Pushing the docs to dev/ for branch: main, commit 2cf3c55807bd644a14b87d3cc1d5d2cc652c995b

Author: sklearn-ci

dev/_downloads/07fcc19ba03226cd3d83d4e40ec44385/auto_examples_python.zip

@@ -15,7 +15,7 @@
       "cell_type": "markdown",
       "metadata": {},
       "source": [
-        "\n# Gaussian process regression (GPR) with noise-level estimation\n\nThis example shows the ability of the\n:class:`~sklearn.gaussian_process.kernels.WhiteKernel` to estimate the noise\nlevel in the data. Moreover, we show the importance of kernel hyperparameters\ninitialization.\n"
+        "\n# Gaussian process regression (GPR) with noise-level estimation\n\nThis example illustrates that GPR with a sum-kernel including a WhiteKernel can\nestimate the noise level of data. An illustration of the\nlog-marginal-likelihood (LML) landscape shows that there exist two local\nmaxima of LML. The first corresponds to a model with a high noise level and a\nlarge length scale, which explains all variations in the data by noise. The\nsecond one has a smaller noise level and shorter length scale, which explains\nmost of the variation by the noise-free functional relationship. The second\nmodel has a higher likelihood; however, depending on the initial value for the\nhyperparameters, the gradient-based optimization might also converge to the\nhigh-noise solution. It is thus important to repeat the optimization several\ntimes for different initializations.\n"
       ]
     },
     {
@@ -26,177 +26,7 @@
       },
       "outputs": [],
       "source": [
-        "# Authors: Jan Hendrik Metzen <[email protected]>\n#          Guillaume Lemaitre <[email protected]>\n# License: BSD 3 clause"
-      ]
-    },
-    {
-      "cell_type": "markdown",
-      "metadata": {},
-      "source": [
-        "## Data generation\n\nWe will work in a setting where `X` will contain a single feature. We create a\nfunction that will generate the target to be predicted. We will add an\noption to add some noise to the generated target.\n\n"
-      ]
-    },
-    {
-      "cell_type": "code",
-      "execution_count": null,
-      "metadata": {
-        "collapsed": false
-      },
-      "outputs": [],
-      "source": [
-        "import numpy as np\n\n\ndef target_generator(X, add_noise=False):\n    target = 0.5 + np.sin(3 * X)\n    if add_noise:\n        rng = np.random.RandomState(1)\n        target += rng.normal(0, 0.3, size=target.shape)\n    return target.squeeze()"
-      ]
-    },
-    {
-      "cell_type": "markdown",
-      "metadata": {},
-      "source": [
-        "Let's have a look to the target generator where we will not add any noise to\nobserve the signal that we would like to predict.\n\n"
-      ]
-    },
-    {
-      "cell_type": "code",
-      "execution_count": null,
-      "metadata": {
-        "collapsed": false
-      },
-      "outputs": [],
-      "source": [
-        "X = np.linspace(0, 5, num=30).reshape(-1, 1)\ny = target_generator(X, add_noise=False)"
-      ]
-    },
-    {
-      "cell_type": "code",
-      "execution_count": null,
-      "metadata": {
-        "collapsed": false
-      },
-      "outputs": [],
-      "source": [
-        "import matplotlib.pyplot as plt\n\nplt.plot(X, y, label=\"Expected signal\")\nplt.legend()\nplt.xlabel(\"X\")\n_ = plt.ylabel(\"y\")"
-      ]
-    },
-    {
-      "cell_type": "markdown",
