Neural Network Compression Framework (NNCF)

Key Features • Installation • Documentation • Usage • Tutorials and Samples • Third-party integration • Model Zoo

Neural Network Compression Framework (NNCF) provides a suite of post-training and training-time algorithms for optimizing inference of neural networks in OpenVINO™ with a minimal accuracy drop.

NNCF is designed to work with models from PyTorch, TorchFX, TensorFlow, ONNX and OpenVINO™.

NNCF provides samples that demonstrate the usage of compression algorithms for different use cases and models. See compression results achievable with the NNCF-powered samples on the NNCF Model Zoo page.

The framework is organized as a Python* package that can be built and used in a standalone mode. The framework architecture is unified to make it easy to add different compression algorithms for both PyTorch and TensorFlow deep learning frameworks.

Key Features

Post-Training Compression Algorithms

Compression algorithm	OpenVINO	PyTorch	TorchFX	TensorFlow	ONNX
Post-Training Quantization	Supported	Supported	Experimental	Supported	Supported
Weights Compression	Supported	Supported	Experimental	Not supported	Not supported
Activation Sparsity	Not supported	Experimental	Not supported	Not supported	Not supported

Training-Time Compression Algorithms

Compression algorithm	PyTorch	TensorFlow
Quantization Aware Training	Supported	Supported
Mixed-Precision Quantization	Supported	Not supported
Sparsity	Supported	Supported
Filter pruning	Supported	Supported
Movement pruning	Experimental	Not supported

• Automatic, configurable model graph transformation to obtain the compressed model.

  NOTE: Limited support for TensorFlow models. Only models created using Sequential or Keras Functional API are supported.

• Common interface for compression methods.
• GPU-accelerated layers for faster compressed model fine-tuning.
• Distributed training support.
• Git patch for prominent third-party repository (huggingface-transformers) demonstrating the process of integrating NNCF into custom training pipelines.
• Seamless combination of pruning, sparsity, and quantization algorithms. Please refer to optimum-intel for examples of joint (movement) pruning, quantization, and distillation (JPQD), end-to-end from NNCF optimization to compressed OpenVINO IR.
• Exporting PyTorch compressed models to ONNX* checkpoints and TensorFlow compressed models to SavedModel or Frozen Graph format, ready to use with OpenVINO™ toolkit.
• Support for Accuracy-Aware model training pipelines via the Adaptive Compression Level Training and Early Exit Training.

Documentation

This documentation covers detailed information about NNCF algorithms and functions needed for the contribution to NNCF.

The latest user documentation for NNCF is available here.

NNCF API documentation can be found here.

Usage

Post-Training Quantization

The NNCF PTQ is the simplest way to apply 8-bit quantization. To run the algorithm you only need your model and a small (~300 samples) calibration dataset.

OpenVINO is the preferred backend to run PTQ with, while PyTorch, TensorFlow, and ONNX are also supported.

OpenVINO

import nncf
import openvino.runtime as ov
import torch
from torchvision import datasets, transforms

# Instantiate your uncompressed model
model = ov.Core().read_model("/model_path")

# Provide validation part of the dataset to collect statistics needed for the compression algorithm
val_dataset = datasets.ImageFolder("/path", transform=transforms.Compose([transforms.ToTensor()]))
dataset_loader = torch.utils.data.DataLoader(val_dataset, batch_size=1)

# Step 1: Initialize transformation function
def transform_fn(data_item):
    images, _ = data_item
    return images

# Step 2: Initialize NNCF Dataset
calibration_dataset = nncf.Dataset(dataset_loader, transform_fn)
# Step 3: Run the quantization pipeline
quantized_model = nncf.quantize(model, calibration_dataset)

PyTorch

import nncf
import torch
from torchvision import datasets, models

# Instantiate your uncompressed model
model = models.mobilenet_v2()

# Provide validation part of the dataset to collect statistics needed for the compression algorithm
val_dataset = datasets.ImageFolder("/path", transform=transforms.Compose([transforms.ToTensor()]))
dataset_loader = torch.utils.data.DataLoader(val_dataset)

# Step 1: Initialize the transformation function
def transform_fn(data_item):
    images, _ = data_item
    return images

# Step 2: Initialize NNCF Dataset
calibration_dataset = nncf.Dataset(dataset_loader, transform_fn)
# Step 3: Run the quantization pipeline
quantized_model = nncf.quantize(model, calibration_dataset)

NOTE If the Post-Training Quantization algorithm does not meet quality requirements you can fine-tune the quantized pytorch model. You can find an example of the Quantization-Aware training pipeline for a pytorch model here.

