Skip to content

Release 1.4.0

Choose a tag to compare

View all tags
@dylan-fan dylan-fan released this 15 May 02:32
v1.4.0
d64c73f
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Verified
Learn about vigilant mode.

By downloading, installing or using the software, you accept and agree to be bound by all of the terms and conditions of the LICENSE and DISCLAIMER.

Major Features and Improvements

FederatedML

  • Support Homo Secureboost
  • Support AIC/BIC-based Stepwise for Linear Models
  • Add Hetero Optimal Feature Binning, support iv/gini/chi-square/ks,and allow asymmetric binning methods
  • Interoperate with AI ecosystem: Add pytorch backend for Homo NN
  • Homo Framework factorization, simplify developing homo algorithms
  • Early stopping strategy for hetero algorithms.
  • Local Baseline supports multi-class classification
  • Add consistency check to Predict function
  • Optimize validation strategy，3x speed up in-training validation

FATE-Flow

  • Refactoring model management, native file directory storage, storage structure is more flexible, more information
  • Support model import and export, store and restore with reliable distributed system(Redis is currently supported)
  • Using MySQL instead of Redis to implement Job Queue, reducing system complexity
  • Support for uploading client local files
  • Automatically detects the existence of the table and provides the destroy option
  • Separate system, algorithm, scheduling command log, scheduling command log can be independently audited

Eggroll

Stability Boosts:

  • New resource management components introduce the brand new session mechanism. Processors can be cleaned up with a simple method call, even the session goes wrong.
  • Removes storage service. No C++ / native library compilation is needed.
  • Federated learning algorithms can still work at a 28% packet loss rate.

Performance Boosts:

  • Performances of federated learning algorithms are improved on Eggroll 2. Some algorithms get 10x performance boost.
  • Join interface is 16x faster than pyspark under federated learning scenarios.

User Experiences Boosts:

  • Quick deployment. Maven, pip, config and start.
  • Light dependencies. Check our requirements.txt / pom.xml and see.
  • Easy debugging. Necessary running contexts are provided. Runtime status are kept in log files and databases.
  • Few daemon processes. And they are all JVM applications.
Assets 2
Loading