[Deprecated] Quantizing Models


Post-training Optimization Tool is deprecated since OpenVINO 2023.0. Neural Network Compression Framework (NNCF) is recommended for the post-training quantization instead.

This guide describes how to apply model quantization with the Default Quantization method without accuracy control, using an unannotated dataset. To use this method, create a Python script using an API of Post-Training Optimization Tool (POT) and implement data preparation logic and quantization pipeline. If you are not familiar with Python, try command-line interface of POT which is designed to quantize models from OpenVINO Model Zoo. The figure below shows the common workflow of the quantization script implemented with POT API.


The script should include three basic steps:

  1. Prepare data and dataset interface.

  2. Select quantization parameters.

  3. Define and run the quantization process.

Prepare data and dataset interface

In most cases, it is required to implement only the openvino.tools.pot.DataLoader interface, which allows acquiring data from a dataset and applying model-specific pre-processing providing access by index. Any implementation should override the following methods:

  • The __len__(), returns the size of the dataset.

  • The __getitem__(), provides access to the data by index in the range of 0 to len(self). It can also encapsulate the logic of model-specific pre-processing. This method should return data in the (data, annotation) format, in which:

    • The data is the input that is passed to the model at inference so that it should be properly preprocessed. It can be either the numpy.array object or a dictionary, where the key is the name of the model input and the value is numpy.array which corresponds to this input.

    • The annotation is not used by the Default Quantization method. Therefore, this object can be None in this case.

Framework data loading classes can be wrapped by the openvino.tools.pot.DataLoader interface which is usually straightforward. For example, the torch.utils.data.Dataset has a similar interface as the openvino.tools.pot.DataLoader, so that its TorchVision implementations can be easily wrapped by POT API.


Model-specific preprocessing (for example, mean/scale normalization), can be embedded into the model at the conversion step, using Model Optimizer. This should be considered during the implementation of the DataLoader interface to avoid “double” normalization, which can lead to the loss of accuracy after optimization.

The example code below defines the DataLoader for three popular use cases: images, text, and audio.

import os

import numpy as np
import cv2 as cv

from openvino.tools.pot import DataLoader

class ImageLoader(DataLoader):
    """ Loads images from a folder """
    def __init__(self, dataset_path):
        # Use OpenCV to gather image files
        # Collect names of image files
        self._files = []
        all_files_in_dir = os.listdir(dataset_path)
        for name in all_files_in_dir:
            file = os.path.join(dataset_path, name)
            if cv.haveImageReader(file):

        # Define shape of the model
        self._shape = (224,224)

    def __len__(self):
        """ Returns the length of the dataset """
        return len(self._files)

    def __getitem__(self, index):
        """ Returns image data by index in the NCHW layout
        Note: model-specific preprocessing is omitted, consider adding it here
        if index >= len(self):
            raise IndexError("Index out of dataset size")

        image = cv.imread(self._files[index]) # read image with OpenCV
        image = cv.resize(image, self._shape) # resize to a target input size
        image = np.expand_dims(image, 0)  # add batch dimension
        image = image.transpose(0, 3, 1, 2)  # convert to NCHW layout
        return image, None   # annotation is set to None
import os
from pathlib import Path

from datasets import load_dataset      #pip install datasets
from transformers import AutoTokenizer #pip install transformers

from openvino.tools.pot import DataLoader

class TextLoader(DataLoader):
    """ Loads content of .txt files from a folder """
    def __init__(self, dataset_path):
        # HuggingFace dataset API is used to process text files
        # Collect names of text files
        extension = ".txt"
        files = sorted(str(p.stem) for p in
            Path(dataset_path).glob("*" + extension))
        files = [os.path.join(dataset_path, file + extension) for file in files]
        self._dataset = load_dataset('text', data_files=files)
        # replace with your tokenizer
        self._tokenizer = AutoTokenizer.from_pretrained('bert-base-cased')
        self._dataset = self._dataset.map(self._encode, batched=False)
        # replace with names of model inputs
                    columns=['input_ids', 'token_type_ids', 'attention_mask'])

    def _encode(self, examples):
        """ Tokenization of the input text """
        return self._tokenizer(examples['text'], truncation=True, padding='max_length')

    def __len__(self):
        """ Returns the length of the dataset """
        return len(self._dataset['train'])

    def __getitem__(self, index):
        """ Returns data by index as a (dict[str, np.array], None) """
        if index >= len(self):
            raise IndexError("Index out of dataset size")

