Merge pull request lukas#52 from charlesfrye/cf-by-hand

Adds interactive nbs for (over-)fitting by hand

Author: charlesfrye

examples/by-hand/data/xs.npy

@@ -0,0 +1,95 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Fitting a Model by Hand"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import matplotlib.pyplot as plt\n",
+    "import numpy as np\n",
+    "\n",
+    "from utils.models import Parameters, LinearModel\n",
+    "\n",
+    "plt.rcParams.update({'font.size': 18})"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%matplotlib notebook"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "xs, ys = np.load(\"data/xs.npy\")[:30], np.load(\"data/ys.npy\")[:30]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "parameters = Parameters([0.5, 0], [[-1, 1], [-1, 1]], [\"bias\", \"weight\"])\n",
+    "lm = LinearModel(input_values=np.linspace(0, 1), parameters=parameters)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "lm.plot()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "lm.set_data(xs, ys)\n",
+    "\n",
+    "lm.show_MSE = True\n",
+    "lm.make_interactive(log=True)"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "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": "3.7.3"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}

examples/by-hand/data/ys.npy

@@ -0,0 +1,76 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# (Over-)Fitting a Model by Hand"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import matplotlib.pyplot as plt\n",
+    "import numpy as np\n",
+    "\n",
+    "from utils.landmarks import LandmarksModel, setup_plot\n",
+    "\n",
+    "plt.rcParams.update({'font.size': 18})"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%matplotlib notebook"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "xs, ys = np.load(\"data/xs.npy\"), np.load(\"data/ys.npy\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "line = setup_plot()\n",
+    "landmarks_model = LandmarksModel(line, xs, ys, log=True)\n",
+    "\n",
+    "plt.show()"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "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": "3.7.3"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}

examples/by-hand/fitting.ipynb

@@ -0,0 +1,103 @@
+from matplotlib.lines import Line2D
+import matplotlib.pyplot as plt
+import numpy as np
+import wandb
+
+
+class LocalLandmarks(object):
+
+    def __init__(self, points, num=200):
+        self.points = np.array(points)
+        self.grid = np.linspace(0, 1, num=num)
+        self.pred_fun = np.array([self.grid, self.predict(self.grid)]).T
+
+    def predict(self, xs):
+        idxs_closest_x = []
+        for x in xs:
+            idxs_closest_x.append(np.argmin(np.abs(self.points[:, 0] - x)))
+        return self.points[idxs_closest_x, 1]
+
+
+class LandmarksModel(object):
+    """Interactively build a landmark-based model.
+    """
+    def __init__(self, landmarks_polygon, xs, ys, train_size=15, log=True):
+        self.landmarks_polygon = landmarks_polygon
+        self.xp = list(landmarks_polygon.get_xdata())
+        self.yp = list(landmarks_polygon.get_ydata())
+
+        self.canvas = landmarks_polygon.figure.canvas
+        self.ax_main = landmarks_polygon.axes
+
+        self.observed_xs = xs[:train_size]
+        self.observed_ys = ys[:train_size]
+
+        self.make_scatter(self.observed_xs, self.observed_ys, self.ax_main)
+        self.test_xs = xs[train_size:]
+        self.test_ys = ys[train_size:]
+
+        self.log = log
+
+        self.cid = self.canvas.mpl_connect('button_press_event', self)
+
+        prediction_line = Line2D([], [],
+                                 c=np.divide([255, 204, 51], 256), lw=4)
+        self.prediction_line_plot = self.ax_main.add_line(prediction_line)
+
+        if self.log:
+            wandb.init()
+
+    def __call__(self, event):
+        # Ignore clicks outside axes
+        if event.inaxes != self.landmarks_polygon.axes:
+            return
+
+        # Add point
+        self.xp.append(event.xdata)
+        self.yp.append(event.ydata)
+
+        self.landmarks_polygon.set_data(self.xp, self.yp)
+
+        # Rebuild prediction curve and update canvas
+        self.prediction_line_plot.set_data(*self._rebuild_predictor())
+        self._update()
+
+    def _update(self):
+        self.canvas.draw()
+
+    def _rebuild_predictor(self):
+        self.local_landmarks = LocalLandmarks(list(zip(self.xp, self.yp)))
+
+        train_MSE = self.compute_MSE(self.observed_xs, self.observed_ys)
+
+        if self.log:
+            test_MSE = self.compute_MSE(self.test_xs, self.test_ys)
+
+            wandb.log({"train_loss": train_MSE,
+                       "test_loss": test_MSE})
+
+        x, y = self.local_landmarks.pred_fun.T
+
+        return x, y
+
+    def compute_MSE(self, xs, ys):
+        predictions = self.local_landmarks.predict(xs)
+        MSE = np.mean(np.square(ys - predictions))
+        return MSE
+
+    def make_scatter(self, xs, ys, ax):
+        ax.scatter(xs, ys, color='k', alpha=0.5, s=72)
+
+
+def setup_plot():
+    fig, ax1 = plt.subplots(figsize=(10, 10))
+
+    line = Line2D([], [], ls='none', c='#616666',
+                  marker='x', mew=4, mec='k', ms=10, zorder=3)
+    ax1.add_line(line)
+
+    ax1.set_xlim(0, 1)
+    ax1.set_ylim(0, 1)
+    ax1.axis("off")
+
+    return line