notebooks/302-pytorch-quantization-aware-training/302-pytorch-quantization-aware-training.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "git68adWeq4l"
   },
   "source": [
    "# Quantization Aware Training with NNCF, using PyTorch framework\n",
    "\n",
    "This notebook is based on 'ImageNet training in PyTorch' [example](https://github.com/pytorch/examples/blob/master/imagenet/main.py).\n",
    "\n",
    "The goal of this notebook to demonstrate how to use the Neural Network Compression Framework [NNCF](https://github.com/openvinotoolkit/nncf) 8-bit quantization to optimize a PyTorch model for inference with OpenVINO Toolkit. The optimization process contains the following steps:\n",
    "* Transform the original FP32 model to INT8\n",
    "* Use fine-tuning to restore the accuracy\n",
    "* Export optimized and original models to ONNX and then to OpenVINO\n",
    "* Measure and compare the performance of models\n",
    "For more advanced usage, please refer to these [examples](https://github.com/openvinotoolkit/nncf/tree/develop/examples).\n",
    "\n",
    "We selected the ResNet-18 model with Tiny ImageNet-200 dataset. ResNet-18 is the version of ResNet models that contains the fewest layers (18) and Tiny ImageNet-200 is a subset of the larger ImageNet dataset that should be downloaded in the notebook. Using the smaller model and dataset will speed up training and download time. To see other ResNet models, visit [PyTorch hub](https://pytorch.org/hub/pytorch_vision_resnet/).\n",
    "\n",
    "> NOTE: _This notebook requires a C++ compiler._"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "6M1xndNu-z_2"
   },
   "source": [
    "## Imports and Settings\n",
    "\n",
    "On Windows, add the required C++ directories to the system's path.\n",
    "\n",
    "Import NNCF and all auxiliary packages from your Python* code.\n",
    "Set a name for the model, and the image width and height that will be used for the network. Also define paths where PyTorch, ONNX and OpenVINO IR versions of the models will be stored. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# On Windows, add the directory that contains cl.exe to the PATH to enable PyTorch to find the\n",
    "# required C++ tools. This code assumes that Visual Studio 2019 is installed in the default\n",
    "# directory. If you have a different C++ compiler, please add the correct path to os.environ[\"PATH\"]\n",
    "# directly.\n",
    "\n",
    "# Adding the path to os.environ[\"LIB\"] is not always required - it depends on the system's configuration\n",
    "\n",
    "import sys\n",
    "\n",
    "if sys.platform == \"win32\":\n",
    "    import distutils.command.build_ext\n",
    "    import os\n",
    "    from pathlib import Path\n",
    "\n",
    "    VS_INSTALL_DIR = r\"C:/Program Files (x86)/Microsoft Visual Studio\"\n",
    "    cl_paths = sorted(list(Path(VS_INSTALL_DIR).glob(\"**/Hostx86/x64/cl.exe\")))\n",
    "    if len(cl_paths) == 0:\n",
    "        raise ValueError(\n",
    "            \"Cannot find Visual Studio. This notebook requires C++. If you installed \"\n",
    "            \"a C++ compiler, please add the directory that contains cl.exe to \"\n",
    "            \"`os.environ['PATH']`\"\n",
    "        )\n",
    "    else:\n",
    "        # If multiple versions of MSVC are installed, get the most recent version\n",
    "        cl_path = cl_paths[-1]\n",
    "        vs_dir = str(cl_path.parent)\n",
    "        os.environ[\"PATH\"] += f\"{os.pathsep}{vs_dir}\"\n",
    "        # Code for finding the library dirs from\n",
    "        # https://stackoverflow.com/questions/47423246/get-pythons-lib-path\n",
    "        d = distutils.core.Distribution()\n",
    "        b = distutils.command.build_ext.build_ext(d)\n",
    "        b.finalize_options()\n",
