PyTorch Performance Analysis

PyTorch-Performance-Analysis.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1. Libraries"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'1.4.0'"
      ]
     },
     "execution_count": 1,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import cProfile\n",
    "import time\n",
    "\n",
    "import numpy as np\n",
    "import torch\n",
    "import torch.utils.data as torchdata\n",
    "\n",
    "from multiprocessing import cpu_count\n",
    "\n",
    "\n",
    "torch.__version__"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 2. DataLoaderでnum_workers > 0を設定して早くなるのはどんな時か?"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Dataset定義\n",
    "\n",
    "オプションをつけることでわざと遅くする機能を持たせます。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "class DatasetWrapper(torchdata.Dataset):\n",
    "    def __init__(self, slow=False, wait=0.1):\n",
    "        self.slow = slow\n",
    "        self.wait = wait\n",
    "        \n",
    "    def __len__(self):\n",
    "        return 128\n",
    "    \n",
    "    def __getitem__(self, idx):\n",
    "        if self.slow:\n",
    "            time.sleep(self.wait)\n",
    "        return np.random.random(10)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### num_workes > 0に指定して効果検証"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "# parameters\n",
    "batch_size = 32\n",
    "shuffle = False\n",
    "pin_memory = False"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "cpuの数: 8\n"
     ]
    }
   ],
   "source": [
    "print(\"cpuの数:\", cpu_count())\n",
    "fastloader = torchdata.DataLoader(\n",
    "    DatasetWrapper(), \n",
    "    num_workers=0, \n",
    "    batch_size=batch_size,\n",
    "    shuffle=shuffle,\n",
    "    pin_memory=pin_memory)\n",
    "\n",
    "fastloader_multi = torchdata.DataLoader(\n",
    "    DatasetWrapper(), \n",
    "    num_workers=cpu_count() // 2, \n",
    "    batch_size=batch_size,\n",
    "    shuffle=shuffle,\n",
    "    pin_memory=pin_memory)\n",
    "\n",
    "slowloader = torchdata.DataLoader(\n",
    "    DatasetWrapper(slow=True, wait=0.01), \n",
    "    num_workers=0, \n",
    "    batch_size=batch_size,\n",
    "    shuffle=shuffle,\n",
    "    pin_memory=pin_memory)\n",
    "\n",
    "slowloader_multi = torchdata.DataLoader(\n",
    "    DatasetWrapper(slow=True, wait=0.01), \n",
    "    num_workers=cpu_count() // 2, \n",
    "    batch_size=batch_size,\n",
    "    shuffle=shuffle,\n",
    "    pin_memory=pin_memory)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### `DataLoader`のロード時間 < `for`文内の場合"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 4.31 ms, sys: 1.77 ms, total: 6.09 ms\n",
      "Wall time: 4.01 s\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "for batch in fastloader:\n",
    "    time.sleep(1.0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 14.3 ms, sys: 18.9 ms, total: 33.1 ms\n",
      "Wall time: 4.16 s\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "for batch in fastloader_multi:\n",
    "    time.sleep(1.0)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### `DataLoader`のロード時間 > `for`文内の場合"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 13.3 ms, sys: 4.77 ms, total: 18.1 ms\n",
      "Wall time: 5.52 s\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "for batch in slowloader:\n",
    "    time.sleep(1.0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 13.2 ms, sys: 20.3 ms, total: 33.5 ms\n",
      "Wall time: 4.4 s\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "for batch in slowloader_multi:\n",
    "    time.sleep(1.0)"
   ]
  },
  {
   "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": "3.6.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}