For the installation instructions, click here.
NNCF provides a suite of advanced algorithms for Neural Networks inference optimization in OpenVINO™ with minimal accuracy drop.
NNCF is designed to work with models from PyTorch and TensorFlow.
NNCF provides samples that demonstrate the usage of compression algorithms for three different use cases on public PyTorch and TensorFlow models and datasets: Image Classification, Object Detection and Semantic Segmentation. 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.
- Support of various compression algorithms, applied during a model fine-tuning process to achieve a better performance-accuracy trade-off:
| Compression algorithm | PyTorch | TensorFlow |
|---|---|---|
| Quantization | Supported | Supported |
| Mixed-Precision Quantization | Supported | Not supported |
| Binarization | Supported | Not supported |
| Sparsity | Supported | Supported |
| Filter pruning | Supported | 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.
- Configuration file examples for each supported compression algorithm.
- Git patches for prominent third-party repositories (huggingface-transformers) demonstrating the process of integrating NNCF into custom training pipelines
- 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.
The NNCF is organized as a regular Python package that can be imported in your target training pipeline script.
The basic workflow is loading a JSON configuration script containing NNCF-specific parameters determining the compression to be applied to your model, and then passing your model along with the configuration script to the create_compressed_model function.
This function returns a model with additional modifications necessary to enable algorithm-specific compression during fine-tuning and handle to the object allowing you to control the compression during the training process:
import torch
import nncf # 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.
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. For in-depth examples of NNCF integration, browse the sample scripts code, or the example patches to third-party repositories. For FAQ, visit this link.
For a quicker start with NNCF-powered compression, you can also try the sample scripts, each of which provides a basic training pipeline for classification, semantic segmentation and object detection neural network training correspondingly.
To run the samples please refer to the corresponding tutorials:
- PyTorch samples:
- TensorFlow samples:
A collection of ready-to-run Jupyter* notebooks are also available to demonstrate how to use NNCF compression algorithms to optimize models for inference with the OpenVINO Toolkit.
- Optimizing PyTorch models with NNCF of OpenVINO by 8-bit quantization
- Optimizing TensorFlow models with NNCF of OpenVINO by 8-bit quantization
- Post-Training Quantization of Pytorch model with NNCF
NNCF may be straightforwardly integrated into training/evaluation pipelines of third-party repositories.
-
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.
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:
- Ubuntu* 18.04 or later (64-bit)
- Python* 3.7 or later
- Supported frameworks:
- PyTorch* 1.12.1
- TensorFlow* >=2.4.0, <=2.8.2
This repository is tested on Python* 3.8.10, PyTorch* 1.12.1 (NVidia CUDA* Toolkit 11.6) and TensorFlow* 2.8.2 (NVidia CUDA* Toolkit 11.2).
We suggest to install or use the package in the Python virtual environment.
If you want to optimize a model from PyTorch, install PyTorch by following PyTorch installation guide. If you want to optimize a model from TensorFlow, install TensorFlow by following TensorFlow installation guide.
Install the package and its dependencies by running the following in the repository root directory:
pip install .
Note that if you install NNCF in this manner, the backend frameworks supported by NNCF will not be explicitly installed. NNCF will try to work with whatever backend versions you have installed in your Python environment.
If you want to install both NNCF and the supported PyTorch version in one line, you can do this by running:
pip install .[torch]
For installation of NNCF along with TensorFlow, run:
pip install .[tf]
(Experimental) For installation of NNCF for ONNXRuntime-OpenVINO, run:
pip install .[onnx]NB: For launching example scripts in this repository, we recommend setting the PYTHONPATH variable to the root of the checked-out repository once the installation is completed.
NNCF can be installed as a regular PyPI package via pip:
pip install nncf
Use the same pip install syntax as above to install NNCF along with the backend package versions in one go, i.e. for NNCF with PyTorch, run:
pip install nncf[torch]
For installation of NNCF along with TensorFlow, run:
pip install nncf[tf]
(Experimental) For installation of NNCF for ONNXRuntime-OpenVINO, run:
pip install nncf[onnx]NNCF is also available via conda:
conda install -c conda-forge nncf
pip install git+https://github.com/openvinotoolkit/nncf@bd189e2#egg=nncfNote that in order for this to work for pip versions >= 21.3, your Git version must be at least 2.22.
