zhouleidcc/neural-compressor

Intel® Neural Compressor (formerly known as Intel® Low Precision Optimization Tool), targeting to provide unified APIs for network compression technologies, such as low precision, sparsity, pruning, knowledge distillation, cross different deep learning frameworks to purse best inference performance.

Introduction to Intel® Neural Compressor

Intel® Neural Compressor (formerly known as Intel® Low Precision Optimization Tool) is an open-source Python library running on Intel CPUs and GPUs, which delivers unified interfaces across multiple deep learning frameworks for popular network compression technologies, such as quantization, pruning, knowledge distillation. This tool supports automatic accuracy-driven tuning strategies to help user quickly find out the best quantized model. It also implements different weight pruning algorithms to generate pruned model with predefined sparsity goal and supports knowledge distillation to distill the knowledge from the teacher model to the student model.

Note

GPU support is under development.

Visit the Intel® Neural Compressor online document website at: https://intel.github.io/neural-compressor.

Architecture

Intel® Neural Compressor features an infrastructure and workflow that aids in increasing performance and faster deployments across architectures.

Infrastructure

Click the image to enlarge it.

Workflow

Click the image to enlarge it.

Supported Frameworks

Supported deep learning frameworks are:

• TensorFlow*, including 1.15.0 UP3, 1.15.0 UP2, 1.15.0 UP1, 2.1.0, 2.2.0, 2.3.0, 2.4.0, 2.5.0, Official TensorFlow 2.6.0

Note: Intel Optimized TensorFlow 2.5.0 requires to set environment variable TF_ENABLE_MKL_NATIVE_FORMAT=0 before running Neural Compressor quantization or deploying the quantized model.

Note: From the official TensorFlow 2.6.0, oneDNN support has been upstreamed. Download the official TensorFlow 2.6.0 binary for the CPU device and set the environment variable TF_ENABLE_ONEDNN_OPTS=1 before running the quantization process or deploying the quantized model.

• PyTorch*, including 1.5.0+cpu, 1.8.0+cpu, 1.9.0+cpu
• Apache* MXNet, including 1.6.0, 1.7.0, 1.8.0
• ONNX* Runtime, including 1.6.0, 1.7.0, 1.8.0
• Engine, which is a built-in bare metal acceleration library for domain-specific NLP models.

Installation

Select the installation based on your operating system.

Linux Installation

You can install Neural Compressor using one of three options: Install just the library from binary or source, or get the Intel-optimized framework together with the library by installing the Intel® oneAPI AI Analytics Toolkit.

Option 1 Install from binary

# install stable version from pip
pip install neural-compressor

# install nightly version from pip
pip install -i https://test.pypi.org/simple/ neural-compressor

# install stable version from from conda
conda install neural-compressor -c conda-forge -c intel 

Option 2 Install from source

git clone https://github.com/intel/neural-compressor.git
cd neural-compressor
pip install -r requirements.txt
python setup.py install

Option 3 Install from AI Kit

The Intel® Neural Compressor library is released as part of the Intel® oneAPI AI Analytics Toolkit (AI Kit). The AI Kit provides a consolidated package of Intel's latest deep learning and machine optimizations all in one place for ease of development. Along with Neural Compressor, the AI Kit includes Intel-optimized versions of deep learning frameworks (such as TensorFlow and PyTorch) and high-performing Python libraries to streamline end-to-end data science and AI workflows on Intel architectures.

The AI Kit is distributed through many common channels, including from Intel's website, YUM, APT, Anaconda, and more. Select and download the AI Kit distribution package that's best suited for you and follow the Get Started Guide for post-installation instructions.

Download AI Kit	AI Kit Get Started Guide

Windows Installation

Prerequisites

The following prerequisites and requirements must be satisfied for a successful installation:

• Python version: 3.6 or 3.7 or 3.8 or 3.9

• Download and install anaconda.

