nncf 2.5.0

Neural Networks Compression Framework

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 neural networks inference optimization in OpenVINO™ with minimal accuracy drop.

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

NNCF provides samples that demonstrate the usage of compression algorithms for different use cases and models. Compression results achievable with the NNCF-powered samples can be found in a table at the end of this document.

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	TensorFlow	ONNX
Post-Training Quantization	Supported	Supported	Supported	Supported

Training-Time Compression Algorithms

Compression algorithm	PyTorch	TensorFlow
Quantization Aware Training	Supported	Supported
Mixed-Precision Quantization	Supported	Not supported
Binarization	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. The models created using Sequential or Keras Functional API are only 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, and PyTorch, TensorFlow and ONNX are also supported.

OpenVINO

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

# 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")
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")
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)

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")
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 Compression

Below 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 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_int8.json")

# Provide data loaders for compression algorithm initialization, if necessary
import torchvision.datasets as datasets
representative_dataset = datasets.ImageFolder("/path")
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_int8.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.

Model Compression Tutorials and Samples

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

Model Compression Tutorials

A collection of ready-to-run Jupyter* notebooks are available to demonstrate how to use NNCF compression algorithms to optimize models for inference with the OpenVINO Toolkit:

• Accelerate Inference of NLP models with Post-Training Qunatization API of NNCF
• Convert and Optimize YOLOv8 with OpenVINO
• Convert and Optimize YOLOv7 with OpenVINO
• NNCF Post-Training Optimization of Segment Anything Model
• Quantize a Segmentation Model and Show Live Inference
• Training to Deployment with TensorFlow and OpenVINO
• Migrate quantization from POT API to NNCF API
• Post-Training Quantization of Pytorch model with NNCF
• Optimizing PyTorch models with NNCF of OpenVINO by 8-bit quantization
• Optimizing TensorFlow models with NNCF of OpenVINO by 8-bit quantization
• Accelerate Inference of Sparse Transformer Models with OpenVINO and 4th Gen Intel Xeon Scalable Processors

Post-Training Quantization Samples

Compact scripts demonstrating quantization and corresponding inference speed boost:

• Post-Training Quantization of MobileNet v2 OpenVINO Model
• Post-Training Quantization of YOLOv8 OpenVINO Model
• Post-Training Quantization of Anomaly Classification OpenVINO model with control of accuracy metric
• Post-Training Quantization of YOLOv8 OpenVINO Model with control of accuracy metric
• Post-Training Quantization of MobileNet v2 PyTorch Model
• Post-Training Quantization of SSD PyTorch Model
• Post-Training Quantization of MobileNet v2 ONNX Model
• Post-Training Quantization of MobileNet v2 TensorFlow Model

Training-Time Compression Samples

These examples provide full pipelines including compression, training and inference for classification, object detection and segmentation tasks.

• PyTorch samples:
  • Image Classification sample
  • Object Detection sample
  • Semantic Segmentation sample
• TensorFlow samples:
  • Image Classification sample
  • Object Detection sample
  • Instance Segmentation sample

Third-party repository integration

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

Used by

• OpenVINO Training Extensions

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

• HuggingFace Optimum Intel

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

Git patches for third-party repository

See third_party_integration for examples of code modifications (Git patches and base commit IDs are provided) that are necessary to integrate NNCF into the following repositories:

• huggingface-transformers

Installation Guide

For detailed installation instructions please refer to the Installation page.

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

pip install nncf

If you want to install both NNCF and the supported PyTorch version in one line, you can do this by simply running:

pip install nncf[torch]

Other viable options besides [torch] are [tf], [onnx] and [openvino].

NNCF is also available via conda:

conda install -c conda-forge nncf

You may also use one of the Dockerfiles in the docker directory to build an image with an environment already set up and ready for running NNCF sample scripts.

System requirements

• Ubuntu* 18.04 or later (64-bit)
• Python* 3.7 or later
• Supported frameworks:
  • PyTorch* >=1.9.1, <1.14
  • TensorFlow* >=2.4.0, <=2.11.1
  • ONNX* ~=1.13.1
  • OpenVINO* >=2022.3.0

This repository is tested on Python* 3.8.10, PyTorch* 1.13.1 (NVidia CUDA* Toolkit 11.6) and TensorFlow* 2.11.1 (NVidia CUDA* Toolkit 11.2).

NNCF Compressed Model Zoo

Results achieved using sample scripts, example patches to third-party repositories and NNCF configuration files provided with this repository. See README.md files for sample scripts and example patches to find instruction and links to exact configuration files and final checkpoints.

