(X1,X2,var,lscale)\n",
    "K_rbf = rbf_kernel(x.reshape(-1,1), None, 1.0, lscale)\n",
    "# mu = 0\n",
    "mu = np.zeros(x.shape[0])\n",
    "# functions from rbf kernel\n",
    "f_rbf = np.random.multivariate_normal(mu, K_rbf, 100)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "becd6ca0",
   "metadata": {},
   "outputs": [],
   "source": [
    "# plot 100 GP samples\n",
    "index=100\n",
    "fig, ax = plt.subplots(figsize=(10,5))\n",
    "ax.plot(x, f_rbf.T[:,0:index], color='C0', alpha=0.2)\n",
    "ax.set_xlim([-6,6])\n",
    "# depending you you function choice you might need to comment out the ylim line below \n",
    "ax.set_ylim([-10,15])\n",
    "ax.set_xlabel(\"x\")\n",
    "ax.set_ylabel(\"f(x)\")\n",
    "ax.set_title('100 GP samples from  RBF kernel where lscale ='+str(lscale))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1fcbf842",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "    \n",
    "as the mean $\\mu$ is set to zero, all functions are distributed normally around zero.\n",
    "</div>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "38c26c7d",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "\n",
    "## Posterior distribution\n",
    "\n",
    "Now we're going to activate training data which we can add back into our distribution to give the posterior distribution (where we have incorporated the training data into our model).\n",
    "\n",
    "Let $\\mathbf{K}$ denote the prior covariance matrix of the training points derived from our RBF kernel. We define $\\mathbf{K}_*$ as the prior cross-covariance matrix of our training points and the points at which we wish to make predictions. Finally, let $\\mathbf{K}_{*,*}$ be the prior covariance matrix of the points at which to make predictions.\n",
    "\n",
    "1. First, we form the joint distribution of all the x points `x` and the training points `x_samp` which gives cross-covariance matrix `k_starX` .\n",
    "2. We can the use `k_xx` (covariance matrix of training points) `k_starstar` covariance matrix of all x to calculate $\\boldsymbol \\mu_{\\text{pred}}$ and $\\mathbf{K}_{\\text{pred}}$\n",
    "    \n",
    "\\begin{equation}\n",
    "\\boldsymbol \\mu_{\\text{pred}} = \\mathbf{K}_*^\\top \\left[\\mathbf{K} + \\sigma^2 \\mathbf{I}\\right]^{-1} \\mathbf{y}\n",
    "\\end{equation}\n",
    "    \n",
    "\\begin{equation}\n",
    "\\mathbf{K}_{\\text{pred}} = \\mathbf{K}_{*,*} - \\mathbf{K}_*^\\top \\left[\\mathbf{K} + \\sigma^2 \\mathbf{I}\\right]^{-1} \\mathbf{K}_*.\n",
    "\\end{equation}\n",
    " \n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "a11ec1cb",
   "metadata": {},
   "outputs": [],
   "source": [
    "# cross-covariance matrix\n",
    "k_starX = rbf_kernel(x.reshape(-1,1), x_samp.reshape(-1,1), 3, lscale)\n",
    "\n",
    "# prior covariance matrix of training points\n",
    "k_xx = rbf_kernel(x_samp.reshape(-1,1), None, 3, lscale)\n",
    "\n",
    "# prior covariance matrix of test points\n",
    "k_starstar = rbf_kernel(x.reshape(-1,1), None, 3, lscale)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c9eb330e",
   "metadata": {},
   "outputs": [],
   "source": [
    "print('test points prior covariance matrix shape = ' +str(k_starstar.shape))\n",
    "print('training-test point prior cross-covariance matrix shape = ' +str(k_starX.shape))\n",
    "print('training points prior covariance matrix shape = ' +str(k_xx.shape))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "98e08ec9",
   "metadata": {},
   "outputs": [],
   "source": [
    "try:\n",
    "    mu = k_starX.dot(np.linalg.inv(k_xx)).dot(y_samp).flatten()\n",
    "    var = k_starstar - k_starX.dot(np.linalg.inv(k_xx)).dot(k_starX.T)\n",
    "except:\n",
    "    print('ERROR PLEASE RETURN TO GENERATE SAMPLE')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "65afd2d7-8175-4707-8791-f16bd555550c",
   "metadata": {},
   "outputs": [],
   "source": [
    "print('test points posterior covariance matrix shape = ' +str(var.shape))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b65b08da",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #cce5ff; padding: 10px;\">\n",
    "\n",
    "If you get an error here please return to [generate sample](#generate-sample) section to check for duplicated sample points.\n",
    "\n",
    "</div>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b3b2bc70",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "    \n",
