Post-Training Quantization of PyTorch models with NNCF

This tutorial is also available as a Jupyter notebook that can be cloned directly from GitHub. See the installation guide for instructions to run this tutorial locally on Windows, Linux or macOS.


The goal of this tutorial is to demonstrate how to use the NNCF (Neural Network Compression Framework) 8-bit quantization in post-training mode (without the fine-tuning pipeline) to optimize a PyTorch model for the high-speed inference via OpenVINO™ Toolkit. The optimization process contains the following steps:

  1. Evaluate the original model.

  2. Transform the original model to a quantized one.

  3. Export optimized and original models to OpenVINO IR.

  4. Compare performance of the obtained FP32 and INT8 models.

This tutorial uses a ResNet-50 model, pre-trained on Tiny ImageNet, which contains 100000 images of 200 classes (500 for each class) downsized to 64×64 colored images. The tutorial will demonstrate that only a tiny part of the dataset is needed for the post-training quantization, not demanding the fine-tuning of the model.

NOTE: This notebook requires that a C++ compiler is accessible on the default binary search path of the OS you are running the notebook.


# On Windows, this script adds the directory that contains cl.exe to the PATH to enable PyTorch to find the
# required C++ tools. This code assumes that Visual Studio 2019 is installed in the default
# directory. If you have a different C++ compiler, add the correct path to os.environ["PATH"]
# directly.

# Adding the path to os.environ["LIB"] is not always required - it depends on the system configuration.

import sys

if sys.platform == "win32":
    import distutils.command.build_ext
    import os
    from pathlib import Path

    VS_INSTALL_DIR = r"C:/Program Files (x86)/Microsoft Visual Studio"
    cl_paths = sorted(list(Path(VS_INSTALL_DIR).glob("**/Hostx86/x64/cl.exe")))
    if len(cl_paths) == 0:
        raise ValueError(
            "Cannot find Visual Studio. This notebook requires C++. If you installed "
            "a C++ compiler, please add the directory that contains cl.exe to "
        # If multiple versions of MSVC are installed, get the most recent one.
        cl_path = cl_paths[-1]
        vs_dir = str(cl_path.parent)
        os.environ["PATH"] += f"{os.pathsep}{vs_dir}"
        # The code for finding the library dirs is from
        d = distutils.core.Distribution()
        b = distutils.command.build_ext.build_ext(d)
        os.environ["LIB"] = os.pathsep.join(b.library_dirs)
        print(f"Added {vs_dir} to PATH")


import os
import time
import zipfile
from pathlib import Path
from typing import List, Tuple

import nncf
from openvino.runtime import Core, serialize
from import mo

import torch
from torchvision.datasets import ImageFolder
from torchvision.models import resnet50
import torchvision.transforms as transforms

from notebook_utils import download_file
2023-05-17 22:55:25.724897: I tensorflow/core/util/] 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.
2023-05-17 22:55:25.759432: I tensorflow/core/platform/] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
2023-05-17 22:55:26.311117: W tensorflow/compiler/tf2tensorrt/utils/] TF-TRT Warning: Could not find TensorRT
/opt/home/k8sworker/cibuilds/ov-notebook/OVNotebookOps-408/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/openvino/offline_transformations/ FutureWarning: The module is private and following namespace offline_transformations will be removed in the future, use openvino.runtime.passes instead!
INFO:nncf:NNCF initialized successfully. Supported frameworks detected: torch, tensorflow, onnx, openvino


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using {device} device")

MODEL_DIR = Path("model")
OUTPUT_DIR = Path("output")
BASE_MODEL_NAME = "resnet50"
IMAGE_SIZE = [64, 64]


# Paths where PyTorch and OpenVINO IR models will be stored.
fp32_checkpoint_filename = Path(BASE_MODEL_NAME + "_fp32").with_suffix(".pth")
fp32_ir_path = OUTPUT_DIR / Path(BASE_MODEL_NAME + "_fp32").with_suffix(".xml")
int8_ir_path = OUTPUT_DIR / Path(BASE_MODEL_NAME + "_int8").with_suffix(".xml")

fp32_pth_url = ""
download_file(fp32_pth_url, directory=MODEL_DIR, filename=fp32_checkpoint_filename)
Using cpu device
model/resnet50_fp32.pth:   0%|          | 0.00/91.5M [00:00<?, ?B/s]

