Intel Model Zoo v1.8.1

New Topologies and Models:

• DLRM (BFloat16) Training for Recommendation

DL frameworks (TensorFlow):

TensorFlow models in v1.8.1 release are validated on the following TensorFlow versions:

• Intel Optimizations for TensorFlow v2.3.0
• Intel Optimizations for TensorFlow serving v2.2

DL frameworks (PyTorch):

PyTorch models in v1.8.1 release are validated on the following PyTorch version:

• PyTorch v1.5.0-rc3

Supported Configurations:

Intel Model Zoo v1.8.1 is validated on the following operating system:

• Ubuntu 18.04 LTS
• Python 3.6
• Docker Server v19+
• Docker Client v18+

Intel Model Zoo v1.8.0

New Topologies and Models:

• New pre-trained model for SSD-MobileNet with more optimizations
• BERT (FP32) Inference for Language Translation

New in TensorFlow Serving:

• InceptionV3
• ResNet50v1.5
• Transformer-LT (Official)
• SSD-MobileNet

New Tutorials:

• TensorFlow – BERT Large BFloat16 Training
• TensorFlow Serving – Installation Guide
• TensorFlow Serving – General Best Practices
• TensorFlow Serving – InceptionV3 and ResNet50v1.5
• TensorFlow Serving – SSD-MobileNet and R-FCN
• TensorFlow Serving – Transformer-LT (Official)

Bug fixes:

• Fixed MLpref GNMT with TensorFlow 2.x version
• Fixed SSD-ResNet34 FP32 Inference libGL.so.1 import error
• Fixed unit tests and linting
• Several minor bug fixes and improvements.

DL frameworks (TensorFlow):

Intel Model Zoo v1.8.0 is validated on the following TensorFlow versions:

• Intel Optimizations for TensorFlow v2.2
• Intel Optimizations for TensorFlow serving v2.2

Supported Configurations:

Intel Model Zoo v1.8.0 is validated on the following operating system:

• Ubuntu 18.04 LTS
• Python 3.6
• Docker Server v19+
• Docker Client v18+

Release v1.6.1

New functionality:

• Added experimental BFloat16 support for following models
  • Resnet50 v1.5 Training & Inference
  • BERT-Large (Squad) Training & Inference
  • SSD-ResNet34 Training
  • Transformer LT Training
• Added multi instance support training for following models
  • Resnet50 v1.5 Training
  • BERT-Large (Squad) Training
  • SSD-ResNet34 Training
  • Transformer LT Training
• Added following new FP32 model scripts
  • BERT-Large (Squad) Training & Inference
  • SSD-ResNet34 Training
  • Transformer LT Training

Bug fixes:

• Fixed Resnet50 v1.5 training issue while rerunning in same workspace without deleting checkpoints
  Several minor bug fixes and improvements.

Validated Configurations

• Model Zoo v1.6.1 is validated on the following operating system:
  • Ubuntu 18.04

Release 1.6.0

New functionality:

• Support Intel® Optimized TensorFlow v2.1.0 (most models are also compatible with Intel® Optimized TensorFlow v2.0.0)
• Porting 13 supported models to TF2.0 API
• Add the management of pre-trained models with version control and update some INT8 pre-trained models
• Add multi-instance training mode support for the following models:
  • ResNet50 v1.5
• Add MiniGo 9x9 support
• Updated mobilenet fp32 pb for better performance
• Add batch parallel support for Wide_Deep_Large_DS
• Enhance bare metal benchmark support
• Align the default container image to TF 2.1.0 image in intel namespace

Bug fixes:

• Fixed the SSD-Mobilenet to run in TF2.1.0
• Fixed the MLPerf-GNMT tensorflow-addons issue
• Fixed some security issues found by bandit scan
• Fixed for Pillow CVE-2019-19911 and CVE-2020-5313

Supported Models

Use Case	Model	Mode	Instructions
Image Recognition	DenseNet169	Inference	FP32
Image Recognition	Inception V3	Inference	Int8 FP32
Image Recognition	Inception V4	Inference	Int8 FP32
Image Recognition	MobileNet V1*	Inference	Int8 FP32
Image Recognition	ResNet 101	Inference	Int8 FP32
Image Recognition	ResNet 50	Inference	Int8 FP32
Image Recognition	ResNet 50v1.5*	Inference	Int8 FP32
Image Recognition	ResNet 50v1.5*	Training	FP32
Reinforcement	MiniGo	Training	FP32
Language Translation	GNMT*	Inference	FP32
Language Translation	Transformer_LT_Official	Inference	FP32
Object Detection	R-FCN	Inference	Int8 FP32
Object Detection	SSD-MobileNet*	Inference	Int8 FP32
Object Detection	SSD-ResNet34*	Inference	Int8 FP32
Recommendation	Wide & Deep Large Dataset	Inference	Int8 FP32
Recommendation	Wide & Deep	Inference	FP32

Validated Configurations

Model Zoo v1.6.0 were validated on the following operating systems:

• Ubuntu 18.04, 16.04
• CentOS 7.6

Release 1.5.0

New functionality:

• Support Intel® Optimized TensorFlow v1.15.0 (most models are also compatible with Intel® Optimized TensorFlow v1.14.0)

• Add the management of pre-trained models with version control and update some INT8 pre-trained models