-      "metadata": {},
-      "source": [
-        "The target is transforming the input `X` using a sine function. Now, we will\ngenerate few noisy training samples. To illustrate the noise level, we will\nplot the true signal together with the noisy training samples.\n\n"
-      ]
-    },
-    {
-      "cell_type": "code",
-      "execution_count": null,
-      "metadata": {
-        "collapsed": false
-      },
-      "outputs": [],
-      "source": [
-        "rng = np.random.RandomState(0)\nX_train = rng.uniform(0, 5, size=20).reshape(-1, 1)\ny_train = target_generator(X_train, add_noise=True)"
-      ]
-    },
-    {
-      "cell_type": "code",
-      "execution_count": null,
-      "metadata": {
-        "collapsed": false
-      },
-      "outputs": [],
-      "source": [
-        "plt.plot(X, y, label=\"Expected signal\")\nplt.scatter(\n    x=X_train[:, 0],\n    y=y_train,\n    color=\"black\",\n    alpha=0.4,\n    label=\"Observations\",\n)\nplt.legend()\nplt.xlabel(\"X\")\n_ = plt.ylabel(\"y\")"
-      ]
-    },
-    {
-      "cell_type": "markdown",
-      "metadata": {},
-      "source": [
-        "## Optimisation of kernel hyperparameters in GPR\n\nNow, we will create a\n:class:`~sklearn.gaussian_process.GaussianProcessRegressor`\nusing an additive kernel adding a\n:class:`~sklearn.gaussian_process.kernels.RBF` and\n:class:`~sklearn.gaussian_process.kernels.WhiteKernel` kernels.\nThe :class:`~sklearn.gaussian_process.kernels.WhiteKernel` is a kernel that\nwill able to estimate the amount of noise present in the data while the\n:class:`~sklearn.gaussian_process.kernels.RBF` will serve at fitting the\nnon-linearity between the data and the target.\n\nHowever, we will show that the hyperparameter space contains several local\nminima. It will highlights the importance of initial hyperparameter values.\n\nWe will create a model using a kernel with a high noise level and a large\nlength scale, which will explain all variations in the data by noise.\n\n"
-      ]
-    },
-    {
-      "cell_type": "code",
-      "execution_count": null,
-      "metadata": {
-        "collapsed": false
-      },
-      "outputs": [],
-      "source": [
-        "from sklearn.gaussian_process import GaussianProcessRegressor\nfrom sklearn.gaussian_process.kernels import RBF, WhiteKernel\n\nkernel = 1.0 * RBF(length_scale=1e1, length_scale_bounds=(1e-2, 1e3)) + WhiteKernel(\n    noise_level=1, noise_level_bounds=(1e-5, 1e1)\n)\ngpr = GaussianProcessRegressor(kernel=kernel, alpha=0.0)\ngpr.fit(X_train, y_train)\ny_mean, y_std = gpr.predict(X, return_std=True)"
-      ]
-    },
-    {
-      "cell_type": "code",
-      "execution_count": null,
-      "metadata": {
-        "collapsed": false
-      },
-      "outputs": [],
-      "source": [
-        "plt.plot(X, y, label=\"Expected signal\")\nplt.scatter(x=X_train[:, 0], y=y_train, color=\"black\", alpha=0.4, label=\"Observsations\")\nplt.errorbar(X, y_mean, y_std)\nplt.legend()\nplt.xlabel(\"X\")\nplt.ylabel(\"y\")\n_ = plt.title(\n    f\"Initial: {kernel}\\nOptimum: {gpr.kernel_}\\nLog-Marginal-Likelihood: \"\n    f\"{gpr.log_marginal_likelihood(gpr.kernel_.theta)}\",\n    fontsize=8,\n)"
-      ]
-    },
-    {
-      "cell_type": "markdown",
-      "metadata": {},
-      "source": [