• Create a virtual environment named nc in anaconda:

  # Here we install python 3.7 for instance. You can also choose python 3.6, 3.8, or 3.9.
  conda create -n nc python=3.7
  conda activate nc

Installation options

Option 1 Install from binary

# install stable version from pip
pip install neural-compressor

# install nightly version from pip
pip install -i https://test.pypi.org/simple/ neural-compressor

# install from conda
conda install neural-compressor -c conda-forge -c intel 

Option 2 Install from source

git clone https://github.com/intel/neural-compressor.git
cd neural-compressor
pip install -r requirements.txt
python setup.py install

Note: To run examples from neural-compressor version<1.7 with neural-compressor binary version>=1.7, please replace model name in main.py.

sed -i "s|lpot|neural_compressor|g" main.py

Documentation

Get Started

• APIs explains Intel® Neural Compressor's API.
• Transform introduces how to utilize Neural Compressor's built-in data processing and how to develop a custom data processing method.
• Dataset introduces how to utilize Neural Compressor's built-in dataset and how to develop a custom dataset.
• Metric introduces how to utilize Neural Compressor's built-in metrics and how to develop a custom metric.
• Tutorial provides comprehensive instructions on how to utilize Neural Compressor's features with examples.
• Examples are provided to demonstrate the usage of Neural Compressor in different frameworks: TensorFlow, PyTorch, MXNet, and ONNX Runtime.
• Intel® Neural Compressor Bench is a web-based system used to simplify Intel® Neural Compressor usage.
• Intel oneAPI AI Analytics Toolkit Get Started Guide explains the AI Kit components, installation and configuration guides, and instructions for building and running sample apps.
• AI and Analytics Samples includes code samples for Intel oneAPI libraries.

Deep Dive

• Quantization are processes that enable inference and training by performing computations at low-precision data types, such as fixed-point integers. Neural Compressor supports Post-Training Quantization (PTQ) with different quantization capabilities and Quantization-Aware Training (QAT). Note that (Dynamic Quantization) currently has limited support.
• Pruning provides a common method for introducing sparsity in weights and activations.
• Knowledge Distillation provides a common method for distilling knowledge from teacher model to student model.
• Distributed Training introduces how to leverage Horovod to do multi-node training in Intel® Neural Compressor to speed up the training time.
• Benchmarking introduces how to utilize the benchmark interface of Neural Compressor.
• Mixed precision introduces how to enable mixed precision, including BFP16 and int8 and FP32, on Intel platforms during tuning.
• Graph Optimization introduces how to enable graph optimization for FP32 and auto-mixed precision.
• Model Conversion introduces how to convert TensorFlow QAT model to quantized model running on Intel platforms.
• TensorBoard provides tensor histograms and execution graphs for tuning debugging purposes.

Advanced Topics

• Engine is a new backend supported by Intel® Neural Compressor to support domain-specific acceleration for NLP models.
• Adaptor is the interface between components and framework. The method to develop adaptor extension is introduced with ONNX Runtime as example.
• Strategy can automatically optimized low-precision recipes for deep learning models to achieve optimal product objectives like inference performance and memory usage with expected accuracy criteria. The method to develop a new strategy is introduced.

Publications

• MLPerf™ Performance Gains Abound with latest 3rd Generation Intel® Xeon® Scalable Processors (Apr 2021)
• 3D Digital Face Reconstruction Solution enabled by 3rd Gen Intel® Xeon® Scalable Processors (Apr 2021)
• Accelerating Alibaba Transformer model performance with 3rd Gen Intel® Xeon® Scalable Processors (Ice Lake) and Intel® Deep Learning Boost (Apr 2021)
• Using Low-Precision Optimizations for High-Performance DL Inference Applications (Apr 2021)
• DL Boost Quantization with CERN's 3D-GANs model (Feb 2021)

Full publication list please refers to here

System Requirements

Intel® Neural Compressor supports systems based on Intel 64 architecture or compatible processors, specially optimized for the following CPUs:

• Intel Xeon Scalable processor (formerly Skylake, Cascade Lake, Cooper Lake, and Icelake)
• future Intel Xeon Scalable processor (code name Sapphire Rapids)

Intel® Neural Compressor requires installing the Intel-optimized framework version for the supported DL framework you use: TensorFlow, PyTorch, MXNet, or ONNX runtime.

Note: Intel Neural Compressor supports Intel-optimized and official frameworks for some TensorFlow versions. Refer to Supported Frameworks for specifics.

Validated Hardware/Software Environment

Platform	OS	Python	Framework	Version
Cascade Lake Cooper Lake Skylake Ice Lake	CentOS 8.3 Ubuntu 18.04	3.6 3.7 3.8 3.9	TensorFlow	2.6.0
				2.5.0
				2.4.0
				2.3.0
				2.2.0
				2.1.0
				1.15.0 UP1
				1.15.0 UP2
				1.15.0 UP3
				1.15.2
			PyTorch	1.5.0+cpu
				1.8.0+cpu
				1.9.0+cpu
				IPEX
			MXNet	1.8.0
				1.7.0
				1.6.0
			ONNX Runtime	1.6.0
				1.7.0
				1.8.0

Validated Models

Intel® Neural Compressor provides numerous examples to show promising accuracy loss with the best performance gain. A full quantized model list on various frameworks is available in the Model List.