    "In the constrained (posterior) covariance matrix `var`, we can see that the correlation of neighbouring points is affected by the training data. If a predicted point lies on the training data, there is no correlation with other points. Therefore, the function must pass directly through it. Predicted values further away are also affected by the training data to an extent which decreases with distance.\n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "cc008204",
   "metadata": {},
   "outputs": [],
   "source": [
    "# heatmap of test points posterior covariance matrix\n",
    "fig, ax = plt.subplots()\n",
    "ax.imshow(var, cmap=\"BuPu\", interpolation='None')\n",
    "ax.axis(False)\n",
    "ax.set_title(\"Posterior covariance matrix\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "75426e81",
   "metadata": {},
   "outputs": [],
   "source": [
    "# functions from new mu and covariance matrix\n",
    "f_star = np.random.multivariate_normal(mu, var, 100)\n",
    "f_star_mean = f_star.mean(axis=0)\n",
    "f_star_std = f_star.std(axis=0)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3e49d531",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "\n",
    "Now we have our function predictions `f_star` that intercept all training points.\n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "86d813ac",
   "metadata": {},
   "outputs": [],
   "source": [
    "# plot predictions\n",
    "fig, ax = plt.subplots(figsize=(10,5))\n",
    "#plt.scatter(x, y, 100, 'k', 'o', zorder=100)\n",
    "ax.scatter(x_samp, y_samp, s=50, c=\"C1\", zorder=10)\n",
    "ax.plot(x, f_star.T, color='C0', alpha=0.1)\n",
    "# if you have altered the function you may need to alter the y lim\n",
    "ax.set_xlim([-6,6])\n",
    "ax.set_ylim([-10,15])\n",
    "ax.set_xlabel(\"x\")\n",
    "ax.set_ylabel(\"f(x)\")\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "9398b3a5",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "\n",
    "We can plot below the mean prediction function and the standard deviation. Away from training points, the standard deviation is much higher, reflecting the lack of knowledge in these regions of the x axis. \n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ee7df77f",
   "metadata": {},
   "outputs": [],
   "source": [
    "fig, ax = plt.subplots(figsize=(10,5))\n",
    "ax.scatter(x_samp, y_samp, s=50, c=\"C1\", zorder=10)\n",
    "ax.plot(x,y,color='r',alpha=0.2)\n",
    "ax.plot(x, f_star_mean, color='C0')\n",
    "ax.fill_between(x.flatten(), f_star_mean+f_star_std, f_star_mean-f_star_std, alpha=0.1, color='C0')\n",
    "ax.set_xlim([-6,6])\n",
    "# if you have altered the function you may need to alter the y lim\n",
    "ax.set_ylim([-10,15])\n",
    "ax.set_xlabel(\"x\")\n",
    "ax.set_ylabel(\"f(x)\")\n",
    "ax.legend(['training','actual','predicted'])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a1624a46",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "Since the RBF kernel is stationary, the posterior mean function will always return to $\\mu=0$ in regions further away from observed training data. In our example, this results in decreased accuracy of predictions as x tends to $\\pm \\infty$. This figure indicates that our prior uncertainty was too low, as the true function lies many standard deviations away from the prior mean (zero). Not much can be done about this in practice: it is essentially a consequence of trying to empirically model a rapidly diverging function without using domain-specific knowledge.\n",
    "</div>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "9851a817",
   "metadata": {},
   "source": [
    "# Sea Level Example\n",
    "\n",
    "<div style=\"background-color: #ccffcc; padding: 10px;\">\n",
    "\n",
    "Now let's look at an example using Gaussian Processes to predict sea-level change using a subset of sea-level data from the RISeR dataset, provided in Python NumPy arrays by Oliver Pollard.    \n",
    "\n",
    "Example of a global Relative Sea Level (RSL) change output for the Last Glacial Maximum to present (using the ICE-5G (Peltier et al., 2015) input to an implemented model (Han and Gomez, 2018) of sea-level theory (Kendall et al., 2005)).\n",
    "    \n",
    "<img src=\"images/sealevel_data.png\">\n",
    "\n",
    "        \n",
    "</div>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "hispanic-jones",
   "metadata": {
    "tags": [
     "hide-cell"
    ]
   },
   "source": [