        data = self._dataset['train'][index]
        return {'input_ids': data['input_ids'],
                'token_type_ids': data['token_type_ids'],
                'attention_mask': data['attention_mask']}, None # annotation is set to None
import os
from pathlib import Path

import torchaudio # pip install torch torchaudio

from openvino.tools.pot import DataLoader

class AudioLoader(DataLoader):
    """ Loads content of .wav files from a folder """
    def __init__(self, dataset_path):
        # Collect names of wav files
        self._extension = ".wav"
        self._dataset_path = dataset_path
        self._files = sorted(str(p.stem) for p in
            Path(self._dataset_path).glob("*" + self._extension))

    def __len__(self):
        """ Returns the length of the dataset """
        return len(self._files)

    def __getitem__(self, index):
        """ Returns wav data by index
        Note: model-specific preprocessing is omitted, consider adding it here
        if index >= len(self):
            raise IndexError("Index out of dataset size")

        file_name = self._files[index] + self._extension
        file_path = os.path.join(self._dataset_path, file_name)
        waveform, _ = torchaudio.load(file_path) # use a helper from torchaudio to load data
        return waveform.numpy(), None   # annotation is set to None

Select quantization parameters

Default Quantization algorithm has mandatory and optional parameters which are defined as a dictionary:

    "name": "DefaultQuantization",
    "params": {
        "target_device": "ANY",
        "stat_subset_size": 300,
        "stat_batch_size": 1
  • "target_device" - the following options are available:

    • "ANY" (or "CPU") - default option to quantize models for CPU, GPU, or NPU

    • "CPU_SPR" - to quantize models for CPU SPR (4th Generation Intel® Xeon® Scalable processor family)

    • "GNA", "GNA3", "GNA3.5" - to quantize models for GNA devices respectively.

  • "stat_subset_size" - size of the data subset to calculate activations statistics used for quantization. The whole dataset is used if no parameter is specified. It is recommended to use not less than 300 samples.

  • "stat_batch_size" - size of the batch to calculate activations statistics used for quantization. 1 if no parameter is specified.

For full specification, see the Default Quantization method.

Run quantization

POT API provides methods to load and save model objects from OpenVINO Intermediate Representation: the load_model and save_model. It also has a concept of the Pipeline that sequentially applies specified optimization methods to the model. The create_pipeline method is used to instantiate a Pipeline object. An example code below shows a basic quantization workflow:

from openvino.tools.pot import IEEngine
from openvino.tools.pot import load_model, save_model
from openvino.tools.pot import compress_model_weights
from openvino.tools.pot import create_pipeline

# Model config specifies the name of the model and paths to .xml and .bin files of the model.
model_config =
    "model_name": "model",
    "model": path_to_xml,
    "weights": path_to_bin,

# Engine config.
engine_config = {"device": "CPU"}

algorithms = [
        "name": "DefaultQuantization",
        "params": {
            "target_device": "ANY",
            "stat_subset_size": 300,
            "stat_batch_size": 1

# Step 1: Implement and create a user data loader.
data_loader = ImageLoader("<path_to_images>")

# Step 2: Load a model.
model = load_model(model_config=model_config)

# Step 3: Initialize the engine for metric calculation and statistics collection.
engine = IEEngine(config=engine_config, data_loader=data_loader)

# Step 4: Create a pipeline of compression algorithms and run it.
pipeline = create_pipeline(algorithms, engine)
compressed_model = pipeline.run(model=model)

# Step 5 (Optional): Compress model weights to quantized precision
#                     to reduce the size of the final .bin file.

# Step 6: Save the compressed model to the desired path.
# Set save_path to the directory where the model should be saved.
compressed_model_paths = save_model(

The output of the script is the quantized model that can be used for inference in the same way as the original full-precision model.

If high degradation of accuracy occurs after applying the Default Quantization method, it is recommended to follow the tips from Quantization Best Practices article or use Accuracy-aware Quantization method.

Quantizing cascaded models

When the optimized model is a cascaded one (consists of several submodels, for example, MT-CNN), it will be necessary to implement a complex inference pipeline that can properly handle different submodels and data flow between them. POT API provides the Engine interface for this purpose, which allows customization of the inference logic. However, it is recommended to inherit from IEEngine helper class that already contains all the logic required to do the inference based on OpenVINO Python API. For more details, see the following example.