• PyTorch models
  • Classification
  • Object detection
  • Semantic segmentation
  • Natural language processing (3rd-party training pipelines)
• TensorFlow models
  • Classification
  • Object detection
  • Instance segmentation

PyTorch models

Classification

Model	Compression algorithm	Dataset	Accuracy (drop) %
ResNet-50	INT8	ImageNet	76.46 (-0.31)
ResNet-50	INT8 (per-tensor only)	ImageNet	76.39 (-0.24)
ResNet-50	Mixed, 43.12% INT8 / 56.88% INT4	ImageNet	76.05 (0.10)
ResNet-50	INT8 + Sparsity 61% (RB)	ImageNet	75.42 (0.73)
ResNet-50	INT8 + Sparsity 50% (RB)	ImageNet	75.50 (0.65)
ResNet-50	Filter pruning, 40%, geometric median criterion	ImageNet	75.57 (0.58)
Inception V3	INT8	ImageNet	77.45 (-0.12)
Inception V3	INT8 + Sparsity 61% (RB)	ImageNet	76.36 (0.97)
MobileNet V2	INT8	ImageNet	71.07 (0.80)
MobileNet V2	INT8 (per-tensor only)	ImageNet	71.24 (0.63)
MobileNet V2	Mixed, 58.88% INT8 / 41.12% INT4	ImageNet	70.95 (0.92)
MobileNet V2	INT8 + Sparsity 52% (RB)	ImageNet	71.09 (0.78)
MobileNet V3 small	INT8	ImageNet	66.98 (0.68)
SqueezeNet V1.1	INT8	ImageNet	58.22 (-0.03)
SqueezeNet V1.1	INT8 (per-tensor only)	ImageNet	58.11 (0.08)
SqueezeNet V1.1	Mixed, 52.83% INT8 / 47.17% INT4	ImageNet	57.57 (0.62)
ResNet-18	XNOR (weights), scale/threshold (activations)	ImageNet	61.67 (8.09)
ResNet-18	DoReFa (weights), scale/threshold (activations)	ImageNet	61.63 (8.13)
ResNet-18	Filter pruning, 40%, magnitude criterion	ImageNet	69.27 (0.49)
ResNet-18	Filter pruning, 40%, geometric median criterion	ImageNet	69.31 (0.45)
ResNet-34	Filter pruning, 50%, geometric median criterion + KD	ImageNet	73.11 (0.19)
GoogLeNet	Filter pruning, 40%, geometric median criterion	ImageNet	69.47 (0.30)

Object detection

Model	Compression algorithm	Dataset	mAP (drop) %
SSD300-MobileNet	INT8 + Sparsity 70% (Magnitude)	VOC12+07 train, VOC07 eval	62.95 (-0.72)
SSD300-VGG-BN	INT8	VOC12+07 train, VOC07 eval	77.81 (0.47)
SSD300-VGG-BN	INT8 + Sparsity 70% (Magnitude)	VOC12+07 train, VOC07 eval	77.66 (0.62)
SSD300-VGG-BN	Filter pruning, 40%, geometric median criterion	VOC12+07 train, VOC07 eval	78.35 (-0.07)
SSD512-VGG-BN	INT8	VOC12+07 train, VOC07 eval	80.04 (0.22)
SSD512-VGG-BN	INT8 + Sparsity 70% (Magnitude)	VOC12+07 train, VOC07 eval	79.68 (0.58)

Semantic segmentation

Model	Compression algorithm	Dataset	mIoU (drop) %
UNet	INT8	CamVid	71.89 (0.06)
UNet	INT8 + Sparsity 60% (Magnitude)	CamVid	72.46 (-0.51)
ICNet	INT8	CamVid	67.89 (0.00)
ICNet	INT8 + Sparsity 60% (Magnitude)	CamVid	67.16 (0.73)
UNet	INT8	Mapillary	56.09 (0.15)
UNet	INT8 + Sparsity 60% (Magnitude)	Mapillary	55.69 (0.55)
UNet	Filter pruning, 25%, geometric median criterion	Mapillary	55.64 (0.60)

NLP (HuggingFace Transformers-powered models)

PyTorch Model	Compression algorithm	Dataset	Accuracy (Drop) %
BERT-base-chinese	INT8	XNLI	77.22 (0.46)
BERT-base-cased	INT8	CoNLL2003	99.18 (-0.01)
BERT-base-cased	INT8	MRPC	84.8 (-0.24)
BERT-large (Whole Word Masking)	INT8	SQuAD v1.1	F1: 92.68 (0.53)
RoBERTa-large	INT8	MNLI	matched: 89.25 (1.35)
DistilBERT-base	INT8	SST-2	90.3 (0.8)
MobileBERT	INT8	SQuAD v1.1	F1: 89.4 (0.58)
GPT-2	INT8	WikiText-2 (raw)	perplexity: 20.9 (-1.17)