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.
Refer to the CONTRIBUTING.md file for guidelines on contributions to the NNCF repository.
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 Model | Dataset | Accuracy (Drop) % | |
|---|---|---|---|
| ResNet-50 | INT8 | ImageNet | 76.42 (-0.26) |
| ResNet-50 | INT8 (per-tensor for weights) | ImageNet | 76.37 (-0.21) |
| ResNet-50 | Mixed, 44.8% INT8 / 55.2% INT4 | ImageNet | 76.2 (-0.04) |
| ResNet-50 | INT8 + Sparsity 61% (RB) | ImageNet | 75.43 (0.73) |
| ResNet-50 | INT8 + Sparsity 50% (RB) | ImageNet | 75.55 (0.61) |
| ResNet-50 | Filter pruning, 40%, geometric median criterion | ImageNet | 75.62 (0.54) |
| Inception V3 | INT8 | ImageNet | 78.25 (-0.91) |
| Inception V3 | INT8 + Sparsity 61% (RB) | ImageNet | 77.58 (-0.24) |
| MobileNet V2 | INT8 | ImageNet | 71.35 (0.58) |
| MobileNet V2 | INT8 (per-tensor for weights) | ImageNet | 71.3 (0.63) |
| MobileNet V2 | Mixed, 46.6% INT8 / 53.4% INT4 | ImageNet | 70.92 (1.01) |
| MobileNet V2 | INT8 + Sparsity 52% (RB) | ImageNet | 71.11 (0.82) |
| MobileNet V3 small | INT8 | ImageNet | 66.94 (0.73) |
| SqueezeNet V1.1 | INT8 | ImageNet | 58.28 (-0.04) |
| SqueezeNet V1.1 | INT8 (per-tensor for weights) | ImageNet | 58.26 (-0.02) |
| SqueezeNet V1.1 | Mixed, 54.7% INT8 / 45.3% INT4 | ImageNet | 58.9 (-0.66) |
| ResNet-18 | XNOR (weights), scale/threshold (activations) | ImageNet | 61.63 (8.17) |
| ResNet-18 | DoReFa (weights), scale/threshold (activations) | ImageNet | 61.61 (8.19) |
| ResNet-18 | Filter pruning, 40%, magnitude criterion | ImageNet | 69.26 (0.54) |
| ResNet-18 | Filter pruning, 40%, geometric median criterion | ImageNet | 69.32 (0.48) |
| ResNet-34 | Filter pruning, 50%, geometric median criterion + KD | ImageNet | 73.11 (0.19) |
| GoogLeNet | Filter pruning, 40%, geometric median criterion | ImageNet | 68.82 (0.93) |
| PyTorch Model | Compression algorithm | Dataset | mAP (drop) % |
|---|---|---|---|
| SSD300-MobileNet | INT8 + Sparsity 70% (Magnitude) | VOC12+07 train, VOC07 eval | 62.94 (-0.71) |
| SSD300-VGG-BN | INT8 | VOC12+07 train, VOC07 eval | 77.96 (0.32) |
| SSD300-VGG-BN | INT8 + Sparsity 70% (Magnitude) | VOC12+07 train, VOC07 eval | 77.59 (0.69) |
| SSD300-VGG-BN | Filter pruning, 40%, geometric median criterion | VOC12+07 train, VOC07 eval | 77.72 (0.56) |
| SSD512-VGG-BN | INT8 | VOC12+07 train, VOC07 eval | 80.12 (0.14) |
| SSD512-VGG-BN | INT8 + Sparsity 70% (Magnitude) | VOC12+07 train, VOC07 eval | 79.67 (0.59) |
| PyTorch Model | Dataset | Accuracy (Drop) % | |
|---|---|---|---|
| UNet | INT8 | CamVid | 71.8 (0.15) |
| UNet | INT8 + Sparsity 60% (Magnitude) | CamVid | 72.03 (-0.08) |
| ICNet | INT8 | CamVid | 67.86 (0.03) |
| ICNet | INT8 + Sparsity 60% (Magnitude) | CamVid | 67.18 (0.71) |
| UNet | INT8 | Mapillary | 55.87 (0.36) |
| UNet | INT8 + Sparsity 60% (Magnitude) | Mapillary | 55.65 (0.58) |
| UNet | Filter pruning, 25%, geometric median criterion | Mapillary | 55.62 (0.61) |