Validated MLPerf Models

Model	Framework	Support	Example
ResNet50 v1.5	TensorFlow	Yes	Link
	PyTorch	Yes	Link
DLRM	PyTorch	Yes	Link
BERT-large	TensorFlow	Yes	Link
	PyTorch	Yes	Link
SSD-ResNet34	TensorFlow	Yes	Link
	PyTorch	Yes	Link
RNN-T	PyTorch	WIP
3D-UNet	TensorFlow	WIP
	PyTorch	Yes	Link

Validated Quantized Models

Framework	Version	Model	Accuracy			Performance
			INT8 Tuning Accuracy	FP32 Accuracy Baseline	Acc Ratio [(INT8-FP32)/FP32]	INT8 realtime(ms) CLX8280 1s 4c per instance	FP32 realtime(ms) CLX8280 1s 4c per instance	Realtime Latency Ratio[FP32/INT8]
tensorflow	2.5.0	resnet50v1.0	74.24%	74.27%	-0.04%	7.56	21.24	2.81x
tensorflow	2.5.0	resnet50v1.5	76.94%	76.46%	0.63%	9.64	24.86	2.58x
tensorflow	2.5.0	resnet101	77.21%	76.45%	0.99%	12.73	30.80	2.42x
tensorflow	2.5.0	inception_v1	70.30%	69.74%	0.80%	5.57	9.92	1.78x
tensorflow	2.5.0	inception_v2	74.27%	73.97%	0.41%	6.69	12.33	1.84x
tensorflow	2.5.0	inception_v3	77.29%	76.75%	0.70%	12.90	27.46	2.13x
tensorflow	2.5.0	inception_v4	80.36%	80.27%	0.11%	20.88	54.13	2.59x
tensorflow	2.5.0	inception_resnet_v2	80.42%	80.40%	0.02%	44.47	87.69	1.97x
tensorflow	2.5.0	mobilenetv1	73.93%	70.96%	4.19%	2.95	10.12	3.43x
tensorflow	2.5.0	mobilenetv2	71.96%	71.76%	0.28%	4.97	10.39	2.09x
tensorflow	2.5.0	ssd_resnet50_v1	37.91%	38.00%	-0.24%	140.46	411.03	2.93x
tensorflow	2.5.0	ssd_mobilenet_v1	23.02%	23.13%	-0.48%	12.25	26.90	2.20x
tensorflow	2.5.0	ssd_resnet34	21.97%	22.16%	-0.86%	264.26	960.48	3.63x

Framework	Version	Model	Accuracy			Performance
			INT8 Tuning Accuracy	FP32 Accuracy Baseline	Acc Ratio [(INT8-FP32)/FP32]	INT8 realtime(ms) CLX8280 1s 4c per instance	FP32 realtime(ms) CLX8280 1s 4c per instance	Realtime Latency Ratio[FP32/INT8]
pytorch	1.9.0+cpu	resnet18	69.58%	69.76%	-0.26%	14.21	26.55	1.87x
pytorch	1.9.0+cpu	resnet50	75.87%	76.13%	-0.34%	24.89	53.84	2.16x
pytorch	1.9.0+cpu	resnext101_32x8d	79.09%	79.31%	-0.28%	64.03	147.51	2.30x
pytorch	1.9.0+cpu	bert_base_mrpc	88.16%	88.73%	-0.64%	41.15	81.56	1.98x
pytorch	1.9.0+cpu	bert_base_cola	58.29%	58.84%	-0.93%	39.17	83.42	2.13x
pytorch	1.9.0+cpu	bert_base_sts-b	88.65%	89.27%	-0.70%	39.59	83.07	2.10x
pytorch	1.9.0+cpu	bert_base_sst-2	91.63%	91.86%	-0.25%	39.39	83.17	2.11x
pytorch	1.9.0+cpu	bert_base_rte	69.31%	69.68%	-0.52%	39.51	81.84	2.07x
pytorch	1.9.0+cpu	bert_large_mrpc	87.48%	88.33%	-0.95%	112.80	281.91	2.50x
pytorch	1.9.0+cpu	bert_large_squad	92.78988	93.04683	-0.28%	503.92	934.01	1.85x
pytorch	1.9.0+cpu	bert_large_qnli	91.12%	91.82%	-0.76%	111.08	289.13	2.60x
pytorch	1.9.0+cpu	bert_large_rte	72.92%	72.56%	0.50%	151.93	298.53	1.96x
pytorch	1.9.0+cpu	bert_large_cola	62.85%	62.57%	0.45%	113.04	285.43	2.52x