• Enable the inference accuracy and performance of Faster-RCNN & R-FCN with Python3

• Add training mode support for the following models:
  ** GNMT
  ** SSD-ResNet34
  ** Wide&Deep_Large_Dataset

• Enhance bare metal benchmark support

• Update Wide&Deep_large_Dataset with large feature column optimization

Bug fixes:

• Fix the inceptionv3, resnet101, resnet50, resnetv1.5 to run in TF1.14
• Fix python dependency requirements for SSD-ResNet34, Faster-RCNN, and R-FCN
• Fix MTCC bare metal test issues
• Fix GNMT FP32 inference issues
• Fix Mask-RCNN inference issues
• Fix some code format check issues
• Fix some security issues found by bandit scan
• Correct TensorFlow session config for MobileNet-v1 benchmark
• Fix Pillow version

Other changes:

• Change the FP32 performance test of RFCN and Faster-RCNN from checkpoint to PB
• Remove TensorFlow Server partial support
• Remove models that were failing because of TensorFlow API changes

Release 1.4.1

Bug Fixes

• Fixed Pillow version used by unet model.
• Updated unet model documentation to correct checkpoint path.
• Fixed libsnd dependency issue for wavenet model.

Release 1.4.0

Release 1.4.0

New scripts:

• lm-1b FP32 inference
• MobileNet V1 Int8 inference
• DenseNet 169 FP32 inference
• SSD-VGG16 FP32 and Int8 inference
• SSD-ResNet34 Int8 inference
• ResNet50 v1.5 FP32 and Int8 inference
• Inception V3 FP32 inference using TensorFlow Serving

Other script changes and bug fixes:

• Updated SSD-MobileNet accuracy script to take a full path to the coco_val.records, rather than a directory
• Added a deprecation warning for using checkpoint files
• Changed Inception ResNet V2 FP32 to use a frozen graph rather than checkpoints
• Added support for custom volume mounts when running with docker
• Moved model default env var configs to config.json files
• Added support for dummy data with MobileNet V1 FP32
• Added support for TCMalloc (enabled by default for int8 models)
• Updated model zoo unit test to use json files for model parameters
• Made the reference file optional for Transformer LT performance testing
• Added iteration time to accuracy scripts
• Updated Transformer LT Official to support num_inter and num_intra threads
• Fixed path to the calibration script for ResNet101 Int8

New tutorials:

• Transformer LT inference using TensorFlow
• Transformer LT inference using TensorFlow Serving
• ResNet50 Int8 inference using TensorFlow Serving
• SSD-MobileNet inference using TensorFlow Serving

Documentation updates:

• Added Contribute.md doc with instructions on adding new models
• Added note about setting environment variables when running on bare metal
• Updated model README files to use TensorFlow 1.14.0 docker images (except for Wide and Deep int8)
• Updated FasterRCNN Int8 README file to clarify that performance testing uses raw images
• Fixed docker build command in the TensorFlow Serving Installation Guide
• NCF documentation update to remove line of code that causes an error
• Updated mlperf/inference branch and paths in README file

Known issues:

• RFCN FP32 accuracy is not working with the gcr.io/deeplearning-platform-release/tf-cpu.1-14 docker image
• The TensorFlow Serving Installation Guide still shows example commands that build version 1.13. This will be updated to 1.14 when the official TensorFlow Serving release tag exists. To build version 1.14 now, you can use one of the following values for TF_SERVING_VERSION_GIT_BRANCH in your multi-stage docker build: "1.14.0-rc0" or "r1.14".

v1.3.1

Revised language regarding performance expectations.

v1.3.0

Release 1.3.0

New benchmarking scripts:

• FaceNet FP32 inference
• GNMT FP32 inference
• Inception ResNet V2 Int8 inference
• Inception V4 Int8 inference
• MTCC FP32 inference
• RFCN Int8 inference
• SSD-MobileNet Int8 inference
• SSD-ResNet34 FP32 inference
• Transformer LT Official FP32 inference

Other script changes and bug fixes:

• Renamed Fast RCNN to Faster RCNN
• Fixed SSD-MobileNet FP32 inference container error with python3
• Added python file to download and preprocess the Wide and Deep census dataset
• Added ability for ResNet50 FP32 --output-results to work with benchmarking
• Added --data-num-inter-threads and --data-num-intra-threads to the launch script (currently supported by ResNet50, ResNet101, and InceptionV3)
• Added data layer optimization and calibration option for ResNet50, ResNet101 and InceptionV3
• Bug fixes and an arg update for Wide and Deep large dataset
• Only print lscpu info with verbose logging
• Reduced duplicated code in Wide and Deep inference scripts
• Added ability to run benchmarking script without docker
• ResNet50 fix for the issue of not reporting the average of all segments

New tutorials:

• ResNet101 and Inception V3 tutorial contents
• TensorFlow Serving Object Detection Tutorial
• TensorFlow Recommendation System Tutorial
• ResNet50 Quantization Tutorial

Documentation updates:

• Improved main README with repo purpose and structure
• Updated NCF README file
• Added links to the arXiv papers for each model
• Updated TF Serving BKMs for split parallelism vars
• Added note to TF BKM about KMP_AFFINITY when HT is off

v1.2.1

Release 1.2.1

Benchmarking Scripts

• Fix dummy data performance problem for RN50 FP32 and InceptionV3 FP32