Release 1.1.0

This release of Intel® AI Quantization Tools for TensorFlow* 1.1.0 is released under v1.1.0 tag (https://github.com/IntelAI/tools/tree/v1.1.0). Please notes that Intel® AI Quantization Tools for TensorFlow* requires Intel® Optimizations for TensorFlow. Intel® AI Quantization Tools for Tensorflow 1.1.0 is the last official release and the code has been migrated to Intel® Low Precision Inference Tool (iLiT) as one of backend framework engines. This revision contains the following features and fixes:

New functionality:

• Add experimental KL Divergence and Moving Average algorithms to the tool

• Add support using the Intel® AI Quantization Tool for TensoFlow* Python Programming APIs for the following out-of-box models:

  • ResNet-50v1.0

  • ResNet-50v1.5

  • ResNet-101

  • SSD-ResNet34

  • MobileNetv1

  • SSD-MobileNet

  • Faster-RCNN

  • R-FCN

  • Inception_v3

  • inception_v1

  • inception_v2

  • inception_v4

  • vgg_16

  • vgg_19

  • mobilenet_v2

  • mobilenet_v1

  • resnet_v1_152

  • resnet_v1_50

  • mask_rcnn_resnet50 (not for ITF2.1)

  • mask_rcnn_resnet101 (not for ITF2.1)

  • mask_rcnn_inception_v2

  • mask_rcnn_inception_resnet_v2 (not for ITF2.1)

  • faster_rcnn_resnet101

  • rfcn_resnet101

  • faster_rcnn_resnet50

  • faster_rcnn_inception_v2

  • ssd_resnet50_v1

  • ssd_mobilenet_v1

  • resnet_v2_101

  • resnet_v2_152

  • resnet_v2_50

  • inception_resnet_v2

• Add Convolution/FusedBatchnormv3 fusion supporting

Bug fixes:

• Fix several bugs introduced by quantize_graph refactor.

• Fix the flake8 errors.

Release 1.0.0

Intel® AI Quantization Tools for Tensorflow 1.0 supports Intel® Optimized TensorFlow v1.15.0, v2.0.0 and v2.1.0, which added support using the Intel® AI Quantization Tool for TensoFlow* Python Programming APIs for the out-of-box models. For TensorFlow 2.1.0, we also added s8 support.

New functionality:

• Add Convolution s8 support for TensorFlow 2.1.
• Add optimize_for_inference() equivalent logic to the tool.
• Add support using the Intel® AI Quantization Tool for TensoFlow* Python Programming APIs for the following out-of-box models:
  • inception_v1
  • inception_v2
  • inception_v4
  • vgg_16
  • vgg_19
  • resnet_v1_152
  • resnet_v1_50
• Add support for TensorFlow 2.1.

Bug fixes:

• Fix bugs found in s8 test.
• Fix several bugs introduced by quantize_graph refactor.
• Fix the flake8 errors.

Release 1.0b

This release of Intel® AI Quantization Tools for Tensorflow* 1.0 Beta is released under v1.0b tag (https://github.com/IntelAI/tools/tree/v1.0b). Please notes that Intel® AI Quantization Tools for Tensorflow* must depend on Intel® Optimizations for Tensorflow. This revision contains the following features and fixes:

New functionality:

• Add support .whl Pip and conda installation for Python3.4/3.5/3.6/3.7, and remove the Tensorflow source build dependency in Intel ® optimizations for Tensorflow1.14.0, 1.15.0 and 2.0.0.
• Add three entries to run the quantization for specific models under api/examples/, including bash command for model zoo, bash command for custom model,
and python programming APIs direct call.
• Add Dockerfile for user to build the docker container.
• Add debug mode, and exclude ops and nodes in Python programming APIs.
• Add the Bridge interface with Model Zoo for Intel® Architecture.
• Add the Python implementation of summarize_graph to remove the dependency of Tensorflow source.
• Add the Python implementation of freeze_min/max, freeze_requantization_ranges,
fuse_quantized_conv_and_requantize, rerange_quantized_concat, Insert_logging of Transform_graph to remove the dependency of Tensorflow source build.
• Add per-channel support.
• Add support using the Intel® AI Quantization Tool for Tensoflow* Python Programming APIs for the following models:
**ResNet50
**ResNet50 v1.5
**SSD-MobileNet
**SSD-ResNet34
**ResNet101
**MobileNet
**Inception V3
**Faster-RCNN
**RFCN
• Add new procedures in README
• Support for Tensorflow1.14.0.
• Support for Tensorflow1.15.0.
• Support for Tensorflow 2.0.
• Support for Model Zoo for Intel® Architecture 1.5.

Bug fixes:

• Fix several bugs for Python rewrite transform_graph ops.
• Fix data types issue for optimize_for_inference.
• Fix the bug for MobileNet and ResNet101.
• Clean up the hardcode for Faster-RCNN and RFCN.
• Fix the pylint errors.

v0.4.0

This release has enabled easy-to-use “Intel® AI Quantization Tool for Tensoflow* Python Programming APIs” automating the quantization process and improve user experience, which support for Tensorflow 1.14.0 and 1.15.0 release.

New functionality:

• Enable easy-to-use “Intel® AI Quantization Tool for Tensoflow* Python Programming APIs” automating the quantization process and improve user experience.

• Add support using the Intel® AI Quantization Tool for Tensoflow* Python Programming APIs for the following models:

  • ResNet50

  • ResNet50 v1.5

  • SSD-MobileNet

  • SSD-ResNet34

• Add the Intel® AI Quantization Tool for Tensoflow* quantization procedure document.

• Add support for Tensorflow1.14.0

• Add support for Tensorflow1.15.0.

• Add support for Model Zoo for Intel® Architecture 1.4.

Bug fixes:

• Enhance the fuse PAD op for ResNet50v1.5.

• Fix the docker build.

• Fix the bug for Wide & Deep.

• Fix the bug for ResNet101.

• Fix the pylint errors.

• Fix the BUILD for graph_ transform.

• Fix re-quantize fusion for non-perchannel.

• Fix re-quantize op fusion to be more generic.

• Fix per-channel quantization.

• Fix excluded ops and nodes.

• Fix the bugs for integration test.

• Clean up unused files

v0.1.5

This release fixes issues for requantize operator fusion for MiniGo. It also includes offline bias quantization in the settings of per-channel quantization of weights.

v0.1.0

Release 0.1.0

Overview

• Initial release of Intel AI Quantization Tools for TensorFlow
• Dockerfile to build quantization tools
• Script to launch quantization
• Quantize graph python script
• Test scripts for the quantization workflow

Graph Transforms

• fold_convolutionwithbias_mul
• fold_subdivmul_batch_norms
• fuse_quantized_conv_and_requantize
• mkl_fuse_pad_and_conv
• rerange_quantized_concat

Test Scripts

• faster_rcnn
• inceptionv3
• inceptionv4
• inception_resnet_v2
• rfcn
• resnet101
• resnet50
• ssd_mobilenet