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

{
 "cells": [
  {
   "attachments": {},
   "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](https://github.com/pytorch/examples/blob/master/imagenet/main.py).\n",
    "\n",
    "The goal of this notebook is 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",
    "\n",
    "* Transforming the original `FP32` model to `INT8`\n",
    "* Using fine-tuning to improve the accuracy.\n",
    "* Exporting optimized and original models to OpenVINO IR\n",
    "* Measuring and comparing the performance of models.\n",
    "\n",
    "For more advanced usage, refer to these [examples](https://github.com/openvinotoolkit/nncf/tree/develop/examples).\n",
    "\n",
    "This tutorial uses the ResNet-18 model with the Tiny ImageNet-200 dataset. ResNet-18 is the version of ResNet models that contains the fewest layers (18). Tiny ImageNet-200 is a subset of the larger ImageNet dataset with smaller images. The dataset will 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 for compiling PyTorch custom operations for quantization.\n",
    "> For Windows we recommend to install Visual Studio with C++ support, you can find instruction [here](https://learn.microsoft.com/en-us/cpp/build/vscpp-step-0-installation?view=msvc-170).\n",
    "> For MacOS `xcode-select --install` command installs many developer tools, including C++.\n",
    "> For Linux you can install gcc with your distribution's package manager.\n",
    "\n",
    "\n",
    "<img referrerpolicy=\"no-referrer-when-downgrade\" src=\"https://static.scarf.sh/a.png?x-pxid=5b5a4db0-7875-4bfb-bdbd-01698b5b1a77&file=notebooks/pytorch-quantization-aware-training/pytorch-quantization-aware-training.ipynb\" />\n"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "\n",
    "#### Table of contents:\n",
    "\n",
    "- [Imports and Settings](#Imports-and-Settings)\n",
    "- [Pre-train Floating-Point Model](#Pre-train-Floating-Point-Model)\n",
    "    - [Train Function](#Train-Function)\n",
    "    - [Validate Function](#Validate-Function)\n",
    "    - [Helpers](#Helpers)\n",
    "    - [Get a Pre-trained FP32 Model](#Get-a-Pre-trained-FP32-Model)\n",
    "- [Create and Initialize Quantization](#Create-and-Initialize-Quantization)\n",
    "- [Fine-tune the Compressed Model](#Fine-tune-the-Compressed-Model)\n",
    "- [Export INT8 Model to OpenVINO IR](#Export-INT8-Model-to-OpenVINO-IR)\n",
    "- [Benchmark Model Performance by Computing Inference Time](#Benchmark-Model-Performance-by-Computing-Inference-Time)\n",
    "\n",
    "\n",
    "### Installation Instructions\n",
    "\n",
    "This is a self-contained example that relies solely on its own code.\n",
    "\n",
    "We recommend  running the notebook in a virtual environment. You only need a Jupyter server to start.\n",
    "For details, please refer to [Installation Guide](https://github.com/openvinotoolkit/openvino_notebooks/blob/latest/README.md#-installation-guide)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Note: you may need to restart the kernel to use updated packages.\n",
      "Note: you may need to restart the kernel to use updated packages.\n"
     ]
    }
   ],
   "source": [
    "%pip install -q --extra-index-url https://download.pytorch.org/whl/cpu  \"openvino>=2024.0.0\" \"torch\" \"torchvision\" \"tqdm\"\n",
    "%pip install -q \"nncf>=2.9.0\""
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "6M1xndNu-z_2"
   },
   "source": [
    "## Imports and Settings\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "On Windows, add the required C++ directories to the system 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 and OpenVINO IR versions of the models will be stored. \n",
    "\n",