-        "We see that the optimum kernel found still have a high noise level and\nan even larger length scale. Furthermore, we observe that the\nmodel does not provide faithful predictions.\n\nNow, we will initialize the\n:class:`~sklearn.gaussian_process.kernels.RBF` with a\nlarger `length_scale` and the\n:class:`~sklearn.gaussian_process.kernels.WhiteKernel`\nwith a smaller noise level lower bound.\n\n"
-      ]
-    },
-    {
-      "cell_type": "code",
-      "execution_count": null,
-      "metadata": {
-        "collapsed": false
-      },
-      "outputs": [],
-      "source": [
-        "kernel = 1.0 * RBF(length_scale=1e-1, length_scale_bounds=(1e-2, 1e3)) + WhiteKernel(\n    noise_level=1e-2, noise_level_bounds=(1e-10, 1e1)\n)\ngpr = GaussianProcessRegressor(kernel=kernel, alpha=0.0)\ngpr.fit(X_train, y_train)\ny_mean, y_std = gpr.predict(X, return_std=True)"
-      ]
-    },
-    {
-      "cell_type": "code",
-      "execution_count": null,
-      "metadata": {
-        "collapsed": false
-      },
-      "outputs": [],
-      "source": [
-        "plt.plot(X, y, label=\"Expected signal\")\nplt.scatter(x=X_train[:, 0], y=y_train, color=\"black\", alpha=0.4, label=\"Observations\")\nplt.errorbar(X, y_mean, y_std)\nplt.legend()\nplt.xlabel(\"X\")\nplt.ylabel(\"y\")\n_ = plt.title(\n    f\"Initial: {kernel}\\nOptimum: {gpr.kernel_}\\nLog-Marginal-Likelihood: \"\n    f\"{gpr.log_marginal_likelihood(gpr.kernel_.theta)}\",\n    fontsize=8,\n)"
-      ]
-    },
-    {
-      "cell_type": "markdown",
-      "metadata": {},
-      "source": [
-        "First, we see that the model's predictions are more precise than the\nprevious model's: this new model is able to estimate the noise-free\nfunctional relationship.\n\nLooking at the kernel hyperparameters, we see that the best combination found\nhas a smaller noise level and shorter length scale than the first model.\n\nWe can inspect the Log-Marginal-Likelihood (LML) of\n:class:`~sklearn.gaussian_process.GaussianProcessRegressor`\nfor different hyperparameters to get a sense of the local minima.\n\n"
-      ]
-    },
-    {
-      "cell_type": "code",
-      "execution_count": null,
-      "metadata": {
-        "collapsed": false
-      },
-      "outputs": [],
-      "source": [
-        "from matplotlib.colors import LogNorm\n\nlength_scale = np.logspace(-2, 4, num=50)\nnoise_level = np.logspace(-2, 1, num=50)\nlength_scale_grid, noise_level_grid = np.meshgrid(length_scale, noise_level)\n\nlog_marginal_likelihood = [\n    gpr.log_marginal_likelihood(theta=np.log([0.36, scale, noise]))\n    for scale, noise in zip(length_scale_grid.ravel(), noise_level_grid.ravel())\n]\nlog_marginal_likelihood = np.reshape(\n    log_marginal_likelihood, newshape=noise_level_grid.shape\n)"
-      ]
-    },
-    {
-      "cell_type": "code",
-      "execution_count": null,
-      "metadata": {
-        "collapsed": false
-      },
-      "outputs": [],
-      "source": [
-        "vmin, vmax = (-log_marginal_likelihood).min(), 50\nlevel = np.around(np.logspace(np.log10(vmin), np.log10(vmax), num=50), decimals=1)\nplt.contour(\n    length_scale_grid,\n    noise_level_grid,\n    -log_marginal_likelihood,\n    levels=level,\n    norm=LogNorm(vmin=vmin, vmax=vmax),\n)\nplt.colorbar()\nplt.xscale(\"log\")\nplt.yscale(\"log\")\nplt.xlabel(\"Length-scale\")\nplt.ylabel(\"Noise-level\")\nplt.title(\"Log-marginal-likelihood\")\nplt.show()"