| PyTorch Model | 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 Model | Compression algorithm | Dataset | Accuracy (Drop) % |
|---|---|---|---|
| Inception V3 | INT8 (per-tensor for weights) | ImageNet | 78.36 (-0.44) |
| Inception V3 | Sparsity 54% (Magnitude) | ImageNet | 77.87 (0.03) |
| Inception V3 | INT8 (per-tensor for weights) + Sparsity 61% (RB) | ImageNet | 77.58 (0.32) |
| MobileNet V2 | INT8 (per-tensor for weights) | ImageNet | 71.66 (0.19) |
| MobileNet V2 | Sparsity 50% (RB) | ImageNet | 71.34 (0.51) |
| MobileNet V2 | INT8 (per-tensor for weights) + Sparsity 52% (RB) | ImageNet | 71.0 (0.85) |
| MobileNet V3 small | INT8 (per-channel, symmetric for weights; per-tensor, asymmetric for activations) | ImageNet | 67.75 (0.63) |
| MobileNet V3 small | INT8 (per-channel, symmetric for weights; per-tensor, asymmetric for activations) + Sparsity 42% (RB) | ImageNet | 67.59 (0.79) |
| MobileNet V3 large | INT8 (per-channel, symmetric for weights; per-tensor, asymmetric for activations) | ImageNet | 75.04 (0.77) |
| MobileNet V3 large | INT8 (per-channel, symmetric for weights; per-tensor, asymmetric for activations) + Sparsity 42% (RB) | ImageNet | 75.29 (0.52) |
| ResNet50 | INT8 (per-tensor for weights) | ImageNet | 75.0 (0.04) |
| ResNet50 | Sparsity 80% (RB) | ImageNet | 74.36 (0.68) |
| ResNet50 | INT8 (per-tensor for weightsy) + Sparsity 65% (RB) | ImageNet | 74.3 (0.74) |
| ResNet50 | Filter Pruning 40%, geometric_median criterion | ImageNet | 74.98 (0.06) |
| ResNet50 | Filter Pruning 40%, geometric_median criterion + INT8 (per-tensor for weights) | ImageNet | 75.08 (-0.04) |
| TensorFlow Hub MobileNet V2 | Sparsity 35% (Magnitude) | ImageNet | 71.90 (-0.06) |
| TensorFlow Model | Compression algorithm | Dataset | mAP (drop) % |
|---|---|---|---|
| RetinaNet | INT8 (per-tensor for weights) | COCO2017 | 33.18 (0.26) |
| RetinaNet | Sparsity 50% (Magnitude) | COCO2017 | 33.13 (0.31) |
| RetinaNet | Filter Pruning 40%, geometric_median criterion | COCO2017 | 32.7 (0.74) |
| RetinaNet | Filter Pruning 40%, geometric_median criterion + INT8 (per-tensor for weights) | COCO2017 | 32.68 (0.76) |
| YOLOv4 | INT8 (per-channel, symmetric for weights; per-tensor, asymmetric for activations) | COCO2017 | 46.30 (0.74) |
| YOLOv4 | Sparsity 50% (Magnitude) | COCO2017 | 46.54 (0.50) |
| TensorFlow Model | Dataset | mAP (drop) % | |
|---|---|---|---|
| MaskRCNN | INT8 (per-tensor for weights) | COCO2017 | bbox: 37.27 (0.06) segm: 33.54 (0.02) |
| MaskRCNN | Sparsity 50% (Magnitude) | COCO2017 | bbox: 36.93 (0.40) segm: 33.23 (0.33) |
@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}
}
- Documentation
- Example scripts (model objects available through links in respective README.md files):
- FAQ
- Notebooks
- HuggingFace Optimum Intel utilizes NNCF as a compression backend within the renowned
transformersrepository. - Model Optimization Guide
[*] Other names and brands may be claimed as the property of others.