    "> **NOTE**: All NNCF logging messages below ERROR level (INFO and WARNING) are disabled to simplify the tutorial. For production use, it is recommended to enable logging by removing ```set_log_level(logging.ERROR)```."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "id": "BtaM_i2mEB0z",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Using cuda device\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "aaac9ef8ffe44d138d226367829c94c4",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "model/resnet18_fp32.pth:   0%|          | 0.00/43.1M [00:00<?, ?B/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/plain": [
       "PosixPath('/home/dlyakhov/Projects/openvino_notebooks/notebooks/pytorch-quantization-aware-training/model/resnet18_fp32.pth')"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import time\n",
    "import warnings  # To disable warnings on export model\n",
    "import zipfile\n",
    "from pathlib import Path\n",
    "\n",
    "import torch\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",
    "import openvino as ov\n",
    "from torch.jit import TracerWarning\n",
    "\n",
    "# Fetch `notebook_utils` module\n",
    "import requests\n",
    "\n",
    "if not Path(\"notebook_utils.py\").exists():\n",
    "    r = requests.get(\n",
    "        url=\"https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/notebook_utils.py\",\n",
    "    )\n",
    "\n",
    "    open(\"notebook_utils.py\", \"w\").write(r.text)\n",
    "from notebook_utils import download_file, device_widget\n",
    "\n",
    "torch.manual_seed(0)\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 and OpenVINO IR models will be stored.\n",
    "fp32_pth_path = Path(MODEL_DIR / (BASE_MODEL_NAME + \"_fp32\")).with_suffix(\".pth\")\n",
    "fp32_ir_path = fp32_pth_path.with_suffix(\".xml\")\n",
    "int8_ir_path = Path(MODEL_DIR / (BASE_MODEL_NAME + \"_int8\")).with_suffix(\".xml\")\n",
    "\n",
    "# It is possible to train FP32 model from scratch, but it might be slow. Therefore, the pre-trained weights are downloaded by default.\n",
    "pretrained_on_tiny_imagenet = True\n",
    "fp32_pth_url = \"https://storage.openvinotoolkit.org/repositories/nncf/openvino_notebook_ckpts/302_resnet18_fp32_v1.pth\"\n",
    "if not (MODEL_DIR / fp32_pth_path).exists():\n",
    "    download_file(fp32_pth_url, directory=MODEL_DIR, filename=fp32_pth_path.name)\n",
    "\n",
    "# Read more about telemetry collection at https://github.com/openvinotoolkit/openvino_notebooks?tab=readme-ov-file#-telemetry\n",
    "from notebook_utils import collect_telemetry\n",
    "\n",
    "collect_telemetry(\"pytorch-quantization-aware-training.ipynb\")"
   ]
  },
  {
   "attachments": {},
   "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",
    "\n",
    "* 100k images of shape 3x64x64\n",
    "* 200 different classes: snake, spider, cat, truck, grasshopper, gull, etc."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "id": "-HxsU71bEbLS",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "66722d64b53e40d0a0861c354d39113f",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "data/tiny-imagenet-200.zip:   0%|          | 0.00/237M [00:00<?, ?B/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Successfully downloaded and prepared dataset at: data/tiny-imagenet-200\n"
     ]
    }
   ],
   "source": [
    "def download_tiny_imagenet_200(\n",
    "    data_dir: Path,\n",
    "    url=\"http://cs231n.stanford.edu/tiny-imagenet-200.zip\",\n",
    "    tarname=\"tiny-imagenet-200.zip\",\n",
    "):\n",
    "    archive_path = data_dir / tarname\n",
    "    download_file(url, directory=data_dir, filename=tarname)\n",
    "    zip_ref = zipfile.ZipFile(archive_path, \"r\")\n",
    "    zip_ref.extractall(path=data_dir)\n",
    "    zip_ref.close()\n",
    "\n",
    "\n",
    "def prepare_tiny_imagenet_200(dataset_dir: Path):\n",
    "    # Format validation set the same way as train set is formatted.\n",
    "    val_data_dir = dataset_dir / \"val\"\n",
