hierachicalmodel_edward

hierachicalmodel_edward.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Edwardでの階層モデル"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- https://discourse.edwardlib.org/t/simple-hierarchical-model-fails/196\n",
    "- http://willwolf.io/2017/06/15/random-effects-neural-networks/   (original question)\n",
    "\n",
    "- https://aksarkar.github.io/nwas/klqp.html   (answer)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/usr/local/Cellar/python3/3.6.3/Frameworks/Python.framework/Versions/3.6/lib/python3.6/importlib/_bootstrap.py:219: RuntimeWarning: compiletime version 3.5 of module 'tensorflow.python.framework.fast_tensor_util' does not match runtime version 3.6\n",
      "  return f(*args, **kwds)\n",
      "/Users/apple/Library/Python/3.6/lib/python/site-packages/h5py/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.\n",
      "  from ._conv import register_converters as _register_converters\n"
     ]
    }
   ],
   "source": [
    "import edward as ed\n",
    "from edward.models import Normal\n",
    "import numpy as np\n",
    "import tensorflow as tf"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "# TOY DATA\n",
    "N = 3  # number of groups\n",
    "M = 1000  # samples per group\n",
    "\n",
    "# mean for each group is different\n",
    "# want to infer the group means plus the overall mean\n",
    "actual_group_means = [0.1, 0.2, 0.3]\n",
    "sigma = 0.1\n",
    "\n",
    "observed_groups = np.repeat([0, 1, 2], M)\n",
    "samples = [np.random.normal(actual_group_means[g], sigma, M) for g in range(N)]\n",
    "observed_data = np.concatenate(samples)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# MODEL"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "groups = tf.placeholder(tf.int32, [M * N])\n",
    "\n",
    "overall_mean = Normal(\n",
    "    loc=tf.zeros(1), \n",
    "    scale=tf.ones(1) * 0.05 )\n",
    "\n",
    "group_means = Normal(\n",
    "    loc=tf.ones(N) * overall_mean,\n",
    "    scale=tf.ones(N) * 0.05 )\n",
    "\n",
    "q_overall_mean = Normal(\n",
    "    loc=tf.Variable(tf.zeros(1)),\n",
    "    scale=tf.nn.softplus(tf.Variable(tf.zeros(1))) )\n",
    "\n",
    "q_group_means = Normal(\n",
    "    loc=tf.Variable(tf.zeros(N)),\n",
    "    scale=tf.nn.softplus(tf.Variable(tf.zeros(N))) )\n",
    "\n",
    "data = Normal(\n",
    "    loc=tf.gather(group_means, groups),\n",
    "    scale=tf.ones(shape=[N * M]) * sigma )\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/usr/local/lib/python3.6/site-packages/edward/util/random_variables.py:52: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.\n",
      "  not np.issubdtype(value.dtype, np.float) and \\\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1000/1000 [100%] ██████████████████████████████ Elapsed: 7s | Loss: -2533.115\n",
      "Using <class 'edward.inferences.klqp.ReparameterizationKLqp'>:\n",
      "[0.13311534]\n",
      "[0.0944147  0.20102091 0.2979718 ]\n",
      "1000/1000 [100%] ██████████████████████████████ Elapsed: 4s | Loss: -2493.819\n",
      "Using <class 'edward.inferences.klqp.ReparameterizationKLKLqp'>:\n",
      "[0.]\n",
      "[0.09101488 0.19820164 0.2882782 ]\n"
     ]
    }
   ],
   "source": [
    "#for inference_alg in (ed.KLpq, ed.KLqp):\n",
    "for inference_alg in (ed.ReparameterizationKLqp, ed.ReparameterizationKLKLqp):\n",
    "    inference = inference_alg(\n",
    "        {\n",
    "            overall_mean: q_overall_mean,\n",
    "            group_means: q_group_means\n",
    "        },\n",
    "        data={\n",
    "            groups: observed_groups,\n",
    "            data: observed_data\n",
    "        }\n",
    "    )\n",
    "    \n",
    "    inference.run(n_samples=5, n_iter=1000)\n",
    "    sess = ed.get_session()\n",
    "    print('Using {}:'.format(inference_alg))\n",
    "    print(q_overall_mean.mean().eval())\n",
    "    print(q_group_means.mean().eval())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "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": "3.6.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}