PECOS v0.3.0

Highlights

• Enable distributed training for XLinear
• Enable PECOS for aarch64(arm64) CPU Architecture
• Enhance pecos.ann.hnsw with Function Multi-Versioning (FMV) technique to automatically select the best supported SIMD instructions (SSE, AVX2, AVX512) at runtime
• Reduce CPU memory usage in pecos.xmc.xtransformer training

Enhancements

• Add distilbert model. by @mo-fu in #97
• add CNAME by @jiong-zhang in #104
• Bump numpy from 1.20.3 to 1.21.0 in /examples/qp2q by @dependabot in #110
• enable Function Multi-Versioning (FMV) to support AVX512 by @rofuyu in #111
• Modify supported Python version by @weiliw-amz in #113
• Enabling PECOS for aarch64(arm64) CPU Architecture by @weiliw-amz in #114
• Update OpenBLAS Version for x86 Wheel Build by @weiliw-amz in #117
• SIMD Functions for aarch64(ARM64) by @weiliw-amz in #115
• Add profile_util module by @weiliw-amz in #121
• Fix FMV setup link flag and add test wheel CI by @weiliw-amz in #119
• Fix xlinear.reconstruct_model; Add PII embedding by @weiliw-amz in #120
• Add Distributed PECOS XLinear Modules by @weiliw-amz in #123
• Add distributed PECOS README by @weiliw-amz in #127
• update HNSW README and save/load in Python API by @OctoberChang in #129
• Improve XR-Transformer memory efficiency by @jiong-zhang in #128

Bug Fixes

• properly set Text2Text prediction argument by @OctoberChang in #101
• Fix HiearchicalMLModel pred-params initialization and add bugs by @weiliw-amz in #103
• minor bug fix in XR-Transformer exp script by @jiong-zhang in #106
• fixed multithreading bugs in py hierarchical kmeans by @OctoberChang in #108
• set pytest of hierarchical kmeans with single thread by @OctoberChang in #109
• Fix relative path in distributed README by @weiliw-amz in #130

Experiment Codes for Publications

• add overlap-clustering (Liu et al.) in NeurIPS21 by @xuanqing94 in #98
• add MACLR codes by @xyh97 in #100
• update experiment code for pecos jmlr paper by @OctoberChang in #107
• update Philip's experiment code into example folder by @OctoberChang in #118

New Contributors

• @mo-fu made their first contribution in #97
• @xuanqing94 made their first contribution in #98
• @xyh97 made their first contribution in #100
• @dependabot made their first contribution in #110

Full Changelog: v0.2.3...v0.3.0

PECOS v0.2.3

Bug Fixes

• Fix the index type during accessing sparse matrices [#96] (@hallogameboy)

PECOS v0.2.2

Bug Fixes

• Fix cost-sensitive bug in R normalization [#90] (@jiong-zhang )

PECOS v0.2.1

Highlights

• Remove support of Ubuntu 16.04
• Implemented XR-Transformer
• Enabled HNSW functionality
• Enabled cost-sensitive learning in PECOS

Enhancements

ANN HNSW

• Initial implementation of HNSW in C++ with single-thread [#44] (@OctoberChang)
• Refactor HNSW in C++ to support sparse/dense features and multi-threading [#49] (@rofuyu)
• Initial implementation of HNSW Python interface [#53] (@OctoberChang)
• Refactor HNSW python API and readme markdown [#63] (@OctoberChang)
• Refactor HNSW C++ to reuse priority queue for different inference calls within the same Searcher [#65] (@rofuyu)
• Enable HNSW save/load functionality [#71] (@OctoberChang)
• Add serialization version in HNSW save/load [#77] (@rofuyu)
• Enable HNSW python command line interface [#79] (@OctoberChang)

Cost-sensitive Learning

• Enable Cost-Sensitive Learning via XLinear API/CLI [#64] (@jiong-zhang)
• Enable cost sensitive for text2text CLI [#75] (@jiong-zhang)

XR-Transformer [#27, #64] (@jiong-zhang)

• Refactor pecos.xmc.xtransformer and enable end2end XR-Transformer training
• CLI tool for generating embeddings pecos.xmc.xtransformer.encode
• Faster transformer text tokenizers using huggingface's C implementation
• Allow training XR-Transformer without numerical features.

Better control over parameters for XLinear, XTransformer and Text2text [#64, #78, #80] (@jiong-zhang)

• Enable advanced control of parameters via JSON input file
• Add utility tool to generate parameter skeleton for further modification

Other new functionalities

• Added support for predicting on select outputs [#37, #43, #47] (@bhl00)
• Added new primal solver L2R_L2LOSS_SVC_PRIMAL for XLinear [#67] (@yuhchenlin)
• Add Makefile for easy format, install, clean and unittest. [#12] (@weiliw-amz)

Bug Fixes

• (#17) Fixed issues with github information obtaining when installing from .zip. [#21, #29] (@weiliw-amz)
• (#42) Fixed transformer training issue on single GPU [#14] (@jiong-zhang)
• Removed PECOS source-installation dependency on NumPy BLAS library. [#81] (@weili-amz)

Inital Release

Highlights

• Initial release.