    "    val_annotations_file = val_data_dir / \"val_annotations.txt\"\n",
    "    with open(val_annotations_file, \"r\") as f:\n",
    "        val_annotation_data = map(lambda line: line.split(\"\\t\")[:2], f.readlines())\n",
    "    val_images_dir = val_data_dir / \"images\"\n",
    "    for image_filename, image_label in val_annotation_data:\n",
    "        from_image_filepath = val_images_dir / image_filename\n",
    "        to_image_dir = val_data_dir / image_label\n",
    "        if not to_image_dir.exists():\n",
    "            to_image_dir.mkdir()\n",
    "        to_image_filepath = to_image_dir / image_filename\n",
    "        from_image_filepath.rename(to_image_filepath)\n",
    "    val_annotations_file.unlink()\n",
    "    val_images_dir.rmdir()\n",
    "\n",
    "\n",
    "DATASET_DIR = DATA_DIR / \"tiny-imagenet-200\"\n",
    "if not DATASET_DIR.exists():\n",
    "    download_tiny_imagenet_200(DATA_DIR)\n",
    "    prepare_tiny_imagenet_200(DATASET_DIR)\n",
    "    print(f\"Successfully downloaded and prepared dataset at: {DATASET_DIR}\")"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "eZX2GAh3W7ZT",
    "pycharm": {
     "name": "#%% md\n"
    }
   },
   "source": [
    "## Pre-train Floating-Point Model\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "Using NNCF for model compression assumes that a pre-trained model and a training pipeline are already in use.\n",
    "\n",
    "This tutorial demonstrates 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"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "E01dMaR2_AFL"
   },
   "source": [
    "### Train Function\n",
    "[back to top ⬆️](#Table-of-contents:)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "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),\n",
    "        [batch_time, losses, top1, top5],\n",
    "        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)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "CoNr8qwm_El2"
   },
   "source": [
    "### Validate Function\n",
    "[back to top ⬆️](#Table-of-contents:)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "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"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "qMnYsGo9_MA8"
   },
   "source": [
    "### Helpers\n",
    "[back to top ⬆️](#Table-of-contents:)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "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"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "kcSjyLBwiqBx",
    "pycharm": {
     "name": "#%% md\n"
    }
   },
   "source": [
    "### Get a Pre-trained FP32 Model\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\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, this code has already been run and received good enough weights after 4 epochs (for the sake of simplicity, tuning was not done until the best accuracy). By default, this notebook just loads these weights without launching training. To train the model yourself on a model pre-trained on ImageNet, set `pretrained_on_tiny_imagenet = False` in the Imports and Settings section at the top of this notebook."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "avCsioUYIaL7",
    "outputId": "183bdbb6-4016-463c-8d76-636a6b3a9778",
    "tags": [],
    "test_replace": {
     "train_dataset,": "torch.utils.data.Subset(train_dataset, torch.arange(300)), ",
     "val_dataset, ": "torch.utils.data.Subset(val_dataset, torch.arange(100)), "
    }
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/dlyakhov/env/tmp/lib/python3.8/site-packages/torchvision/models/_utils.py:208: UserWarning: The parameter 'pretrained' is deprecated since 0.13 and may be removed in the future, please use 'weights' instead.\n",
      "  warnings.warn(\n",