    "        os.environ[\"LIB\"] = os.pathsep.join(b.library_dirs)\n",
    "        print(f\"Added {vs_dir} to PATH\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "BtaM_i2mEB0z",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "from pathlib import Path\n",
    "\n",
    "import sys\n",
    "import time\n",
    "import zipfile\n",
    "import platform\n",
    "\n",
    "# to disable warnings on export to ONNX\n",
    "import warnings\n",
    "from torch.jit import TracerWarning\n",
    "\n",
    "from urllib.request import urlretrieve\n",
    "\n",
    "import torch\n",
    "import nncf  # Important - should be imported directly after torch\n",
    "from nncf import NNCFConfig\n",
    "from nncf.torch import create_compressed_model\n",
    "from nncf.torch import register_default_init_args\n",
    "\n",
    "import torch.nn as nn\n",
    "import torch.nn.parallel\n",
    "import torch.optim\n",
    "import torch.utils.data\n",
    "import torch.utils.data.distributed\n",
    "import torchvision.datasets as datasets\n",
    "import torchvision.models as models\n",
    "import torchvision.transforms as transforms\n",
    "\n",
    "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
    "print(f\"Using {device} device\")\n",
    "\n",
    "MODEL_DIR = Path(\"model\")\n",
    "OUTPUT_DIR = Path(\"output\")\n",
    "DATA_DIR = Path(\"data\")\n",
    "BASE_MODEL_NAME = \"resnet18\"\n",
    "image_size = 64\n",
    "\n",
    "OUTPUT_DIR.mkdir(exist_ok=True)\n",
    "MODEL_DIR.mkdir(exist_ok=True)\n",
    "DATA_DIR.mkdir(exist_ok=True)\n",
    "\n",
    "# Paths where PyTorch, ONNX and OpenVINO IR models will be stored\n",
    "fp32_pth_path = Path(MODEL_DIR / (BASE_MODEL_NAME + \"_fp32\")).with_suffix(\".pth\")\n",
    "fp32_onnx_path = Path(OUTPUT_DIR / (BASE_MODEL_NAME + \"_fp32\")).with_suffix(\".onnx\")\n",
    "fp32_ir_path = fp32_onnx_path.with_suffix(\".xml\")\n",
    "int8_onnx_path = Path(OUTPUT_DIR / (BASE_MODEL_NAME + \"_int8\")).with_suffix(\".onnx\")\n",
    "int8_ir_path = int8_onnx_path.with_suffix(\".xml\")\n",
    "\n",
    "# It's possible to train FP32 model from scratch, but it might be slow. So the pre-trained weights are downloaded by default.\n",
    "pretrained = True\n",
    "fp32_pth_url = \"https://storage.openvinotoolkit.org/repositories/nncf/openvino_notebook_ckpts/302_resnet18_fp32.pth\"\n",
    "if pretrained:\n",
    "    urlretrieve(fp32_pth_url, fp32_pth_path)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "EIo5S145S0Ug",
    "outputId": "9a2db892-eb38-4863-dfdb-560aa12c8232",
    "pycharm": {
     "name": "#%% md\n"
    }
   },
   "source": [
    "Download Tiny ImageNet dataset\n",
    "* 100k images of shape 3x64x64\n",
    "* 200 different classes: snake, spider, cat, truck, grasshopper, gull, etc."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "-HxsU71bEbLS",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "def download_tiny_imagenet_200(\n",
    "    data_dir,\n",
    "    url=\"http://cs231n.stanford.edu/tiny-imagenet-200.zip\",\n",
    "    tarname=\"tiny-imagenet-200.zip\",\n",
    "):\n",
    "    data_dir.mkdir(exist_ok=True)\n",
    "    archive_path = data_dir / tarname\n",
    "    urlretrieve(url, archive_path)\n",
    "    zip_ref = zipfile.ZipFile(archive_path, \"r\")\n",
    "    zip_ref.extractall(path=data_dir)\n",
    "    zip_ref.close()\n",
    "    print(f\"Successfully downloaded and extracted dataset to: {data_dir}\")\n",
    "\n",
    "\n",
    "DATASET_DIR = DATA_DIR / \"tiny-imagenet-200\"\n",
    "if not DATASET_DIR.exists():\n",
    "    download_tiny_imagenet_200(DATA_DIR)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "eZX2GAh3W7ZT",
    "pycharm": {
     "name": "#%% md\n"
    }
   },
   "source": [
    "## Pre-train floating-point model\n",
    "Using NNCF for model compression assumes that the user has a pre-trained model and a training pipeline.\n",