    "# Load Data "
   ]
  },
  {
   "cell_type": "markdown",
   "id": "modular-arabic",
   "metadata": {
    "tags": [
     "hide-input"
    ]
   },
   "source": [
    "<div style=\"background-color: #cce5ff; padding: 10px;\">\n",
    "\n",
    "For this example we're using data provided in python [.npy files](https://towardsdatascience.com/what-is-npy-files-and-why-you-should-use-them-603373c78883) containing relative sea-level change at a single point in the southern North Sea at 122 (ka). `highstand` data we will use to test our predictions \n",
    "    \n",
    "<img src=\"images/sealevel_data_ts.png\">\n",
    "  \n",
    " </div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7ef40bb9",
   "metadata": {},
   "outputs": [],
   "source": [
    "highstand = np.load(\"data/highstand_data.npy\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "133dd0ed",
   "metadata": {},
   "source": [
    "<div style=\"background-color: #cce5ff; padding: 10px;\">\n",
    "    \n",
    "We can use 4 parameters provided in `parameter_data.npy` which correspond to the parameters listed below\n",
    " \n",
    "The Rates of Interglacial Sea-level change and Responses (RISeR) dataset covers the early interglacial period and can be used to show the relative contribution to the Penultimate Glacial Period (PGM) Max Ice\n",
    "Volume and Last Interglacial (LIG) Highstand from 4 parameters from sediment cores\n",
    "\n",
    "<img src=\"images/sea_level_parameters.png\">\n",
    "\n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "427b8674",
   "metadata": {},
   "outputs": [],
   "source": [
    "parameters = np.load('data/parameter_data.npy')"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "9c6e1a7b",
   "metadata": {},
   "source": [
    "# Normalise Data\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "6fcc73ef",
   "metadata": {},
   "outputs": [],
   "source": [
    "# useful functions\n",
    "def stack_parameters(*parameters):\n",
    "    return np.column_stack([[*parameters]]).T\n",
    "\n",
    "def normalise(parameters):\n",
    "    normalised_parameters = np.zeros_like(parameters)\n",
    "    normalisation_values = []\n",
    "    for index, parameter in enumerate(parameters.T):\n",
    "        shift = - np.min(parameter)\n",
    "        scale = np.max(parameter + shift)\n",
    "        normalised_parameters[:,index] = (parameter + shift)/scale\n",
    "        normalisation_values.append((shift, scale))\n",
    "        \n",
    "    return normalised_parameters, normalisation_values\n",
    "    "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "breeding-membership",
   "metadata": {},
   "outputs": [],
   "source": [
    "# move parameter ranges down to improve results\n",
    "p1 = parameters[:,0] # Bedrock\n",
    "p2 = parameters[:,1] # Onshore Sediment\n",
    "p3 = parameters[:,2] # Marine Sediment\n",
    "p4 = parameters[:,3] # Ice streams\n",
    "parmlist = ['Bedrock','Onshore Sediment','Marine Sediment','Ice streams']\n",
    "parameters = stack_parameters(p1, p2, p3, p4)\n",
    "highstand = highstand.reshape(-1,1)\n",
    "\n",
    "parameters_norm, parameters_norm_params = normalise(parameters)\n",
    "highstand_norm, highstand_norm_params = normalise(highstand)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b49828d5",
   "metadata": {},
   "source": [
    "# Plot Data\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "proud-edwards",
   "metadata": {},
   "outputs": [],
   "source": [
    "# plot data for parameter\n",
    "scatter_fig = plt.figure(figsize=[16,10])\n",
    "for index in range(4):\n",
    "    ax = scatter_fig.add_subplot(221+index)\n",
    "    ax.scatter(parameters_norm[:,index], highstand, s=15, c=\"#71a8ad\")\n",
    "    ax.set_title(f\"{str(parmlist[index])}\", fontsize=16, fontweight='bold')\n",
    "    ax.set_ylabel(\"Highstand (m)\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "fd286e00",
   "metadata": {},
   "source": [
    "## Plot parameters against other parameters"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "accessory-philip",
   "metadata": {},
   "outputs": [],
   "source": [
    "fig, ax = plt.subplots(nrows=4, ncols=4, figsize=(15,15))\n",
    "for index_1 in range(4):\n",
    "    for index_2 in range(4):\n",
    "        if index_1 == index_2:\n",
    "            ax[index_1, index_2].axis('off')\n",
    "            ax[index_1, index_2].text(0.5,0.5,f\"{str(parmlist[index_1])}\",horizontalalignment=\"center\",\n",
