{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "from torchvision.models import vgg16,densenet121,resnet152\n",
    "from time import time\n",
    "import torch.nn as nn\n",
    "import torch.backends.cudnn as cudnn\n",
    "import torch.optim\n",
    "from torch.autograd import Variable\n",
    "import torchvision.models as models\n",
    "torch.backends.cudnn.benchmark=True\n",
    "model_names = sorted(name for name in models.__dict__\n",
    "    if name.islower() and not name.startswith(\"__\")\n",
    "    and callable(models.__dict__[name]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "cuda version= 9.0.176\n",
      "cudnn version= 7005\n",
      "densenet121 FP 32 avg over 100 runs: 71.4666223526001 ms\n",
      "densenet121 FP 16 avg over 100 runs: 59.28621768951416 ms\n",
      "vgg16 FP 32 avg over 100 runs: 87.8125810623169 ms\n",
      "vgg16 FP 16 avg over 100 runs: 58.735389709472656 ms\n",
      "resnet152 FP 32 avg over 100 runs: 143.52030992507935 ms\n",
      "resnet152 FP 16 avg over 100 runs: 92.21737623214722 ms\n"
     ]
    }
   ],
   "source": [
    "print('cuda version=', torch.version.cuda)\n",
    "print('cudnn version=', torch.backends.cudnn.version())\n",
    "\n",
    "for arch in ['densenet121', 'vgg16', 'resnet152']:\n",
    "    model   = models.__dict__[arch]().cuda()\n",
    "    loss   = nn.CrossEntropyLoss().cuda()\n",
    "    optimizer = torch.optim.SGD(model.parameters(), 0.001,\n",
    "                                        momentum=0.9,\n",
    "                                        weight_decay=1e-5)\n",
    "    durations = []\n",
    "    num_runs = 100\n",
    "\n",
    "    for i in range(num_runs + 1):\n",
    "        x = torch.rand(16, 3, 224, 224)\n",
    "        x_var = torch.autograd.Variable(x).cuda()\n",
    "        target = Variable(torch.LongTensor(16).fill_(1).cuda())\n",
    "        torch.cuda.synchronize()\n",
    "        t1 = time()\n",
    "        out = model(x_var)\n",
    "        err = loss(out, target)\n",
    "        err.backward()\n",
    "        optimizer.step()\n",
    "        torch.cuda.synchronize()\n",
    "        t2 = time()\n",
    "\n",
    "        # treat the initial run as warm up and don't count\n",
    "        if i > 0:\n",
    "            durations.append(t2 - t1)\n",
    "\n",
    "    print('{} FP 32 avg over {} runs: {} ms'.format(arch, len(durations), sum(durations) / len(durations) * 1000)) \n",
    "\n",
    "    model   = models.__dict__[arch]().cuda().half()\n",
    "    loss   = nn.CrossEntropyLoss().cuda()\n",
    "    optimizer = torch.optim.SGD(model.parameters(), 0.001,\n",
    "                                        momentum=0.9,\n",
    "                                        weight_decay=1e-5)\n",
    "    durations = []\n",
    "    num_runs = 100\n",
    "\n",
    "    for i in range(num_runs + 1):\n",
    "        x = torch.rand(16, 3, 224, 224)\n",
    "        x_var = torch.autograd.Variable(x).cuda().half()\n",
    "        target = Variable(torch.LongTensor(16).fill_(1).cuda())\n",
    "        torch.cuda.synchronize()\n",
    "        t1 = time()\n",
    "        out = model(x_var)\n",
    "        err = loss(out, target)\n",
    "        err.backward()\n",
    "        optimizer.step()\n",
    "        torch.cuda.synchronize()\n",
    "        t2 = time()\n",
    "\n",
    "        # treat the initial run as warm up and don't count\n",
    "        if i > 0:\n",
    "            durations.append(t2 - t1)\n",
    "\n",
    "    print('{} FP 16 avg over {} runs: {} ms'.format(arch, len(durations), sum(durations) / len(durations) * 1000))"
   ]
  },
  {
   "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": "3.6.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}