Benchmark_Depthwise_Convolution.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Benchmark Depthwise Convolution in Chainer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import chainer\n",
    "import numpy as np\n",
    "from numpy.testing import assert_allclose"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'4.2.0'"
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "chainer.__version__"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## On CPU"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "input = np.random.randn(1, 32, 112, 112).astype(np.float32)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "input_var = chainer.Variable(input)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "l_1 = chainer.Sequential(\n",
    "    chainer.links.Convolution2D(32, 32, ksize=3, stride=1, pad=1, groups=32, nobias=True),\n",
    "    chainer.links.Convolution2D(32, 32, ksize=3, stride=1, pad=1, groups=32, nobias=True),\n",
    "    chainer.links.Convolution2D(32, 32, ksize=3, stride=1, pad=1, groups=32, nobias=True),\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(32, 1, 3, 3)"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "l_1[0].W.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "weight_1 = np.random.randn(32, 1, 3, 3).astype(np.float32)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "for l in l_1:\n",
    "    l.W.array = weight_1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "231 ms ± 5.42 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "with chainer.using_config('enable_backprop', False):\n",
    "    for _ in range(10):\n",
    "        l_1(input_var)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "693 ms ± 17 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "for _ in range(10):\n",
    "    y = chainer.functions.sum(l_1(input_var))\n",
    "    l_1.cleargrads()\n",
    "    y.backward()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "l_2 = chainer.Sequential(\n",
    "    chainer.links.DepthwiseConvolution2D(32, 1, ksize=3, stride=1, pad=1, nobias=True),\n",
    "    chainer.links.DepthwiseConvolution2D(32, 1, ksize=3, stride=1, pad=1, nobias=True),\n",
    "    chainer.links.DepthwiseConvolution2D(32, 1, ksize=3, stride=1, pad=1, nobias=True),\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1, 32, 3, 3)"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "l_2[0].W.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "for l in l_2:\n",
    "    l.W.array = weight_1.transpose((1, 0, 2, 3))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "108 ms ± 9.6 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "with chainer.using_config('enable_backprop', False):\n",
    "    for _ in range(10):\n",
    "        l_2(input_var)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "430 ms ± 6.09 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "for _ in range(10):\n",
    "    y = chainer.functions.sum(l_2(input_var))\n",
    "    l_2.cleargrads()\n",
    "    y.backward()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "with chainer.using_config('enable_backprop', False):\n",
    "    assert_allclose(l_1(input_var).array, l_2(input_var).array, atol=1e-5, rtol=1e-3)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## On GPU"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<chainer.sequential.Sequential at 0x7f008463c128>"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "input_var.to_gpu()\n",
    "l_1.to_gpu()\n",
    "l_2.to_gpu()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "123 ms ± 17.1 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "with chainer.using_config('enable_backprop', False):\n",
    "    for _ in range(10):\n",
    "        l_1(input_var)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "117 ms ± 94.5 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "for _ in range(10):\n",
    "    y = chainer.functions.sum(l_1(input_var))\n",
    "    l_1.cleargrads()\n",
    "    y.backward()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "73.5 ms ± 83.3 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "with chainer.using_config('enable_backprop', False):\n",
    "    for _ in range(10):\n",
    "        l_2(input_var)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "The slowest run took 7.03 times longer than the fastest. This could mean that an intermediate result is being cached.\n",
      "134 ms ± 80.8 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "for _ in range(10):\n",
    "    y = chainer.functions.sum(l_2(input_var))\n",
    "    l_2.cleargrads()\n",
    "    y.backward()"
   ]
  }
 ],
 "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": "3.6.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}