TensorFlow models

Classification

Model	Compression algorithm	Dataset	Accuracy (drop) %
Inception V3	INT8 (per-tensor symmetric for weights, per-tensor asymmetric half-range for activations)	ImageNet	78.39 (-0.48)
Inception V3	INT8 (per-tensor symmetric for weights, per-tensor asymmetric half-range for activations), Sparsity 61% (RB)	ImageNet	77.52 (0.39)
Inception V3	Sparsity 54% (Magnitude)	ImageNet	77.86 (0.05)
MobileNet V2	INT8 (per-tensor symmetric for weights, per-tensor asymmetric half-range for activations)	ImageNet	71.63 (0.22)
MobileNet V2	INT8 (per-tensor symmetric for weights, per-tensor asymmetric half-range for activations), Sparsity 52% (RB)	ImageNet	70.94 (0.91)
MobileNet V2	Sparsity 50% (RB)	ImageNet	71.34 (0.51)
MobileNet V2 (TensorFlow Hub MobileNet V2)	Sparsity 35% (Magnitude)	ImageNet	71.87 (-0.02)
MobileNet V3 (Small)	INT8 (per-channel symmetric for weights, per-tensor asymmetric half-range for activations)	ImageNet	67.79 (0.59)
MobileNet V3 (Small)	INT8 (per-channel symmetric for weights, per-tensor asymmetric half-range for activations) + Sparsity 42% (Magnitude)	ImageNet	67.44 (0.94)
MobileNet V3 (Large)	INT8 (per-channel symmetric for weights, per-tensor asymmetric half-range for activations)	ImageNet	75.04 (0.76)
MobileNet V3 (Large)	INT8 (per-channel symmetric for weights, per-tensor asymmetric half-range for activations) + Sparsity 42% (RB)	ImageNet	75.24 (0.56)
ResNet-50	INT8	ImageNet	74.99 (0.06)
ResNet-50	INT8 (per-tensor symmetric for weights, per-tensor asymmetric half-range for activations) + Sparsity 65% (RB)	ImageNet	74.36 (0.69)
ResNet-50	Sparsity 80% (RB)	ImageNet	74.38 (0.67)
ResNet-50	Filter pruning, 40%, geometric median criterion	ImageNet	74.96 (0.09)
ResNet-50	INT8 (per-tensor symmetric for weights, per-tensor asymmetric half-range for activations) + Filter pruning, 40%, geometric median criterion	ImageNet	75.09 (-0.04)

Object detection

Model	Compression algorithm	Dataset	mAP (drop) %
RetinaNet	INT8 (per-tensor symmetric for weights, per-tensor asymmetric half-range for activations)	COCO 2017	33.12 (0.31)
RetinaNet	Magnitude sparsity (50%)	COCO 2017	33.10 (0.33)
RetinaNet	Filter pruning, 40%	COCO 2017	32.72 (0.71)
RetinaNet	INT8 (per-tensor symmetric for weights, per-tensor asymmetric half-range for activations) + filter pruning 40%	COCO 2017	32.67 (0.76)
YOLO v4	INT8 (per-channel symmetric for weights, per-tensor asymmetric half-range for activations)	COCO 2017	46.20 (0.87)
YOLO v4	Magnitude sparsity, 50%	COCO 2017	46.49 (0.58)

Instance segmentation

Model	Compression algorithm	Dataset	mAP (drop) %
Mask-R-CNN	INT8 (per-tensor symmetric for weights, per-tensor asymmetric half-range for activations)	COCO 2017	37.19 (0.14)
Mask-R-CNN	Magnitude sparsity, 50%	COCO 2017	36.94 (0.39)

ONNX models

Classification

ONNX Model	Compression algorithm	Dataset	Accuracy (Drop) %
ResNet-50	INT8 (Post-Training)	ImageNet	74.63 (0.21)
ShuffleNet	INT8 (Post-Training)	ImageNet	47.25 (0.18)
GoogleNet	INT8 (Post-Training)	ImageNet	66.36 (0.3)
SqueezeNet V1.0	INT8 (Post-Training)	ImageNet	54.3 (0.54)
MobileNet V2	INT8 (Post-Training)	ImageNet	71.38 (0.49)
DenseNet-121	INT8 (Post-Training)	ImageNet	60.16 (0.8)
VGG-16	INT8 (Post-Training)	ImageNet	72.02 (0.0)

Object Detection

ONNX Model	Compression algorithm	Dataset	mAP (drop) %
SSD1200	INT8 (Post-Training)	COCO2017	20.17 (0.17)
Tiny-YOLOv2	INT8 (Post-Training)	VOC12	29.03 (0.23)

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