    "\n",
    "Here we demonstrate one possible training pipeline: a ResNet-18 model pre-trained on 1000 classes from ImageNet is fine-tuned with 200 classes from Tiny-Imagenet. \n",
    "\n",
    "Subsequently, the training and validation functions will be reused as is for quantization-aware training.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "E01dMaR2_AFL"
   },
   "source": [
    "### Train function"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "940rcAIyiXml"
   },
   "outputs": [],
   "source": [
    "def train(train_loader, model, criterion, optimizer, epoch):\n",
    "    batch_time = AverageMeter(\"Time\", \":3.3f\")\n",
    "    losses = AverageMeter(\"Loss\", \":2.3f\")\n",
    "    top1 = AverageMeter(\"Acc@1\", \":2.2f\")\n",
    "    top5 = AverageMeter(\"Acc@5\", \":2.2f\")\n",
    "    progress = ProgressMeter(\n",
    "        len(train_loader), [batch_time, losses, top1, top5], prefix=\"Epoch:[{}]\".format(epoch)\n",
    "    )\n",
    "\n",
    "    # switch to train mode\n",
    "    model.train()\n",
    "\n",
    "    end = time.time()\n",
    "    for i, (images, target) in enumerate(train_loader):\n",
    "        images = images.to(device)\n",
    "        target = target.to(device)\n",
    "\n",
    "        # compute output\n",
    "        output = model(images)\n",
    "        loss = criterion(output, target)\n",
    "\n",
    "        # measure accuracy and record loss\n",
    "        acc1, acc5 = accuracy(output, target, topk=(1, 5))\n",
    "        losses.update(loss.item(), images.size(0))\n",
    "        top1.update(acc1[0], images.size(0))\n",
    "        top5.update(acc5[0], images.size(0))\n",
    "\n",
    "        # compute gradient and do opt step\n",
    "        optimizer.zero_grad()\n",
    "        loss.backward()\n",
    "        optimizer.step()\n",
    "\n",
    "        # measure elapsed time\n",
    "        batch_time.update(time.time() - end)\n",
    "        end = time.time()\n",
    "\n",
    "        print_frequency = 50\n",
    "        if i % print_frequency == 0:\n",
    "            progress.display(i)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "CoNr8qwm_El2"
   },
   "source": [
    "### Validate function"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "KgnugrWgicWC"
   },
   "outputs": [],
   "source": [
    "def validate(val_loader, model, criterion):\n",
    "    batch_time = AverageMeter(\"Time\", \":3.3f\")\n",
    "    losses = AverageMeter(\"Loss\", \":2.3f\")\n",
    "    top1 = AverageMeter(\"Acc@1\", \":2.2f\")\n",
    "    top5 = AverageMeter(\"Acc@5\", \":2.2f\")\n",
    "    progress = ProgressMeter(len(val_loader), [batch_time, losses, top1, top5], prefix=\"Test: \")\n",
    "\n",
    "    # switch to evaluate mode\n",
    "    model.eval()\n",
    "\n",
    "    with torch.no_grad():\n",
    "        end = time.time()\n",
    "        for i, (images, target) in enumerate(val_loader):\n",
    "            images = images.to(device)\n",
    "            target = target.to(device)\n",
    "\n",
    "            # compute output\n",
    "            output = model(images)\n",
    "            loss = criterion(output, target)\n",
    "\n",
    "            # measure accuracy and record loss\n",
    "            acc1, acc5 = accuracy(output, target, topk=(1, 5))\n",
    "            losses.update(loss.item(), images.size(0))\n",
    "            top1.update(acc1[0], images.size(0))\n",
    "            top5.update(acc5[0], images.size(0))\n",
    "\n",
    "            # measure elapsed time\n",
    "            batch_time.update(time.time() - end)\n",
    "            end = time.time()\n",
    "\n",
    "            print_frequency = 10\n",
    "            if i % print_frequency == 0:\n",
    "                progress.display(i)\n",
    "\n",
