Quantizing Model with Accuracy Control


This document assumes that you already tried Default Quantization for the same model. In case when it introduces a significant accuracy degradation, the Accuracy-aware Quantization algorithm can be used to remain accuracy within the pre-defined range. This may cause a degradation of performance in comparison to Default Quantization algorithm because some layers can be reverted back to the original precision.


In case of GNA target_device, the Accuracy-aware Quantization algorithm behavior is different. It is searching for the best configuration selecting between INT8 and INT16 precisions for weights of each layer. The algorithm works for the performance preset only. For the accuracy preset, this algorithm is not helpful since the whole model is already in INT16 precision.

A script for Accuracy-aware Quantization includes four steps:

  1. Prepare data and dataset interface

  2. Define accuracy metric

  3. Select quantization parameters

  4. Define and run quantization process

Prepare data and dataset interface

This step is the same as in the case of Default Quantization. The only difference is that __getitem__() method should return (data, annotation) or (data, annotation, metadata) where annotation is required and its format should correspond to the expectations of the Metric class. metadata is an optional field that can be used to store additional information required for post-processing.

Define accuracy metric

To control accuracy during the optimization a openvino.tools.pot.Metric interface should be implemented. Each implementation should override the following properties:

  • value - returns the accuracy metric value for the last model output in a format of Dict[str, numpy.array].

  • avg_value - returns the average accuracy metric over collected model results in a format of Dict[str, numpy.array].

  • higher_better should return True if a higher value of the metric corresponds to better performance, otherwise, returns False. Default implementation returns True.

and methods:

  • update(output, annotation) - calculates and updates the accuracy metric value using the last model output and annotation. The model output and annotation should be passed in this method. It should also contain the model-specific post-processing in case the model returns the raw output.

  • reset() - resets collected accuracy metric.

  • get_attributes() - returns a dictionary of metric attributes:

    {metric_name: {attribute_name: value}}

    Required attributes:

    • direction - (higher-better or higher-worse) a string parameter defining whether metric value should be increased in accuracy-aware algorithms.

    • type - a string representation of metric type. For example, ‘accuracy’ or ‘mean_iou’.

Below is an example of the accuracy top-1 metric implementation with POT API:

from openvino.tools.pot import Metric

class Accuracy(Metric):

    # Required methods
    def __init__(self, top_k=1):
        self._top_k = top_k
        self._name = 'accuracy@top{}'.format(self._top_k)
        self._matches = [] # container of the results

    def value(self):
        """ Returns accuracy metric value for all model outputs. """
        return {self._name: self._matches[-1]}

    def avg_value(self):
        """ Returns accuracy metric value for all model outputs. """
        return {self._name: np.ravel(self._matches).mean()}

    def update(self, output, target):
        """ Updates prediction matches.
        :param output: model output
        :param target: annotations
        if len(output) > 1:
            raise Exception('The accuracy metric cannot be calculated '
                            'for a model with multiple outputs')
        if isinstance(target, dict):
            target = list(target.values())
        predictions = np.argsort(output[0], axis=1)[:, -self._top_k:]
        match = [float(t in predictions[i]) for i, t in enumerate(target)]


    def reset(self):
        """ Resets collected matches """
        self._matches = []

    def get_attributes(self):
        Returns a dictionary of metric attributes {metric_name: {attribute_name: value}}.
        Required attributes: 'direction': 'higher-better' or 'higher-worse'
                             'type': metric type
        return {self._name: {'direction': 'higher-better',
                             'type': 'accuracy'}}

An instance of the Metric implementation should be passed to IEEngine object responsible for model inference.

metric = Accuracy()
engine = IEEngine(config=engine_config, data_loader=data_loader, metric=metric)

Select quantization parameters

Accuracy-aware Quantization uses the Default Quantization algorithm at the initialization step so that all its parameters are also valid and can be specified. Here, we describe only Accuracy-aware Quantization required parameters:

  • "maximal_drop" - maximum accuracy drop which has to be achieved after the quantization. Default value is 0.01 (1%).

Run quantization

The code example below shows a basic quantization workflow with accuracy control. UserDataLoader() is a placeholder for the implementation of DataLoader.

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

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

# Engine config
engine_config = Dict({"device": "CPU"})

algorithms = [
        "name": "AccuracyAwareQuantization",
        "params": {
            "target_device": "ANY",
            "stat_subset_size": 300,
            'maximal_drop': 0.02

# Step 1: implement and create user's data loader
data_loader = UserDataLoader()

# Step 2: implement and create user's data loader
metric = Accuracy()

# Step 3: load model
model = load_model(model_config=model_config)

# Step 4: Initialize the engine for metric calculation and statistics collection.
engine = IEEngine(config=engine_config, data_loader=data_loader, metric=metric)

# Step 5: Create a pipeline of compression algorithms and run it.
pipeline = create_pipeline(algorithms, engine)
compressed_model = pipeline.run(model=model)

# Step 6 (Optional): Compress model weights to quantized precision
#                    in order to reduce the size of the final .bin file.

# Step 7: 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(

# Step 8 (Optional): Evaluate the compressed model. Print the results.
metric_results = pipeline.evaluate(compressed_model)

It is worth noting that now the evaluate method that can compute accuracy on demand is also available in the Pipeline object.

In case when Accuracy-aware Quantization does not allow achieving the desired accuracy-performance trade-off, it is recommended to try Quantization-aware Training from NNCF.