      "/home/dlyakhov/env/tmp/lib/python3.8/site-packages/torchvision/models/_utils.py:223: UserWarning: Arguments other than a weight enum or `None` for 'weights' are deprecated since 0.13 and may be removed in the future. The current behavior is equivalent to passing `weights=None`.\n",
      "  warnings.warn(msg)\n"
     ]
    }
   ],
   "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",
    "model = models.resnet18(pretrained=not pretrained_on_tiny_imagenet)\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",
    "val_dir = DATASET_DIR / \"val\"\n",
    "normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])\n",
    "\n",
    "train_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",
    "val_dataset = datasets.ImageFolder(\n",
    "    val_dir,\n",
    "    transforms.Compose(\n",
    "        [\n",
    "            transforms.Resize(image_size),\n",
    "            transforms.ToTensor(),\n",
    "            normalize,\n",
    "        ]\n",
    "    ),\n",
    ")\n",
    "\n",
    "train_loader = torch.utils.data.DataLoader(\n",
    "    train_dataset,\n",
    "    batch_size=batch_size,\n",
    "    shuffle=True,\n",
    "    num_workers=0,\n",
    "    pin_memory=True,\n",
    "    sampler=None,\n",
    ")\n",
    "\n",
    "val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=0, pin_memory=True)\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": 9,
   "metadata": {
    "id": "L0tH9KdwtHhV",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy of FP32 model: 55.520\n"
     ]
    }
   ],
   "source": [
    "if pretrained_on_tiny_imagenet:\n",
    "    #\n",
    "    # ** WARNING: The `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",
    "    acc1_fp32 = checkpoint[\"acc1\"]\n",
    "else:\n",
    "    best_acc1 = 0\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",
    "    acc1_fp32 = best_acc1\n",
    "\n",
    "print(f\"Accuracy of FP32 model: {acc1_fp32:.3f}\")"
   ]
  },
  {
   "attachments": {},
   "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 OpenVINO™ Intermediate Representation, to benchmark it in comparison with the `INT8` model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "id": "9d8LOmKut36x",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "FP32 model was exported to model/resnet18_fp32.xml.\n"
     ]
    }
   ],
   "source": [
    "dummy_input = torch.randn(1, 3, image_size, image_size).to(device)\n",
    "\n",
    "ov_model = ov.convert_model(model, example_input=dummy_input, input=[1, 3, image_size, image_size])\n",
    "ov.save_model(ov_model, fp32_ir_path, compress_to_fp16=False)\n",
    "print(f\"FP32 model was exported to {fp32_ir_path}.\")"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "pobVoHEoKcYp"
   },
   "source": [
    "## Create and Initialize Quantization\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "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 requires only 2 modifications."
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "ENAbqFpdWSlE",
    "outputId": "cd2701e3-e4a2-4a19-86cd-ae37f45cd64a",
    "pycharm": {
     "name": "#%% md\n"
    }
   },
   "source": [
    "1. Create a quantization data loader with batch size equal to one and wrap it by the `nncf.Dataset`, specifying a transformation function which prepares input data to fit into model during quantization. In our case, to pick input tensor from pair (input tensor and label)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "id": "_I_G-g9TPWkl",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "INFO:nncf:NNCF initialized successfully. Supported frameworks detected: torch, tensorflow, onnx, openvino\n"
     ]
    }
   ],
   "source": [
    "import nncf\n",
    "\n",
    "\n",
    "def transform_fn(data_item):\n",
    "    return data_item[0]\n",
    "\n",
    "\n",
    "# Creating separate dataloader with batch size = 1\n",
    "# as dataloaders with batches > 1 is not supported yet.\n",
    "quantization_loader = torch.utils.data.DataLoader(val_dataset, batch_size=1, shuffle=False, num_workers=0, pin_memory=True)\n",