    "        print(\" * Acc@1 {top1.avg:.3f} Acc@5 {top5.avg:.3f}\".format(top1=top1, top5=top5))\n",
    "    return top1.avg"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "qMnYsGo9_MA8"
   },
   "source": [
    "### Helpers"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "R724tbxcidQE"
   },
   "outputs": [],
   "source": [
    "class AverageMeter(object):\n",
    "    \"\"\"Computes and stores the average and current value\"\"\"\n",
    "\n",
    "    def __init__(self, name, fmt=\":f\"):\n",
    "        self.name = name\n",
    "        self.fmt = fmt\n",
    "        self.reset()\n",
    "\n",
    "    def reset(self):\n",
    "        self.val = 0\n",
    "        self.avg = 0\n",
    "        self.sum = 0\n",
    "        self.count = 0\n",
    "\n",
    "    def update(self, val, n=1):\n",
    "        self.val = val\n",
    "        self.sum += val * n\n",
    "        self.count += n\n",
    "        self.avg = self.sum / self.count\n",
    "\n",
    "    def __str__(self):\n",
    "        fmtstr = \"{name} {val\" + self.fmt + \"} ({avg\" + self.fmt + \"})\"\n",
    "        return fmtstr.format(**self.__dict__)\n",
    "\n",
    "\n",
    "class ProgressMeter(object):\n",
    "    def __init__(self, num_batches, meters, prefix=\"\"):\n",
    "        self.batch_fmtstr = self._get_batch_fmtstr(num_batches)\n",
    "        self.meters = meters\n",
    "        self.prefix = prefix\n",
    "\n",
    "    def display(self, batch):\n",
    "        entries = [self.prefix + self.batch_fmtstr.format(batch)]\n",
    "        entries += [str(meter) for meter in self.meters]\n",
    "        print(\"\\t\".join(entries))\n",
    "\n",
    "    def _get_batch_fmtstr(self, num_batches):\n",
    "        num_digits = len(str(num_batches // 1))\n",
    "        fmt = \"{:\" + str(num_digits) + \"d}\"\n",
    "        return \"[\" + fmt + \"/\" + fmt.format(num_batches) + \"]\"\n",
    "\n",
    "\n",
    "def accuracy(output, target, topk=(1,)):\n",
    "    \"\"\"Computes the accuracy over the k top predictions for the specified values of k\"\"\"\n",
    "    with torch.no_grad():\n",
    "        maxk = max(topk)\n",
    "        batch_size = target.size(0)\n",
    "\n",
    "        _, pred = output.topk(maxk, 1, True, True)\n",
    "        pred = pred.t()\n",
    "        correct = pred.eq(target.view(1, -1).expand_as(pred))\n",
    "\n",
    "        res = []\n",
    "        for k in topk:\n",
    "            correct_k = correct[:k].reshape(-1).float().sum(0, keepdim=True)\n",
    "            res.append(correct_k.mul_(100.0 / batch_size))\n",
    "        return res"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "kcSjyLBwiqBx",
    "pycharm": {
     "name": "#%% md\n"
    }
   },
   "source": [
    "### Get a pre-trained FP32 model\n",
    "А pre-trained floating-point model is a prerequisite for quantization. It can be obtained by tuning from scratch  with the code below. However, this usually takes a lot of time. Therefore, we have already run this code and received good enough weights after 4 epochs (for the sake of simplicity, we did not tune until the best accuracy). By default, this notebook just loads these weights without launching training. To tune the model from scratch set `pretrained=False`.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "avCsioUYIaL7",
    "outputId": "183bdbb6-4016-463c-8d76-636a6b3a9778",
    "tags": [],
    "test_replace": {
     "80000, 20000": "300, 100",
     "dataset, [": "torch.utils.data.Subset(dataset, torch.arange(400)), ["
    }
   },
   "outputs": [],
   "source": [
    "num_classes = 200  # 200 is for Tiny ImageNet, default is 1000 for imagenet\n",
    "init_lr = 1e-4\n",
    "batch_size = 128\n",
    "epochs = 4\n",
    "\n",
    "# create model\n",
    "model = models.resnet18(pretrained=True)\n",