TorchFX

import nncf
import torch.fx
from torchvision import datasets, models
from nncf.torch import disable_patching

# Instantiate your uncompressed model
model = models.mobilenet_v2()

# Provide validation part of the dataset to collect statistics needed for the compression algorithm
val_dataset = datasets.ImageFolder("/path", transform=transforms.Compose([transforms.ToTensor()]))
dataset_loader = torch.utils.data.DataLoader(val_dataset)

# Step 1: Initialize the transformation function
def transform_fn(data_item):
    images, _ = data_item
    return images

# Step 2: Initialize NNCF Dataset
calibration_dataset = nncf.Dataset(dataset_loader, transform_fn)

# Step 3: Export model to TorchFX
input_shape = (1, 3, 224, 224)
with nncf.torch.disable_patching():
    fx_model = torch.export.export_for_training(model, args=(ex_input,)).module()
    # or
    # fx_model = torch.export.export(model, args=(ex_input,)).module()

    # Step 4: Run the quantization pipeline
    quantized_fx_model = nncf.quantize(fx_model, calibration_dataset)

TensorFlow

import nncf
import tensorflow as tf
import tensorflow_datasets as tfds

# Instantiate your uncompressed model
model = tf.keras.applications.MobileNetV2()

# Provide validation part of the dataset to collect statistics needed for the compression algorithm
val_dataset = tfds.load("/path", split="validation",
                        shuffle_files=False, as_supervised=True)

# Step 1: Initialize transformation function
def transform_fn(data_item):
    images, _ = data_item
    return images

# Step 2: Initialize NNCF Dataset
calibration_dataset = nncf.Dataset(val_dataset, transform_fn)
# Step 3: Run the quantization pipeline
quantized_model = nncf.quantize(model, calibration_dataset)

ONNX

import onnx
import nncf
import torch
from torchvision import datasets

# Instantiate your uncompressed model
onnx_model = onnx.load_model("/model_path")

# Provide validation part of the dataset to collect statistics needed for the compression algorithm
val_dataset = datasets.ImageFolder("/path", transform=transforms.Compose([transforms.ToTensor()]))
dataset_loader = torch.utils.data.DataLoader(val_dataset, batch_size=1)

# Step 1: Initialize transformation function
input_name = onnx_model.graph.input[0].name
def transform_fn(data_item):
    images, _ = data_item
    return {input_name: images.numpy()}

# Step 2: Initialize NNCF Dataset
calibration_dataset = nncf.Dataset(dataset_loader, transform_fn)
# Step 3: Run the quantization pipeline
quantized_model = nncf.quantize(onnx_model, calibration_dataset)

Training-Time Quantization

Here is an example of Accuracy Aware Quantization pipeline where model weights and compression parameters may be fine-tuned to achieve a higher accuracy.

PyTorch

import nncf
import torch
from torchvision import datasets, models

# Instantiate your uncompressed model
model = models.mobilenet_v2()

# Provide validation part of the dataset to collect statistics needed for the compression algorithm
val_dataset = datasets.ImageFolder("/path", transform=transforms.Compose([transforms.ToTensor()]))
dataset_loader = torch.utils.data.DataLoader(val_dataset)

# Step 1: Initialize the transformation function
def transform_fn(data_item):
    images, _ = data_item
    return images

# Step 2: Initialize NNCF Dataset
calibration_dataset = nncf.Dataset(dataset_loader, transform_fn)
# Step 3: Run the quantization pipeline
quantized_model = nncf.quantize(model, calibration_dataset)

# Now use compressed_model as a usual torch.nn.Module
# to fine-tune compression parameters along with the model weights

# Save quantization modules and the quantized model parameters
checkpoint = {
    'state_dict': model.state_dict(),
    'nncf_config': model.nncf.get_config(),
    ... # the rest of the user-defined objects to save
}
torch.save(checkpoint, path_to_checkpoint)

# ...

# Load quantization modules and the quantized model parameters
resuming_checkpoint = torch.load(path_to_checkpoint)
nncf_config = resuming_checkpoint['nncf_config']
state_dict = resuming_checkpoint['state_dict']

quantized_model = nncf.torch.load_from_config(model, nncf_config, example_input)
model.load_state_dict(state_dict)
# ... the rest of the usual PyTorch-powered training pipeline

Training-Time Compression

Here is an example of Accuracy Aware RB Sparsification pipeline where model weights and compression parameters may be fine-tuned to achieve a higher accuracy.