Download and Prepare Tiny ImageNet dataset

  • 100k images of shape 3x64x64,

  • 200 different classes: snake, spider, cat, truck, grasshopper, gull, etc.

def download_tiny_imagenet_200(
    output_dir: Path,
    url: str = "",
    tarname: str = "",
    archive_path = output_dir / tarname
    download_file(url, directory=output_dir, filename=tarname)
    zip_ref = zipfile.ZipFile(archive_path, "r")
    print(f"Successfully downloaded and extracted dataset to: {output_dir}")

def create_validation_dir(dataset_dir: Path):
    VALID_DIR = dataset_dir / "val"
    val_img_dir = VALID_DIR / "images"

    fp = open(VALID_DIR / "val_annotations.txt", "r")
    data = fp.readlines()

    val_img_dict = {}
    for line in data:
        words = line.split("\t")
        val_img_dict[words[0]] = words[1]

    for img, folder in val_img_dict.items():
        newpath = val_img_dir / folder
        if not newpath.exists():
        if (val_img_dir / img).exists():
            os.rename(val_img_dir / img, newpath / img)

DATASET_DIR = OUTPUT_DIR / "tiny-imagenet-200"
if not DATASET_DIR.exists():
output/   0%|          | 0.00/237M [00:00<?, ?B/s]
Successfully downloaded and extracted dataset to: output

Helpers classes and functions

The code below will help to count accuracy and visualize validation process.

class AverageMeter(object):
    """Computes and stores the average and current value"""

    def __init__(self, name: str, fmt: str = ":f"): = name
        self.fmt = fmt
        self.val = 0
        self.avg = 0
        self.sum = 0
        self.count = 0

    def update(self, val: float, n: int = 1):
        self.val = val
        self.sum += val * n
        self.count += n
        self.avg = self.sum / self.count

    def __str__(self):
        fmtstr = "{name} {val" + self.fmt + "} ({avg" + self.fmt + "})"
        return fmtstr.format(**self.__dict__)

class ProgressMeter(object):
    """Displays the progress of validation process"""

    def __init__(self, num_batches: int, meters: List[AverageMeter], prefix: str = ""):
        self.batch_fmtstr = self._get_batch_fmtstr(num_batches)
        self.meters = meters
        self.prefix = prefix

    def display(self, batch: int):
        entries = [self.prefix + self.batch_fmtstr.format(batch)]
        entries += [str(meter) for meter in self.meters]

    def _get_batch_fmtstr(self, num_batches: int):
        num_digits = len(str(num_batches // 1))
        fmt = "{:" + str(num_digits) + "d}"
        return "[" + fmt + "/" + fmt.format(num_batches) + "]"

def accuracy(output: torch.Tensor, target: torch.Tensor, topk: Tuple[int] = (1,)):
    """Computes the accuracy over the k top predictions for the specified values of k"""
    with torch.no_grad():
        maxk = max(topk)
        batch_size = target.size(0)

        _, pred = output.topk(maxk, 1, True, True)
        pred = pred.t()
        correct = pred.eq(target.view(1, -1).expand_as(pred))

        res = []
        for k in topk:
            correct_k = correct[:k].reshape(-1).float().sum(0, keepdim=True)
            res.append(correct_k.mul_(100.0 / batch_size))
        return res

Validation function

def validate(val_loader:, model: torch.nn.Module):
    """Compute the metrics using data from val_loader for the model"""
    batch_time = AverageMeter("Time", ":3.3f")
    top1 = AverageMeter("Acc@1", ":2.2f")
    top5 = AverageMeter("Acc@5", ":2.2f")
    progress = ProgressMeter(len(val_loader), [batch_time, top1, top5], prefix="Test: ")

    # Switch to evaluate mode.

    with torch.no_grad():
        end = time.time()
        for i, (images, target) in enumerate(val_loader):
            images =
            target =

            # Compute the output.
            output = model(images)

            # Measure accuracy and record loss.
            acc1, acc5 = accuracy(output, target, topk=(1, 5))
            top1.update(acc1[0], images.size(0))
            top5.update(acc5[0], images.size(0))

            # Measure elapsed time.
            batch_time.update(time.time() - end)
            end = time.time()

            print_frequency = 10
            if i % print_frequency == 0:

            " * Acc@1 {top1.avg:.3f} Acc@5 {top5.avg:.3f}".format(top1=top1, top5=top5)
    return top1.avg

Create and load original uncompressed model

ResNet-50 from the torchivision repository is pre-trained on ImageNet with more prediction classes than Tiny ImageNet, so the model is adjusted by swapping the last FC layer to one with fewer output values.

def create_model(model_path: Path):
    """Creates the ResNet-50 model and loads the pretrained weights"""
    model = resnet50()
    # Update the last FC layer for Tiny ImageNet number of classes.
    NUM_CLASSES = 200
    model.fc = torch.nn.Linear(in_features=2048, out_features=NUM_CLASSES, bias=True)
    if model_path.exists():
        checkpoint = torch.load(str(model_path), map_location="cpu")
        model.load_state_dict(checkpoint["state_dict"], strict=True)
        raise RuntimeError("There is no checkpoint to load")
    return model

model = create_model(MODEL_DIR / fp32_checkpoint_filename)

Create train and validation dataloaders

def create_dataloaders(batch_size: int = 128):
    """Creates train dataloader that is used for quantization initialization and validation dataloader for computing the model accruacy"""
    train_dir = DATASET_DIR / "train"
    val_dir = DATASET_DIR / "val" / "images"
    normalize = transforms.Normalize(
        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
    train_dataset = ImageFolder(
    val_dataset = ImageFolder(
            [transforms.Resize(IMAGE_SIZE), transforms.ToTensor(), normalize]

    train_loader =

    val_loader =
    return train_loader, val_loader

train_loader, val_loader = create_dataloaders()

Model quantization and benchmarking

With the validation pipeline, model files, and data-loading procedures for model calibration now prepared, it’s time to proceed with the actual post-training quantization using NNCF.

I. Evaluate the loaded model

acc1 = validate(val_loader, model)
print(f"Test accuracy of FP32 model: {acc1:.3f}")
Test: [ 0/79]   Time 0.269 (0.269)  Acc@1 81.25 (81.25) Acc@5 92.19 (92.19)
Test: [10/79]   Time 0.246 (0.244)  Acc@1 56.25 (66.97) Acc@5 86.72 (87.50)
Test: [20/79]   Time 0.244 (0.245)  Acc@1 67.97 (64.29) Acc@5 85.16 (87.35)
Test: [30/79]   Time 0.243 (0.245)  Acc@1 53.12 (62.37) Acc@5 77.34 (85.33)
Test: [40/79]   Time 0.242 (0.246)  Acc@1 67.19 (60.86) Acc@5 90.62 (84.51)
Test: [50/79]   Time 0.243 (0.246)  Acc@1 60.16 (60.80) Acc@5 88.28 (84.42)
Test: [60/79]   Time 0.242 (0.245)  Acc@1 66.41 (60.46) Acc@5 86.72 (83.79)
Test: [70/79]   Time 0.244 (0.245)  Acc@1 52.34 (60.21) Acc@5 80.47 (83.33)
 * Acc@1 60.740 Acc@5 83.960
Test accuracy of FP32 model: 60.740

II. Create and initialize quantization

NNCF enables post-training quantization by adding the quantization layers into the model graph and then using a subset of the training dataset to initialize the parameters of these additional quantization layers. The framework is designed so that modifications to your original training code are minor. Quantization is the simplest scenario and requires a few modifications. For more information about NNCF Post Training Quantization (PTQ) API, refer to the Basic Quantization Flow Guide.

  1. Create a transformation function that accepts a sample from the dataset and returns data suitable for model inference. This enables the creation of an instance of the nncf.Dataset class, which represents the calibration dataset (based on the training dataset) necessary for post-training quantization.

def transform_fn(data_item):
    images, _ = data_item
    return images

calibration_dataset = nncf.Dataset(train_loader, transform_fn)
  1. Create a quantized model from the pre-trained FP32 model and the calibration dataset.