    "# update the last FC layer for Tiny ImageNet number of classes\n",
    "model.fc = nn.Linear(in_features=512, out_features=num_classes, bias=True)\n",
    "model.to(device)\n",
    "\n",
    "# Data loading code\n",
    "train_dir = DATASET_DIR / \"train\"\n",
    "normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])\n",
    "\n",
    "dataset = datasets.ImageFolder(\n",
    "    train_dir,\n",
    "    transforms.Compose(\n",
    "        [\n",
    "            transforms.Resize(image_size),\n",
    "            transforms.RandomHorizontalFlip(),\n",
    "            transforms.ToTensor(),\n",
    "            normalize,\n",
    "        ]\n",
    "    ),\n",
    ")\n",
    "train_dataset, val_dataset = torch.utils.data.random_split(\n",
    "    dataset, [80000, 20000], generator=torch.Generator().manual_seed(0)\n",
    ")\n",
    "\n",
    "train_loader = torch.utils.data.DataLoader(\n",
    "    train_dataset, batch_size=batch_size, shuffle=True, num_workers=4, pin_memory=True, sampler=None\n",
    ")\n",
    "\n",
    "val_loader = torch.utils.data.DataLoader(\n",
    "    val_dataset, batch_size=batch_size, shuffle=False, num_workers=4, pin_memory=True\n",
    ")\n",
    "\n",
    "# define loss function (criterion) and optimizer\n",
    "criterion = nn.CrossEntropyLoss().to(device)\n",
    "optimizer = torch.optim.Adam(model.parameters(), lr=init_lr)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "L0tH9KdwtHhV",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "acc1 = 0\n",
    "best_acc1 = 0\n",
    "if pretrained:\n",
    "    #\n",
    "    # ** WARNING: torch.load functionality uses Python's pickling module that\n",
    "    # may be used to perform arbitrary code execution during unpickling. Only load data that you\n",
    "    # trust.\n",
    "    #\n",
    "    checkpoint = torch.load(str(fp32_pth_path), map_location=\"cpu\")\n",
    "    model.load_state_dict(checkpoint[\"state_dict\"], strict=True)\n",
    "    best_acc1 = checkpoint[\"acc1\"]\n",
    "else:\n",
    "    # Training loop\n",
    "    for epoch in range(0, epochs):\n",
    "        # run a single training epoch\n",
    "        train(train_loader, model, criterion, optimizer, epoch)\n",
    "\n",
    "        # evaluate on validation set\n",
    "        acc1 = validate(val_loader, model, criterion)\n",
    "\n",
    "        is_best = acc1 > best_acc1\n",
    "        best_acc1 = max(acc1, best_acc1)\n",
    "\n",
    "        if is_best:\n",
    "            checkpoint = {\"state_dict\": model.state_dict(), \"acc1\": acc1}\n",
    "            torch.save(checkpoint, fp32_pth_path)\n",
    "\n",
    "print(f\"Accuracy of FP32 model: {best_acc1:.3f}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "pt_xNDDrJKsy",
    "outputId": "0925c801-0585-4431-98c9-de0decc4ad27",
    "pycharm": {
     "name": "#%% md\n"
    }
   },
   "source": [
    "Export the FP32 model to ONNX, which is supported by OpenVINO™ Toolkit, to benchmark it in comparison with the INT8 model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "9d8LOmKut36x",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "dummy_input = torch.randn(1, 3, image_size, image_size).to(device)\n",
    "\n",
    "torch.onnx.export(model, dummy_input, fp32_onnx_path)\n",
    "print(f\"FP32 ONNX model was exported to {fp32_onnx_path}.\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "qk0Nub3KiqZK",
    "outputId": "a438341a-8211-4750-be34-440ae483a511",
    "pycharm": {
     "name": "#%% md\n"
    }
   },
   "source": [
    "## Create and initialize quantization"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "pobVoHEoKcYp"
   },
   "source": [
    "NNCF enables compression-aware training by integrating into regular training pipelines. The framework is designed so that modifications to your original training code are minor. \n",