PyTorch

import torch
import nncf.torch  # Important - must be imported before any other external package that depends on torch

from nncf import NNCFConfig
from nncf.torch import create_compressed_model, register_default_init_args

# Instantiate your uncompressed model
from torchvision.models.resnet import resnet50
model = resnet50()

# Load a configuration file to specify compression
nncf_config = NNCFConfig.from_json("resnet50_imagenet_rb_sparsity.json")

# Provide data loaders for compression algorithm initialization, if necessary
import torchvision.datasets as datasets
representative_dataset = datasets.ImageFolder("/path", transform=transforms.Compose([transforms.ToTensor()]))
init_loader = torch.utils.data.DataLoader(representative_dataset)
nncf_config = register_default_init_args(nncf_config, init_loader)

# Apply the specified compression algorithms to the model
compression_ctrl, compressed_model = create_compressed_model(model, nncf_config)

# Now use compressed_model as a usual torch.nn.Module
# to fine-tune compression parameters along with the model weights

# ... the rest of the usual PyTorch-powered training pipeline

# Export to ONNX or .pth when done fine-tuning
compression_ctrl.export_model("compressed_model.onnx")
torch.save(compressed_model.state_dict(), "compressed_model.pth")

NOTE (PyTorch): Due to the way NNCF works within the PyTorch backend, import nncf must be done before any other import of torch in your package or in third-party packages that your code utilizes. Otherwise, the compression may be applied incompletely.

Tensorflow

import tensorflow as tf

from nncf import NNCFConfig
from nncf.tensorflow import create_compressed_model, register_default_init_args

# Instantiate your uncompressed model
from tensorflow.keras.applications import ResNet50
model = ResNet50()

# Load a configuration file to specify compression
nncf_config = NNCFConfig.from_json("resnet50_imagenet_rb_sparsity.json")

# Provide dataset for compression algorithm initialization
representative_dataset = tf.data.Dataset.list_files("/path/*.jpeg")
nncf_config = register_default_init_args(nncf_config, representative_dataset, batch_size=1)

# Apply the specified compression algorithms to the model
compression_ctrl, compressed_model = create_compressed_model(model, nncf_config)

# Now use compressed_model as a usual Keras model
# to fine-tune compression parameters along with the model weights

# ... the rest of the usual TensorFlow-powered training pipeline

# Export to Frozen Graph, TensorFlow SavedModel or .h5  when done fine-tuning
compression_ctrl.export_model("compressed_model.pb", save_format="frozen_graph")

For a more detailed description of NNCF usage in your training code, see this tutorial.

Demos, Tutorials and Samples

For a quicker start with NNCF-powered compression, try sample notebooks and scripts presented below.

Jupyter* Notebook Tutorials and Demos

Ready-to-run Jupyter* notebook tutorials and demos are available to explain and display NNCF compression algorithms for optimizing models for inference with the OpenVINO Toolkit:

Notebook Tutorial Name	Compression Algorithm	Backend	Domain
BERT Quantization	Post-Training Quantization	OpenVINO	NLP
MONAI Segmentation Model Quantization	Post-Training Quantization	OpenVINO	Segmentation
PyTorch Model Quantization	Post-Training Quantization	PyTorch	Image Classification
Quantization with Accuracy Control	Post-Training Quantization with Accuracy Control	OpenVINO	Speech-to-Text, Object Detection
PyTorch Training-Time Compression	Training-Time Compression	PyTorch	Image Classification
TensorFlow Training-Time Compression	Training-Time Compression	Tensorflow	Image Classification
Joint Pruning, Quantization and Distillation for BERT	Joint Pruning, Quantization and Distillation	OpenVINO	NLP

A list of notebooks demonstrating OpenVINO conversion and inference together with NNCF compression for models from various domains:

