bnn.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "from jax import lax, jit\n",
    "from jax.config import config; config.update('jax_platform_name', 'gpu')\n",
    "import jax.numpy as np"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "N, D_X, D_H = 50, 3, 5\n",
    "X, Y = np.ones((N, D_X)), np.zeros(N)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "def potential_fn(w1):\n",
    "    z1 = np.matmul(X, w1)\n",
    "    w2 = np.ones((D_H, D_H))\n",
    "    z2 = np.matmul(z1, w2)\n",
    "    w3 = np.ones((D_H, 1))\n",
    "    z3 = np.matmul(z2, w3)\n",
    "    return np.sum((z3 - Y) ** 2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "@jit\n",
    "def loop(w):\n",
    "    def body_fn(i, state):\n",
    "        w, f = state\n",
    "        f = potential_fn(w)\n",
    "        return w, f\n",
    "\n",
    "    w, f = lax.fori_loop(0, 1000, body_fn, (w, 0.))\n",
    "    return f"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### CPU"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 118 ms, sys: 76.3 ms, total: 194 ms\n",
      "Wall time: 212 ms\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "DeviceArray(14062500., dtype=float32)"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "loop(np.ones((D_X, D_H)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 550 µs, sys: 0 ns, total: 550 µs\n",
      "Wall time: 450 µs\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "DeviceArray(14062500., dtype=float32)"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "loop(np.ones((D_X, D_H)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### GPU"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 429 ms, sys: 199 ms, total: 628 ms\n",
      "Wall time: 743 ms\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "DeviceArray(14062500., dtype=float32)"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "loop(np.ones((D_X, D_H)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 35.7 ms, sys: 0 ns, total: 35.7 ms\n",
      "Wall time: 34.4 ms\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "DeviceArray(14062500., dtype=float32)"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "loop(np.ones((D_X, D_H)))"
   ]
  }
 ],
 "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
}