    "Quantization is the simplest scenario and requires only 3 modifications."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "ENAbqFpdWSlE",
    "outputId": "cd2701e3-e4a2-4a19-86cd-ae37f45cd64a",
    "pycharm": {
     "name": "#%% md\n"
    }
   },
   "source": [
    "1. Configure NNCF parameters to specify compression"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "_I_G-g9TPWkl",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "nncf_config_dict = {\n",
    "    \"input_info\": {\"sample_size\": [1, 3, image_size, image_size]},\n",
    "    \"log_dir\": str(OUTPUT_DIR),  # log directory for NNCF-specific logging outputs\n",
    "    \"compression\": {\n",
    "        \"algorithm\": \"quantization\",  # specify the algorithm here\n",
    "    },\n",
    "}\n",
    "nncf_config = NNCFConfig.from_dict(nncf_config_dict)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "2. Provide data loader to initialize the values of quantization ranges and determine which activation should be signed or unsigned from the collected statistics using a given number of samples.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "nncf_config = register_default_init_args(nncf_config, train_loader)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "3. Create a wrapped model ready for compression fine-tuning from a pre-trained FP32 model and configuration object."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "compression_ctrl, model = create_compressed_model(model, nncf_config)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Evaluate the new model on the validation set after initialization of quantization. The accuracy should be close to the accuracy of the floating-point FP32 model for a simple case like the one we are demonstrating now."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "acc1 = validate(val_loader, model, criterion)\n",
    "print(f\"Accuracy of initialized INT8 model: {acc1:.3f}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Fine-tune the compressed model\n",
    "\n",
    "At this step, a regular fine-tuning process is applied to restore accuracy drop. Normally, several epochs of tuning are required with a small learning rate, the same that is usually used at the end of the training of the original model. No other changes in the training pipeline are required. Here is a simple example."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "compression_lr = init_lr / 10\n",
    "optimizer = torch.optim.Adam(model.parameters(), lr=compression_lr)\n",
    "\n",
    "# train for one epoch with NNCF\n",
    "train(train_loader, model, criterion, optimizer, epoch=0)\n",
    "\n",
    "# evaluate on validation set after Quantization-Aware Training (QAT case)\n",
    "acc1 = validate(val_loader, model, criterion)\n",
    "\n",
    "print(f\"Accuracy of tuned INT8 model: {acc1:.3f}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Export INT8 model to ONNX"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "if not int8_onnx_path.exists():\n",
    "    warnings.filterwarnings(\"ignore\", category=TracerWarning)\n",
    "    warnings.filterwarnings(\"ignore\", category=UserWarning)\n",
    "    # Export INT8 model to ONNX that is supported by the OpenVINO™ toolkit\n",
    "    compression_ctrl.export_model(int8_onnx_path)\n",
    "    print(f\"INT8 ONNX model exported to {int8_onnx_path}.\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Export ONNX models to OpenVINO™ Intermediate Representation (IR)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Call the OpenVINO Model Optimizer tool to convert the ONNX model to OpenVINO IR, with FP16 precision. The models are saved to the current directory. We add the mean values to the model and scale the output with the standard deviation by --mean_values and --scale_values arguments. It is not necessary to normalize input data before propagating it through the network with these options.\n",
