initial_centroids_demo.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import faiss\n",
    "from matplotlib import pyplot"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_dataset_2(d, nt, nb, nq):\n",
    "    \"\"\"A dataset that is not completely random but still challenging to\n",
    "    index\n",
    "    \"\"\"\n",
    "    d1 = 10     # intrinsic dimension (more or less)\n",
    "    n = nb + nt + nq\n",
    "    rs = np.random.RandomState(1338)\n",
    "    x = rs.normal(size=(n, d1))\n",
    "    x = np.dot(x, rs.rand(d1, d))\n",
    "    # now we have a d1-dim ellipsoid in d-dimensional space\n",
    "    # higher factor (>4) -> higher frequency -> less linear\n",
    "    x = x * (rs.rand(d) * 4 + 0.1)\n",
    "    x = np.sin(x)\n",
    "    x = x.astype('float32')\n",
    "    return x[:nt], x[nt:nt + nb], x[nt + nb:]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# K-means with initialization"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Generate training sets"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "xt1, xt2, _ = get_dataset_2(32, 10000, 10000, 0)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Initial training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "km = faiss.Kmeans(32, 200, niter=25)\n",
    "km.train(xt1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "bento_obj_id": "139742568626384",
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "pyplot.plot(km.obj)\n",
    "pyplot.xlabel('iterations')\n",
    "pyplot.ylabel('k-means objective')\n",
    "pyplot.grid(True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Fine-tuning with warm-start from initial training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "km2 = faiss.Kmeans(32, 200, niter=10)\n",
    "km2.train(xt2, init_centroids=km.centroids)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "bento_obj_id": "139742544453264",
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "pyplot.plot(np.hstack((km.obj, km2.obj)))\n",
    "pyplot.xlabel('iterations')\n",
    "pyplot.ylabel('k-means objective')\n",
    "pyplot.grid(True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Initialization for IVF quantizer"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This is more involved because the clusters cannot be set directly. However, we can do the training process manually. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [],
   "source": [
    "xt, xb, xq = get_dataset_2(32, 10000, 10000, 100)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We want to initialize with this"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [],
   "source": [
    "init_centroids = xb[:200]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "index = faiss.index_factory(32, 'PCA16,IVF200,Flat')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Train VectorTransform manually"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [],
   "source": [
    "pca = index.chain.at(0)\n",
    "pca.train(xt)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Manual k-means"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "metadata": {},
   "outputs": [],
   "source": [
    "km = faiss.Kmeans(16, 200)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "49671.68359375"
      ]
     },
     "execution_count": 50,
     "metadata": {
      "bento_obj_id": "139742942267408"
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "init_centroids_pca = pca.apply_py(init_centroids)\n",
    "xt_pca = pca.apply_py(xt)\n",
    "km.train(xt_pca, init_centroids=init_centroids_pca)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Set the centroids manually"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [],
   "source": [
    "index_ivf = faiss.downcast_index(faiss.extract_index_ivf(index))\n",
    "index_ivf.quantizer.add(km.centroids)\n",
    "index.is_trained = index_ivf.is_trained = True"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now the index is ready"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "metadata": {},
   "outputs": [],
   "source": [
    "index.add(xb)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 55,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(array([[4.6906796, 5.5750985],\n",
       "        [2.8803923, 3.8141544],\n",
       "        [4.432521 , 4.4821587],\n",
       "        [3.9308074, 4.6096234]], dtype=float32), array([[2080, 8407],\n",
       "        [4281, 1780],\n",
       "        [2927, 8969],\n",
       "        [2089, 9907]]))"
      ]
     },
     "execution_count": 55,
     "metadata": {
      "bento_obj_id": "139742558282528"
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "index.search(xq[:4], 2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "bento_stylesheets": {
   "bento/extensions/flow/main.css": true,
   "bento/extensions/kernel_selector/main.css": true,
   "bento/extensions/kernel_ui/main.css": true,
   "bento/extensions/new_kernel/main.css": true,
   "bento/extensions/system_usage/main.css": true,
   "bento/extensions/theme/main.css": true
  },
  "kernelspec": {
   "display_name": "faiss",
   "language": "python",
   "name": "bento_kernel_faiss"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.5+"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}