-      ]
-    },
-    {
-      "cell_type": "markdown",
-      "metadata": {},
-      "source": [
-        "We see that there are two local minima that correspond to the combination\nof hyperparameters previously found. Depending on the initial values for the\nhyperparameters, the gradient-based optimization might converge whether or\nnot to the best model. It is thus important to repeat the optimization\nseveral times for different initializations.\n\n"
+        "# Authors: Jan Hendrik Metzen <[email protected]>\n#\n# License: BSD 3 clause\n\nimport numpy as np\n\nfrom matplotlib import pyplot as plt\nfrom matplotlib.colors import LogNorm\n\nfrom sklearn.gaussian_process import GaussianProcessRegressor\nfrom sklearn.gaussian_process.kernels import RBF, WhiteKernel\n\n\nrng = np.random.RandomState(0)\nX = rng.uniform(0, 5, 20)[:, np.newaxis]\ny = 0.5 * np.sin(3 * X[:, 0]) + rng.normal(0, 0.5, X.shape[0])\n\n# First run\nplt.figure()\nkernel = 1.0 * RBF(length_scale=100.0, length_scale_bounds=(1e-2, 1e3)) + WhiteKernel(\n    noise_level=1, noise_level_bounds=(1e-10, 1e1)\n)\ngp = GaussianProcessRegressor(kernel=kernel, alpha=0.0).fit(X, y)\nX_ = np.linspace(0, 5, 100)\ny_mean, y_cov = gp.predict(X_[:, np.newaxis], return_cov=True)\nplt.plot(X_, y_mean, \"k\", lw=3, zorder=9)\nplt.fill_between(\n    X_,\n    y_mean - np.sqrt(np.diag(y_cov)),\n    y_mean + np.sqrt(np.diag(y_cov)),\n    alpha=0.5,\n    color=\"k\",\n)\nplt.plot(X_, 0.5 * np.sin(3 * X_), \"r\", lw=3, zorder=9)\nplt.scatter(X[:, 0], y, c=\"r\", s=50, zorder=10, edgecolors=(0, 0, 0))\nplt.title(\n    \"Initial: %s\\nOptimum: %s\\nLog-Marginal-Likelihood: %s\"\n    % (kernel, gp.kernel_, gp.log_marginal_likelihood(gp.kernel_.theta))\n)\nplt.tight_layout()\n\n# Second run\nplt.figure()\nkernel = 1.0 * RBF(length_scale=1.0, length_scale_bounds=(1e-2, 1e3)) + WhiteKernel(\n    noise_level=1e-5, noise_level_bounds=(1e-10, 1e1)\n)\ngp = GaussianProcessRegressor(kernel=kernel, alpha=0.0).fit(X, y)\nX_ = np.linspace(0, 5, 100)\ny_mean, y_cov = gp.predict(X_[:, np.newaxis], return_cov=True)\nplt.plot(X_, y_mean, \"k\", lw=3, zorder=9)\nplt.fill_between(\n    X_,\n    y_mean - np.sqrt(np.diag(y_cov)),\n    y_mean + np.sqrt(np.diag(y_cov)),\n    alpha=0.5,\n    color=\"k\",\n)\nplt.plot(X_, 0.5 * np.sin(3 * X_), \"r\", lw=3, zorder=9)\nplt.scatter(X[:, 0], y, c=\"r\", s=50, zorder=10, edgecolors=(0, 0, 0))\nplt.title(\n    \"Initial: %s\\nOptimum: %s\\nLog-Marginal-Likelihood: %s\"\n    % (kernel, gp.kernel_, gp.log_marginal_likelihood(gp.kernel_.theta))\n)\nplt.tight_layout()\n\n# Plot LML landscape\nplt.figure()\ntheta0 = np.logspace(-2, 3, 49)\ntheta1 = np.logspace(-2, 0, 50)\nTheta0, Theta1 = np.meshgrid(theta0, theta1)\nLML = [\n    [\n        gp.log_marginal_likelihood(np.log([0.36, Theta0[i, j], Theta1[i, j]]))\n        for i in range(Theta0.shape[0])\n    ]\n    for j in range(Theta0.shape[1])\n]\nLML = np.array(LML).T\n\nvmin, vmax = (-LML).min(), (-LML).max()\nvmax = 50\nlevel = np.around(np.logspace(np.log10(vmin), np.log10(vmax), 50), decimals=1)\nplt.contour(Theta0, Theta1, -LML, levels=level, norm=LogNorm(vmin=vmin, vmax=vmax))\nplt.colorbar()\nplt.xscale(\"log\")\nplt.yscale(\"log\")\nplt.xlabel(\"Length-scale\")\nplt.ylabel(\"Noise-level\")\nplt.title(\"Log-marginal-likelihood\")\nplt.tight_layout()\n\nplt.show()"
       ]
     }
   ],