    "\n",
    "quantization_dataset = nncf.Dataset(quantization_loader, transform_fn)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "2. Run `nncf.quantize` for Getting an Optimized Model.\n",
    "\n",
    "`nncf.quantize` function accepts model and prepared quantization dataset for performing basic quantization. Optionally, additional parameters like `subset_size`, `preset`, `ignored_scope` can be provided to improve quantization result if applicable. More details about supported parameters can be found on this [page](https://docs.openvino.ai/2024/openvino-workflow/model-optimization-guide/quantizing-models-post-training/basic-quantization-flow.html#tune-quantization-parameters)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "2024-01-17 15:43:43.543878: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.\n",
      "2024-01-17 15:43:43.579576: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.\n",
      "To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.\n",
      "2024-01-17 15:43:44.170538: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "WARNING:nncf:NNCF provides best results with torch==2.1.0, while current torch version is 1.13.0+cu117. If you encounter issues, consider switching to torch==2.1.0\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "efcf6cd485e745bb920296275e772aab",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Output()"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "<pre style=\"white-space:pre;overflow-x:auto;line-height:normal;font-family:Menlo,'DejaVu Sans Mono',consolas,'Courier New',monospace\"></pre>\n"
      ],
      "text/plain": []
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "<pre style=\"white-space:pre;overflow-x:auto;line-height:normal;font-family:Menlo,'DejaVu Sans Mono',consolas,'Courier New',monospace\">\n",
       "</pre>\n"
      ],
      "text/plain": [
       "\n"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "INFO:nncf:Compiling and loading torch extension: quantized_functions_cuda...\n",
      "INFO:nncf:Finished loading torch extension: quantized_functions_cuda\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "0f4d0f4cbf774827aaa7123ebd6b9bbb",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Output()"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "<pre style=\"white-space:pre;overflow-x:auto;line-height:normal;font-family:Menlo,'DejaVu Sans Mono',consolas,'Courier New',monospace\"></pre>\n"
      ],
      "text/plain": []
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "<pre style=\"white-space:pre;overflow-x:auto;line-height:normal;font-family:Menlo,'DejaVu Sans Mono',consolas,'Courier New',monospace\">\n",
       "</pre>\n"
      ],
      "text/plain": [
       "\n"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "quantized_model = nncf.quantize(model, quantization_dataset)"
   ]
  },
  {
   "attachments": {},
   "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 being demonstrated here."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Test: [ 0/79]\tTime 0.110 (0.110)\tLoss 0.992 (0.992)\tAcc@1 78.12 (78.12)\tAcc@5 89.06 (89.06)\n",
      "Test: [10/79]\tTime 0.069 (0.074)\tLoss 1.990 (1.623)\tAcc@1 44.53 (60.37)\tAcc@5 79.69 (83.95)\n",
      "Test: [20/79]\tTime 0.068 (0.072)\tLoss 1.814 (1.704)\tAcc@1 60.16 (58.26)\tAcc@5 80.47 (82.63)\n",
      "Test: [30/79]\tTime 0.068 (0.071)\tLoss 2.284 (1.794)\tAcc@1 52.34 (56.75)\tAcc@5 67.97 (80.90)\n",
      "Test: [40/79]\tTime 0.070 (0.072)\tLoss 1.618 (1.831)\tAcc@1 61.72 (55.64)\tAcc@5 82.03 (80.37)\n",
      "Test: [50/79]\tTime 0.069 (0.071)\tLoss 1.951 (1.832)\tAcc@1 57.81 (55.70)\tAcc@5 75.00 (80.06)\n",
      "Test: [60/79]\tTime 0.070 (0.071)\tLoss 1.795 (1.855)\tAcc@1 56.25 (55.28)\tAcc@5 84.38 (79.75)\n",
      "Test: [70/79]\tTime 0.069 (0.071)\tLoss 2.359 (1.888)\tAcc@1 47.66 (54.79)\tAcc@5 74.22 (79.08)\n",