Demo Model	Compression Algorithm	Backend	Domain
YOLOv8	Post-Training Quantization	OpenVINO	Object Detection, KeyPoint Detection, Instance Segmentation
EfficientSAM	Post-Training Quantization	OpenVINO	Image Segmentation
Segment Anything Model	Post-Training Quantization	OpenVINO	Image Segmentation
OneFormer	Post-Training Quantization	OpenVINO	Image Segmentation
InstructPix2Pix	Post-Training Quantization	OpenVINO	Image-to-Image
CLIP	Post-Training Quantization	OpenVINO	Image-to-Text
BLIP	Post-Training Quantization	OpenVINO	Image-to-Text
Latent Consistency Model	Post-Training Quantization	OpenVINO	Text-to-Image
ControlNet QR Code Monster	Post-Training Quantization	OpenVINO	Text-to-Image
SDXL-turbo	Post-Training Quantization	OpenVINO	Text-to-Image, Image-to-Image
Distil-Whisper	Post-Training Quantization	OpenVINO	Speech-to-Text
Whisper	Post-Training Quantization	OpenVINO	Speech-to-Text
MMS Speech Recognition	Post-Training Quantization	OpenVINO	Speech-to-Text
Grammar Error Correction	Post-Training Quantization	OpenVINO	NLP, Grammar Correction
LLM Instruction Following	Weight Compression	OpenVINO	NLP, Instruction Following
LLM Chat Bots	Weight Compression	OpenVINO	NLP, Chat Bot

Post-Training Quantization Examples

Compact scripts demonstrating quantization and corresponding inference speed boost:

Example Name	Compression Algorithm	Backend	Domain
OpenVINO MobileNetV2	Post-Training Quantization	OpenVINO	Image Classification
OpenVINO YOLOv8	Post-Training Quantization	OpenVINO	Object Detection
OpenVINO YOLOv8 QwAС	Post-Training Quantization with Accuracy Control	OpenVINO	Object Detection
OpenVINO Anomaly Classification	Post-Training Quantization with Accuracy Control	OpenVINO	Anomaly Classification
PyTorch MobileNetV2	Post-Training Quantization	PyTorch	Image Classification
PyTorch SSD	Post-Training Quantization	PyTorch	Object Detection
TorchFX Resnet18	Post-Training Quantization	TorchFX	Image Classification
TensorFlow MobileNetV2	Post-Training Quantization	TensorFlow	Image Classification
ONNX MobileNetV2	Post-Training Quantization	ONNX	Image Classification

Training-Time Compression Examples

Examples of full pipelines including compression, training, and inference for classification, detection, and segmentation tasks:

Example Name	Compression Algorithm	Backend	Domain
PyTorch Image Classification	Training-Time Compression	PyTorch	Image Classification
PyTorch Object Detection	Training-Time Compression	PyTorch	Object Detection
PyTorch Semantic Segmentation	Training-Time Compression	PyTorch	Semantic Segmentation
TensorFlow Image Classification	Training-Time Compression	TensorFlow	Image Classification
TensorFlow Object Detection	Training-Time Compression	TensorFlow	Object Detection
TensorFlow Instance Segmentation	Training-Time Compression	TensorFlow	Instance Segmentation

Third-party repository integration

NNCF may be easily integrated into training/evaluation pipelines of third-party repositories.

Used by

• OpenVINO Training Extensions

  NNCF is integrated into OpenVINO Training Extensions as a model optimization backend. You can train, optimize, and export new models based on available model templates as well as run the exported models with OpenVINO.

• HuggingFace Optimum Intel

  NNCF is used as a compression backend within the renowned transformers repository in HuggingFace Optimum Intel.

Installation Guide

For detailed installation instructions, refer to the Installation guide.

NNCF can be installed as a regular PyPI package via pip:

pip install nncf

NNCF is also available via conda:

conda install -c conda-forge nncf

System requirements of NNCF correspond to the used backend. System requirements for each backend and the matrix of corresponding versions can be found in installation.md.

NNCF Compressed NNCF Model Zoo

List of models and compression results for them can be found at our NNCF Model Zoo page.

Citing

@article{kozlov2020neural,
    title =   {Neural network compression framework for fast model inference},
    author =  {Kozlov, Alexander and Lazarevich, Ivan and Shamporov, Vasily and Lyalyushkin, Nikolay and Gorbachev, Yury},
    journal = {arXiv preprint arXiv:2002.08679},
    year =    {2020}
}

Contributing Guide

Refer to the CONTRIBUTING.md file for guidelines on contributions to the NNCF repository.

Useful links

• Documentation
• Example scripts (model objects available through links in respective README.md files):
  • PyTorch
  • TensorFlow
• FAQ
• Notebooks
• HuggingFace Optimum Intel
• OpenVINO Model Optimization Guide

Telemetry

NNCF as part of the OpenVINO™ toolkit collects anonymous usage data for the purpose of improving OpenVINO™ tools. You can opt-out at any time by running the following command in the Python environment where you have NNCF installed:

opt_in_out --opt_out

More information available on OpenVINO telemetry.