quantized_model = nncf.quantize(model, calibration_dataset)
No CUDA runtime is found, using CUDA_HOME='/usr/local/cuda'
INFO:nncf:Collecting tensor statistics |█████           | 1 / 3
INFO:nncf:Collecting tensor statistics |██████████      | 2 / 3
INFO:nncf:Collecting tensor statistics |████████████████| 3 / 3
INFO:nncf:Compiling and loading torch extension: quantized_functions_cpu...
INFO:nncf:Finished loading torch extension: quantized_functions_cpu
  1. 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 now.

acc1 = validate(val_loader, quantized_model)
print(f"Accuracy of initialized INT8 model: {acc1:.3f}")
Test: [ 0/79]   Time 0.441 (0.441)  Acc@1 79.69 (79.69) Acc@5 92.19 (92.19)
Test: [10/79]   Time 0.422 (0.435)  Acc@1 54.69 (66.34) Acc@5 86.72 (87.29)
Test: [20/79]   Time 0.426 (0.432)  Acc@1 67.19 (63.84) Acc@5 84.38 (87.13)
Test: [30/79]   Time 0.419 (0.429)  Acc@1 52.34 (62.15) Acc@5 75.00 (85.18)
Test: [40/79]   Time 0.424 (0.428)  Acc@1 67.97 (60.65) Acc@5 89.84 (84.20)
Test: [50/79]   Time 0.426 (0.428)  Acc@1 60.94 (60.69) Acc@5 86.72 (83.95)
Test: [60/79]   Time 0.425 (0.427)  Acc@1 64.84 (60.32) Acc@5 85.94 (83.43)
Test: [70/79]   Time 0.424 (0.427)  Acc@1 53.12 (60.08) Acc@5 79.69 (83.07)
 * Acc@1 60.600 Acc@5 83.670
Accuracy of initialized INT8 model: 60.600

It should be noted that the inference time for the quantized PyTorch model is longer than that of the original model, as fake quantizers are added to the model by NNCF. However, the model’s performance will significantly improve when it is in the OpenVINO Intermediate Representation (IR) format.

III. Convert the models to OpenVINO Intermediate Representation (OpenVINO IR)

Use Model Optimizer Python API to convert the Pytorch models to OpenVINO IR. Pass 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. The models will be saved to the ‘OUTPUT’ directory for latter benchmarking.

For more information about Model Optimizer, refer to the Model Optimizer Developer Guide.

Executing the following command may take a while.

model_ir = mo.convert_model(
    input_shape=[1, 3, *IMAGE_SIZE],
    mean_values=[123.675, 116.28 , 103.53],
    scale_values=[58.395, 57.12 , 57.375],

serialize(model_ir, str(fp32_ir_path))
WARNING:tensorflow:Please fix your imports. Module has been moved to tensorflow.python.trackable.base. The old module will be deleted in version 2.11.
[ WARNING ] Please fix your imports. Module has been moved to tensorflow.python.trackable.base. The old module will be deleted in version 2.11.
[ WARNING ]  Please fix your imports. Module %s has been moved to %s. The old module will be deleted in version %s.
quantized_model_ir = mo.convert_model(
    input_shape=[1, 3, *IMAGE_SIZE],
    mean_values=[123.675, 116.28 , 103.53],
    scale_values=[58.395, 57.12 , 57.375],

serialize(quantized_model_ir, str(int8_ir_path))

IV. Compare perfomance of INT8 model and FP32 model in OpenVINO

Finally, measure the inference performance of the FP32 and INT8 models, using Benchmark Tool - an 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.

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.

def parse_benchmark_output(benchmark_output: str):
    """Prints the output from benchmark_app in human-readable format"""
    parsed_output = [line for line in benchmark_output if 'FPS' in line]
    print(*parsed_output, sep='\n')

print('Benchmark FP32 model (OpenVINO IR)')
benchmark_output = ! benchmark_app -m "$fp32_ir_path" -d CPU -api async -t 15

print('Benchmark INT8 model (OpenVINO IR)')
benchmark_output = ! benchmark_app -m "$int8_ir_path" -d CPU -api async -t 15
Benchmark FP32 model (OpenVINO IR)
[ INFO ] Throughput:   1273.98 FPS
Benchmark INT8 model (OpenVINO IR)
[ INFO ] Throughput:   4191.76 FPS

Show CPU Information for reference:

ie = Core()
ie.get_property("CPU", "FULL_DEVICE_NAME")
'Intel(R) Core(TM) i9-10920X CPU @ 3.50GHz'