    "\n",
    "See the [Model Optimizer Developer Guide](https://docs.openvinotoolkit.org/latest/openvino_docs_MO_DG_Deep_Learning_Model_Optimizer_DevGuide.html) for more information about Model Optimizer.\n",
    "\n",
    "Executing this command may take a while. There may be some errors or warnings in the output. Model Optimization succesfully export to IR if the last lines of the output include: `[ SUCCESS ] Generated IR version 10 model`"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "if not fp32_ir_path.exists():\n",
    "    !mo --input_model $fp32_onnx_path --input_shape \"[1,3, $image_size, $image_size]\" --mean_values \"[123.675, 116.28 , 103.53]\" --scale_values \"[58.395, 57.12 , 57.375]\" --data_type FP16 --output_dir $OUTPUT_DIR"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "if not int8_ir_path.exists():\n",
    "    !mo --input_model $int8_onnx_path --input_shape \"[1,3, $image_size, $image_size]\" --mean_values \"[123.675, 116.28 , 103.53]\" --scale_values \"[58.395, 57.12 , 57.375]\" --data_type FP16 --output_dir $OUTPUT_DIR"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Benchmark model performance by computing inference time"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Finally, we will measure the inference performance of the FP32 and INT8 models. To do this, we use [Benchmark Tool](https://docs.openvinotoolkit.org/latest/openvino_inference_engine_tools_benchmark_tool_README.html) - OpenVINO's inference performance measurement tool. By default, Benchmark Tool runs inference for 60 seconds in asynchronous mode on CPU. It returns inference speed as latency (milliseconds per image) and throughput (frames per second) values."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "> **NOTE**: In this notebook we run benchmark_app for 15 seconds to give a quick indication of performance. For more accurate performance, we recommended running benchmark_app in a terminal/command prompt after closing other applications. Run benchmark_app -m model.xml -d CPU to benchmark async inference on CPU for one minute. Change CPU to GPU to benchmark on GPU. Run benchmark_app --help to see an overview of all command line options.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "def parse_benchmark_output(benchmark_output):\n",
    "    parsed_output = [line for line in benchmark_output if not (line.startswith(r\"[\") or line.startswith(\"  \") or line==\"\")]\n",
    "    print(*parsed_output, sep='\\n')\n",
    "\n",
    "print(f'Benchmark FP32 model (IR)')\n",
    "benchmark_output = ! benchmark_app -m $fp32_ir_path -d CPU -api async -t 15\n",
    "parse_benchmark_output(benchmark_output)\n",
    "\n",
    "print(f'Benchmark INT8 model (IR)')\n",
    "benchmark_output = ! benchmark_app -m $int8_ir_path -d CPU -api async -t 15\n",
    "parse_benchmark_output(benchmark_output)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Show CPU Information for reference"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "try:\n",
    "    import cpuinfo\n",
    "\n",
    "    print(cpuinfo.get_cpu_info()[\"brand_raw\"])\n",
    "except Exception:\n",
    "    # OpenVINO installs cpuinfo, but if a different version is installed\n",
    "    # the command above may not work\n",
    "    print(platform.processor())"
   ]
  }
 ],
 "metadata": {
  "accelerator": "GPU",
  "colab": {
   "collapsed_sections": [
    "K5HPrY_d-7cV",
    "E01dMaR2_AFL",
    "qMnYsGo9_MA8",
    "L0tH9KdwtHhV"
   ],
   "name": "NNCF Quantization PyTorch Demo (tiny-imagenet/resnet-18)",
   "provenance": []
  },
  "kernelspec": {
   "display_name": "openvino_env",
   "language": "python",
   "name": "openvino_env"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.9"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}