Validated Pruning Models

Tasks	FWK	Model	fp32 baseline	gradient sensitivity with 20% sparsity			+onnx dynamic quantization on pruned model
				accuracy%	drop%	perf gain (sample/s)	accuracy%	drop%	perf gain (sample/s)
SST-2	pytorch	bert-base	accuracy = 92.32	accuracy = 91.97	-0.38	1.30x	accuracy = 92.20	-0.13	1.86x
QQP	pytorch	bert-base	[accuracy, f1] = [91.10, 88.05]	[accuracy, f1] = [89.97, 86.54]	[-1.24, -1.71]	1.32x	[accuracy, f1] = [89.75, 86.60]	[-1.48, -1.65]	1.81x

Tasks	FWK	Model	fp32 baseline	Pattern Lock on 70% Unstructured Sparsity		Pattern Lock on 50% 1:2 Structured Sparsity
				accuracy%	drop%	accuracy%	drop%
MNLI	pytorch	bert-base	[m, mm] = [84.57, 84.79]	[m, mm] = [82.45, 83.27]	[-2.51, -1.80]	[m, mm] = [83.20, 84.11]	[-1.62, -0.80]
SST-2	pytorch	bert-base	accuracy = 92.32	accuracy = 91.51	-0.88	accuracy = 92.20	-0.13
QQP	pytorch	bert-base	[accuracy, f1] = [91.10, 88.05]	[accuracy, f1] = [90.48, 87.06]	[-0.68, -1.12]	[accuracy, f1] = [90.92, 87.78]	[-0.20, -0.31]
QNLI	pytorch	bert-base	accuracy = 91.54	accuracy = 90.39	-1.26	accuracy = 90.87	-0.73
QnA	pytorch	bert-base	[em, f1] = [79.34, 87.10]	[em, f1] = [77.27, 85.75]	[-2.61, -1.54]	[em, f1] = [78.03, 86.50]	[-1.65, -0.69]

Framework	Model	fp32 baseline	Compression	dataset	acc(drop)%
Pytorch	resnet18	69.76	30% sparsity on magnitude	ImageNet	69.47(-0.42)
Pytorch	resnet18	69.76	30% sparsity on gradient sensitivity	ImageNet	68.85(-1.30)
Pytorch	resnet50	76.13	30% sparsity on magnitude	ImageNet	76.11(-0.03)
Pytorch	resnet50	76.13	30% sparsity on magnitude and post training quantization	ImageNet	76.01(-0.16)
Pytorch	resnet50	76.13	30% sparsity on magnitude and quantization aware training	ImageNet	75.90(-0.30)

Validated Knowledge Distillation Examples

Example Name	Dataset	Student (Accuracy)	Teacher (Accuracy)	Student With Distillation (Accuracy Improvement)
ResNet example	ImageNet	ResNet18 (0.6739)	ResNet50 (0.7399)	0.6845 (0.0106)
BlendCnn example	MRPC	BlendCnn (0.7034)	BERT-Base (0.8382)	0.7034 (0)
BiLSTM example	SST-2	BiLSTM (0.7913)	RoBERTa-Base (0.9404)	0.8085 (0.0172)

Validated Engine Examples

Model	Accuracy			Performance
	INT8 Tuning Accuracy	FP32 Accuracy Baseline	Acc Ratio [(INT8-FP32)/FP32]	INT8 realtime(ms) CLX8280 1s 4c per instance	FP32 realtime(ms) CLX8280 1s 4c per instance	Realtime Latency Ratio[FP32/INT8]
bert_base_mrpc	0.8235	83.09%	-0.89%	21.91	71.53	3.26x
bert_large	90.6648	90.87	-0.23%	232.38	954.96	4.11x
distilbert_base_uncased_mrpc	0.8407	84.31%	-0.28%	10.41	36.42	3.50x

Additional Content

• Release Information
• Contribution Guidelines
• Legal
• Security Policy
• Intel® Neural Compressor Website