    "                                      verticalalignment=\"center\", fontsize=15, \n",
    "                                      fontweight='bold')\n",
    "        elif index_2 < index_1:\n",
    "            ax[index_1, index_2].axis('off')\n",
    "            \n",
    "        else:\n",
    "            ax[index_1, index_2].scatter(parameters_norm[:,index_2], parameters_norm[:,index_1],s=2,c=\"red\")\n",
    "            ax[index_1, index_2].set_xlim([0,1])\n",
    "            ax[index_1, index_2].set_ylim([0,1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "b896cfa8",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Over lay sparse points in red over the points to be predicted in blue"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "polish-direction",
   "metadata": {},
   "outputs": [],
   "source": [
    "predict_coords = np.load(\"data/predict_points.npy\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "4f08a106",
   "metadata": {},
   "outputs": [],
   "source": [
    "fig, ax = plt.subplots(nrows=4, ncols=4, figsize=(15,15))\n",
    "for index_1 in range(4):\n",
    "    for index_2 in range(4):\n",
    "        if index_1 == index_2:\n",
    "            ax[index_1, index_2].axis('off')\n",
    "            ax[index_1, index_2].text(0.5,0.5,f\"{str(parmlist[index_1])}\",\n",
    "                                      horizontalalignment=\"center\",\n",
    "                                      verticalalignment=\"center\", fontsize=15)\n",
    "        elif index_2 < index_1:\n",
    "            ax[index_1, index_2].axis('off')\n",
    "            \n",
    "        else:\n",
    "            ax[index_1, index_2].scatter(predict_coords[:,index_2], predict_coords[:,index_1],s=2,c=\"C0\", alpha=0.2)\n",
    "            ax[index_1, index_2].scatter(parameters_norm[:,index_2], parameters_norm[:,index_1],s=2,c=\"red\")\n",
    "            ax[index_1, index_2].set_xlim([0,1])\n",
    "            ax[index_1, index_2].set_ylim([0,1])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "96445033",
   "metadata": {},
   "source": [
    "## Define GP flow model\n",
    "\n",
    "<div style=\"background-color: #cce5ff; padding: 10px;\">\n",
    "\n",
    "We're going to use the Python library [GPflow](https://www.gpflow.org/) to build a more complex Gaussian-process model than that in our previous example, and in fewer lines of code.\n",
    "\n",
    "`k = gpflow.kernels.Matern52(lengthscales=lscale)`  selects a [Matérn covariance function](https://en.wikipedia.org/wiki/Mat%C3%A9rn_covariance_function) for the [GPflow kernel](https://gpflow.github.io/GPflow/develop/notebooks/getting_started/kernels.html).\n",
    "\n",
    "`m = gpflow.models.GPR(data=(X, Y), kernel=k, mean_function=None)` constructs a regression model from data points and the selected kernel.\n",
    "\n",
    "To inspect the chosen kernel, you can run the `print_summary` command, which should show you a list of hyperparameters: `variance` and `lengthscale`, and will display information about those hyperparameters (which will start at default values of 1), as well as the [transformation](https://en.wikipedia.org/wiki/Rectifier_(neural_networks)#Softplus) applied.\n",
    "\n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5be48c64",
   "metadata": {},
   "outputs": [],
   "source": [
    "gpflow.config.set_default_summary_fmt(\"notebook\")\n",
    "k = gpflow.kernels.Matern52(lengthscales=np.ones((len(parameters_norm.T))))\n",
    "print_summary(k)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7a60f04b",
   "metadata": {},
   "outputs": [],
   "source": [
    "m = gpflow.models.GPR(data=(parameters_norm, highstand_norm), kernel=k, mean_function=None)\n",
    "print_summary(m)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "2847d55e",
   "metadata": {},
   "source": [
    "## Optimization\n",
    "\n",
    "<div style=\"background-color: #cce5ff; padding: 10px;\">\n",
    "    \n",
    "`opt = gpflow.optimizers.Scipy()` uses the Scipy optimizer, which by default implements the [Limited-memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS-B)](https://en.wikipedia.org/wiki/Limited-memory_BFGS) algorithm.\n",
    "    \n",
    "`opt.minimize(m.training_loss, m.trainable_variables, options=dict(maxiter=100))` calls the minimize method of an optimizer which uses the training loss function defined by the GPflow model, the variables to train, and the number of training iterations.\n",