      " * Acc@1 55.130 Acc@5 79.680\n",
      "Accuracy of initialized INT8 model: 55.130\n"
     ]
    }
   ],
   "source": [
    "acc1 = validate(val_loader, quantized_model, criterion)\n",
    "print(f\"Accuracy of initialized INT8 model: {acc1:.3f}\")"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Fine-tune the Compressed Model\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "At this step, a regular fine-tuning process is applied to further improve quantized model accuracy. 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": 14,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch:[0][  0/782]\tTime 0.284 (0.284)\tLoss 0.876 (0.876)\tAcc@1 78.12 (78.12)\tAcc@5 92.97 (92.97)\n",
      "Epoch:[0][ 50/782]\tTime 0.112 (0.116)\tLoss 0.796 (0.808)\tAcc@1 80.47 (79.96)\tAcc@5 94.53 (94.27)\n",
      "Epoch:[0][100/782]\tTime 0.111 (0.114)\tLoss 0.785 (0.788)\tAcc@1 82.81 (80.52)\tAcc@5 92.19 (94.56)\n",
      "Epoch:[0][150/782]\tTime 0.114 (0.113)\tLoss 0.653 (0.785)\tAcc@1 84.38 (80.69)\tAcc@5 95.31 (94.45)\n",
      "Epoch:[0][200/782]\tTime 0.109 (0.113)\tLoss 0.804 (0.780)\tAcc@1 80.47 (80.92)\tAcc@5 94.53 (94.45)\n",
      "Epoch:[0][250/782]\tTime 0.111 (0.113)\tLoss 0.756 (0.777)\tAcc@1 83.59 (80.98)\tAcc@5 94.53 (94.47)\n",
      "Epoch:[0][300/782]\tTime 0.112 (0.112)\tLoss 0.665 (0.772)\tAcc@1 82.03 (81.07)\tAcc@5 96.88 (94.53)\n",
      "Epoch:[0][350/782]\tTime 0.115 (0.112)\tLoss 0.661 (0.767)\tAcc@1 82.81 (81.14)\tAcc@5 97.66 (94.57)\n",
      "Epoch:[0][400/782]\tTime 0.111 (0.113)\tLoss 0.661 (0.764)\tAcc@1 78.91 (81.24)\tAcc@5 96.09 (94.60)\n",
      "Epoch:[0][450/782]\tTime 0.119 (0.113)\tLoss 0.904 (0.762)\tAcc@1 79.69 (81.27)\tAcc@5 89.06 (94.60)\n",
      "Epoch:[0][500/782]\tTime 0.113 (0.113)\tLoss 0.609 (0.757)\tAcc@1 84.38 (81.46)\tAcc@5 96.88 (94.62)\n",
      "Epoch:[0][550/782]\tTime 0.112 (0.113)\tLoss 0.833 (0.753)\tAcc@1 76.56 (81.59)\tAcc@5 95.31 (94.69)\n",
      "Epoch:[0][600/782]\tTime 0.112 (0.113)\tLoss 0.768 (0.751)\tAcc@1 82.81 (81.63)\tAcc@5 95.31 (94.69)\n",
      "Epoch:[0][650/782]\tTime 0.112 (0.113)\tLoss 0.750 (0.751)\tAcc@1 82.81 (81.61)\tAcc@5 93.75 (94.71)\n",
      "Epoch:[0][700/782]\tTime 0.110 (0.113)\tLoss 0.654 (0.749)\tAcc@1 84.38 (81.62)\tAcc@5 96.09 (94.71)\n",
      "Epoch:[0][750/782]\tTime 0.110 (0.113)\tLoss 0.575 (0.748)\tAcc@1 86.72 (81.67)\tAcc@5 97.66 (94.73)\n",
      "Test: [ 0/79]\tTime 0.070 (0.070)\tLoss 1.028 (1.028)\tAcc@1 78.91 (78.91)\tAcc@5 86.72 (86.72)\n",
      "Test: [10/79]\tTime 0.070 (0.070)\tLoss 1.827 (1.514)\tAcc@1 46.88 (63.35)\tAcc@5 79.69 (84.02)\n",
      "Test: [20/79]\tTime 0.073 (0.070)\tLoss 1.628 (1.594)\tAcc@1 64.06 (60.97)\tAcc@5 82.03 (83.78)\n",
      "Test: [30/79]\tTime 0.069 (0.070)\tLoss 2.061 (1.688)\tAcc@1 57.03 (59.25)\tAcc@5 71.88 (82.26)\n",
      "Test: [40/79]\tTime 0.070 (0.070)\tLoss 1.495 (1.738)\tAcc@1 66.41 (57.93)\tAcc@5 85.16 (81.59)\n",
      "Test: [50/79]\tTime 0.069 (0.070)\tLoss 1.863 (1.741)\tAcc@1 58.59 (57.83)\tAcc@5 76.56 (81.31)\n",
      "Test: [60/79]\tTime 0.069 (0.070)\tLoss 1.571 (1.779)\tAcc@1 65.62 (57.21)\tAcc@5 84.38 (80.74)\n",
      "Test: [70/79]\tTime 0.069 (0.070)\tLoss 2.505 (1.809)\tAcc@1 46.09 (56.78)\tAcc@5 75.00 (80.22)\n",
      " * Acc@1 57.200 Acc@5 80.880\n",
      "Accuracy of tuned INT8 model: 57.200\n",
      "Accuracy drop of tuned INT8 model over pre-trained FP32 model: -1.680\n"
     ]
    }
   ],
   "source": [
    "compression_lr = init_lr / 10\n",
    "optimizer = torch.optim.Adam(quantized_model.parameters(), lr=compression_lr)\n",
    "\n",
    "# Train for one epoch with NNCF.\n",
    "train(train_loader, quantized_model, criterion, optimizer, epoch=0)\n",
    "\n",
    "# Evaluate on validation set after Quantization-Aware Training (QAT case).\n",
    "acc1_int8 = validate(val_loader, quantized_model, criterion)\n",
