iterative_nuts_memory.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import jax.numpy as np\n",
    "from jax import jit, lax, random\n",
    "from numpyro.hmc_util import build_tree, velocity_verlet"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "def kinetic_fn(m_inv, p):\n",
    "    return 0.5 * np.dot(m_inv, p ** 2)\n",
    "\n",
    "def potential_fn(q):\n",
    "    return 0.5 * np.sum(q ** 2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/fehiepsi/miniconda3/envs/pydata/lib/python3.6/site-packages/jax/lib/xla_bridge.py:144: UserWarning: No GPU/TPU found, falling back to CPU.\n",
      "  warnings.warn('No GPU/TPU found, falling back to CPU.')\n"
     ]
    }
   ],
   "source": [
    "vv_init, vv_update = velocity_verlet(potential_fn, kinetic_fn)\n",
    "step_size = 0.001  # force build full tree ~ 1000 leaves\n",
    "rng = random.PRNGKey(0)\n",
    "make_test = lambda dim: (vv_init(np.zeros(dim), np.ones(dim)), np.ones(dim))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "%load_ext memory_profiler"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "vv_state, inverse_mass_matrix = make_test(10000)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 4.18 s, sys: 23.3 ms, total: 4.2 s\n",
      "Wall time: 4.19 s\n",
      "136 ms ± 1.8 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)\n"
     ]
    }
   ],
   "source": [
    "@jit\n",
    "def f(vv_state):\n",
    "    return build_tree(vv_update, kinetic_fn, vv_state,\n",
    "                      inverse_mass_matrix, step_size, rng, iterative_build=True)\n",
    "\n",
    "%time f(vv_state)\n",
    "%timeit f(vv_state)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "def g(vv_state):\n",
    "    return build_tree(vv_update, kinetic_fn, vv_state,\n",
    "                      inverse_mass_matrix, step_size, rng, iterative_build=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "vv_state = vv_update(step_size, inverse_mass_matrix, vv_state)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "peak memory: 199.18 MiB, increment: 0.09 MiB\n"
     ]
    }
   ],
   "source": [
    "%%memit\n",
    "f(vv_state)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "peak memory: 200.36 MiB, increment: 1.18 MiB\n"
     ]
    }
   ],
   "source": [
    "%%memit\n",
    "g(vv_state)"
   ]
  }
 ],
 "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": "3.6.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}