    "\n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d55b5932",
   "metadata": {},
   "outputs": [],
   "source": [
    "opt = gpflow.optimizers.Scipy()\n",
    "opt.minimize(m.training_loss, m.trainable_variables, options=dict(maxiter=100))\n",
    "print_summary(m)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5eb5be2f",
   "metadata": {},
   "source": [
    "## Prediction\n",
    "\n",
    "\n",
    "<div style=\"background-color: #cce5ff; padding: 10px;\">\n",
    "    \n",
    "`m.predict_f` make predictions at the new points, comprising a predictive mean and variance. We'll create an array of predicted means `predict` and variances `predict_var` by looping over the normalised training data from our parameters.     \n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7c51e8b3",
   "metadata": {},
   "outputs": [],
   "source": [
    "mean, var = m.predict_f(np.asarray([[0.2,0.5,0.6,0.7]]))\n",
    "print(mean.numpy()[0][0], var.numpy()[0][0])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "b9abf715",
   "metadata": {},
   "outputs": [],
   "source": [
    "mean, var = m.predict_f(predict_coords)\n",
    "print(mean.numpy()[:,0], var)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "roman-chase",
   "metadata": {},
   "outputs": [],
   "source": [
    "def coords_flattern(*coords):\n",
    "    meshs = np.meshgrid(*coords)\n",
    "    return np.column_stack(([mesh.ravel() for mesh in meshs]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "58b53e4f",
   "metadata": {},
   "outputs": [],
   "source": [
    "predict = []\n",
    "predict_var = []\n",
    "for index, point in enumerate(parameters_norm):\n",
    "    parameters_norm_training = np.delete(parameters_norm, index, axis=0)\n",
    "    highstand_norm_training = np.delete(highstand_norm, index, axis=0)\n",
    "    \n",
    "    k = gpflow.kernels.Matern52(lengthscales=np.ones((len(parameters.T))))\n",
    "    m = gpflow.models.GPR(data=(parameters_norm_training, highstand_norm_training), kernel=k, mean_function=None)\n",
    "    opt.minimize(m.training_loss, m.trainable_variables, options=dict(maxiter=100))\n",
    "    \n",
    "    lop = parameters_norm[index]\n",
    "    meanp, varp = m.predict_f(coords_flattern([lop[0]],[lop[1]],[lop[2]],[lop[3]]))\n",
    "    predict_var.append(varp.numpy()[0][0])\n",
    "    predict.append(meanp.numpy()[0][0])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f61ab311",
   "metadata": {},
   "source": [
    "## Cross Validation\n",
    "\n",
    "\n",
    "<div style=\"background-color: #cce5ff; padding: 10px;\">\n",
    "    \n",
    "We can now plot our predicted values (`predict`) against our actual values (`highstand`) *normalised*.\n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "bf1da56c",
   "metadata": {},
   "outputs": [],
   "source": [
    "fig, ax = plt.subplots(figsize=[6,6])\n",
    "ax.plot(np.linspace(0,1,num=100), np.linspace(0,1,num=100), c=\"#031f1d\")\n",
    "ax.plot(highstand_norm, predict, 'x', c=\"#71a8ad\")\n",
    "ax.errorbar(highstand_norm, predict, yerr=predict_var, linestyle=\"None\")\n",
    "ax.set_ylabel(\"Predicted Highstand Normalised\")\n",
    "ax.set_xlabel(\"Actual Highstand Normalised\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "794b84f8",
   "metadata": {},
   "source": [
    "## Plot Highstands\n",
    "\n",
    "<div style=\"background-color: #cce5ff; padding: 10px;\">\n",
    "    \n",
    "We can now plot the predictions (`mean`) in purple with the grey dots agreement with the actual data (`highstand_norm`)   \n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "e7c9ee16",
   "metadata": {},
   "outputs": [],
   "source": [
    "# plot data for parameter\n",
    "scatter_fig = plt.figure(figsize=[16,10])\n",
    "for index in range(4):\n",
    "    ax = scatter_fig.add_subplot(221+index)\n",
    "    ax.scatter(predict_coords[:,index], mean.numpy()[:,0], s=15, c=\"purple\")\n",
    "    ax.scatter(parameters_norm[:,index], highstand_norm, s=15, c=\"#71a8ad\")\n",
    "    ax.set_title(f\"{str(parmlist[index])}\", fontsize=16, fontweight='bold')\n",
    "    ax.set_ylabel(\"Highstand (m)\")\n",
    "    plt.legend(['mean','highstand_norm'])\n",
    "    \n",
    "plt.tight_layout()"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.11.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}