    "\n",
    "print(f\"Accuracy of tuned INT8 model: {acc1_int8:.3f}\")\n",
    "print(f\"Accuracy drop of tuned INT8 model over pre-trained FP32 model: {acc1_fp32 - acc1_int8:.3f}\")"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Export INT8 Model to OpenVINO IR\n",
    "[back to top ⬆️](#Table-of-contents:)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "WARNING:tensorflow:Please fix your imports. Module tensorflow.python.training.tracking.base has been moved to tensorflow.python.trackable.base. The old module will be deleted in version 2.11.\n",
      "INT8 Omodel exported to model/resnet18_int8.xml.\n"
     ]
    }
   ],
   "source": [
    "if not int8_ir_path.exists():\n",
    "    warnings.filterwarnings(\"ignore\", category=TracerWarning)\n",
    "    warnings.filterwarnings(\"ignore\", category=UserWarning)\n",
    "    # Export INT8 model to OpenVINO™ IR\n",
    "    ov_model = ov.convert_model(quantized_model, example_input=dummy_input, input=[1, 3, image_size, image_size])\n",
    "    ov.save_model(ov_model, int8_ir_path)\n",
    "    print(f\"INT8 model exported to {int8_ir_path}.\")"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Benchmark Model Performance by Computing Inference Time\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "Finally, measure the inference performance of the `FP32` and `INT8` models, using [Benchmark Tool](https://docs.openvino.ai/2024/learn-openvino/openvino-samples/benchmark-tool.html) - inference performance measurement tool in OpenVINO. 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.\n",
    "\n",
    "> **NOTE**: This notebook runs `benchmark_app` for 15 seconds to give a quick indication of performance. For more accurate performance, it is recommended to run `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."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "device = device_widget()\n",
    "\n",
    "device"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Benchmark FP32 model (IR)\n",
      "[ INFO ] Throughput:   3755.92 FPS\n",
      "Benchmark INT8 model (IR)\n",
      "[ INFO ] Throughput:   15141.53 FPS\n"
     ]
    }
   ],
   "source": [
    "def parse_benchmark_output(benchmark_output):\n",
    "    parsed_output = [line for line in benchmark_output if \"FPS\" in line]\n",
    "    print(*parsed_output, sep=\"\\n\")\n",
    "\n",
    "\n",
    "print(\"Benchmark FP32 model (IR)\")\n",
    "benchmark_output = ! benchmark_app -m $fp32_ir_path -d $device.value -api async -t 15\n",
    "parse_benchmark_output(benchmark_output)\n",
    "\n",
    "print(\"Benchmark INT8 model (IR)\")\n",
    "benchmark_output = ! benchmark_app -m $int8_ir_path -d $device.value -api async -t 15\n",
    "parse_benchmark_output(benchmark_output)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Show Device Information for reference."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'Intel(R) Core(TM) i9-10980XE CPU @ 3.00GHz'"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import openvino.properties as props\n",
    "\n",
    "\n",
    "core = ov.Core()\n",
    "core.get_property(device.value, props.device.full_name)"
   ]
  }
 ],
 "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": "Python 3 (ipykernel)",
   "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": "3.8.10"
  },
  "openvino_notebooks": {
   "imageUrl": "",
   "tags": {
    "categories": [
     "Model Training",
     "Optimize"
    ],
    "libraries": [],
    "other": [],
    "tasks": [
     "Image Classification"
    ]
   }
  },
  "vscode": {
   "interpreter": {
    "hash": "916dbcbb3f70747c44a77c7bcd40155683ae19c65e1c03b4aa3499c5328201f1"
   }
  },
  "widgets": {
   "application/vnd.jupyter.widget-state+json": {
    "state": {},
    "version_major": 2,
    "version_minor": 0
   }
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}