Skip to content
This repository has been archived by the owner on Apr 17, 2023. It is now read-only.

Latest commit

 

History

History
86 lines (73 loc) · 3.56 KB

CHANGELOG.md

File metadata and controls

86 lines (73 loc) · 3.56 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

This project is based on mmsegmentation project by OpenMMLab. With respect to it we made the following changes.

[2022-05-09]

Added

  • HPO support
  • Feature dump support
  • NNCF support
  • Lite-HRNet-18-mod2 (middle)
  • Lite-HRNet-x-mod3 (heavy)

Removed

  • Tasks & model templates (moved to OTE)

[2021-12-27]

Added

  • Support of datasets: COCO Stuff, Kvasir-Seg, Kvasir-Instrument.
  • Implemented MaskCompose and ProbCompose composers to merge different augmentation pipelines.
  • Implemented augmentations: MixUp, CrossNorm.
  • Support of the pixel-weighting method to focus training on class borders.
  • Support of backbone architectures (including the appropriate config files): BiSeNet V2, CABiNet, DABNet, DDRNet, EfficientNet, ICNet, ShelfNet, STDCNet, Lite-HRNet.
  • Support of head architectures: BiSeHead, DDRHead, HamburgerHead, HyperSegHead, ICHead, MemoryHead, ShelfHead.
  • Support of EMA hook.
  • MMSegmentation can now use custom optimizer hook with Adaptive Gradient Clipping and custom learning rate hooks (cos, step) with support of three-stage training: freeze, warm-up and default.
  • Support of loss miners: ClassWeightingPixelSampler, LossMaxPooling.
  • Implemented AngularPWConv layer to support ML-based heads.
  • Implemented LocalContrastNormalization layer to normalize the input of a network.
  • Implemented loss factory which supports the following pixel-level losses: CrossEntropy, CrossEntropySmooth, NormalizedCrossEntropy, ReverseCrossEntropy, SymmetricCrossEntropy, ActivePassiveLoss.
  • Implemented Tversky and Boundary losses.
  • Implemented module to freeze the pattern-matched layers during training.
  • Export to InferenceEngine format which allow to run on edge-oriented devices.
  • Integration of NNCF model optimization.
  • Support of OTE tasks which allow to run the following commands through the API: train, eval, export, optimization.
  • Implemented scalar schedulers for ML-related scalar values (e.g. scale, regularization weight, loss weight): constant, step, poly.
  • Implemented script init_venv.sh to initialize the whole mmsegmentation-related environment.
  • Support of CPU-only training mode.

Changed

  • The following datasets have been updated to support MaskCompose augmentation: ADE20k, CHASE, Cityscapes, Drive, HRF, Stare, Pascal VOC12, Pascal VOC12 Aug.
  • Unified the head architectures: FCNHead, DepthwiseSeparableFCNHead.
  • Updated OCRHead to support depthwise separable convolutions.
  • Updated OHEM loss miner to support valid ratio hyperparameter.
  • Refactored the base network class to support a set of losses per head with adaptive loss re-weighting.
  • Refactored base loss class to support: loss-independent pixel miners, PR-product, MaxEntropy regularization, pixel-level losses re-weighting according to the weight mask, loss-jitter regularization.
  • Unified CrossEntropy and Dice losses.
  • Updated Dice loss to support General Dice and Dice++ losses.
  • Updated tools/export.py for the model export to support the implemented API-based export method.
  • Added support of fvcore in tools/get_flops.py tool.

Deprecated

  • TBD

Removed

  • TBD

Fixed

  • TBD

Security

  • TBD