Skip to content

Latest commit

 

History

History
83 lines (74 loc) · 4.06 KB

CHANGELOG.md

File metadata and controls

83 lines (74 loc) · 4.06 KB

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.

[Unreleased]

Added

  • Changelog (this file).
  • Backbone models (ResNet variants, HRNEt, DarkNet53) under classification directory
  • Backbone training modules with support for mixup augmentation
  • V2 MomentumOptimizer for object detection framework
  • Support for Sync Batchnorm
  • Support for loading backbone models in SavedModel format
  • Support to freeze model variables in object detection framework via configuration
  • biases in detection framework get 2x the gradients (based on original detectron implementation)

Changed

  • Bugfix for ROIAlign layer
  • Sampling of anchor targets changed to include low quality matches as well
  • Weight initialization for FPN layer
  • Use L1 loss instead of Smooth L1 for FRCNN family of models
  • Cascade loss calculation update
  • Mask RCNN segmentation crop/resize fix
  • Switched to use crop and resize modification from TensorPack instead of tf.crop_and_resize
  • Object detection configuration system changed from flat definitions to heirarchical structure where individual sections can be overwritten as per training requirement
  • Horovod, MPI and NCCL options updated for train scripts and SM job launcher - this improves scaling efficiency in multinode settings over existing
  • Update README and posted top-1 accuracy for backbone training modules under classification directory
  • Support for CI service via custom configurations

[Unreleased]

Added

  • Changelog (this file).
  • BERT model.
  • Weights & Biases integration.
  • ELECTRA model.
  • Option in bbox target to return foreground assignments. Vector of indices of target within GT
  • Ability for eval hooks to automatically detect masks in runner
  • Mask target class that matches mask head output with GT masks
  • Option for coco dataset to return masks
  • Mask head and extractor options to faster RCNN
  • Mask loss
  • Mask head
  • Profiler hook
  • Mask rcnn configuration files
  • RetinaNet model
  • docs under vision/detection directory has README with results and setup instructions per model
  • Generic AWSDet tutorial
  • Ability to use Keras released backbone or custom SavedModel format backbone
  • Ability to resume complete training state for object detection trainer to restart training from a saved trainer state

Changed

  • SageMaker and Kubernetes Dockerfiles have been merged into one.
  • Use the module system rather than $PYTHONPATH, so jobs are launched with python -m albert.run_pretraining instead of python albert/run_pretraining.py.
  • NLP models use --per_gpu_batch_size instead of --batch_size.
  • NLP models use --squad_version instead of --task_name.
  • NLP models use --filesystem_prefix instead of --fsx_prefix. This option is mostly hidden from the user and should be a no-op.
  • NLP scripts start training at step 1 instead of step 0. So a job with --total_steps=100 will run steps [1..100] instead of [0..99].
  • NLP transformers dependency is now pinned to 2.11.0 instead of a custom fork. This removes the --pre_layer_norm=true option.
  • Removed the --pretrain_dataset argument, now pass directly --filesystem_prefix, --train_dir and --val_dir.
  • Background assignment in box target now uses while loop to handle rare case of too few backgrounds after initial duplicate assignment
  • Switched COCO utils segmentation assignment to use yxyx instead of xyxy
  • Matplotlib backend for visualization
  • Directory structure has changed for vision models
  • Per model configurations for EC2 and SageMaker have been introduced
  • Now we have a single training entrypoint for both EC2 and SM training jobs
  • Changes to core to support single stage detectors

Removed

  • NGC GPUMonitor Dockerfile.
  • Duplicate code for schedulers, sagemaker trainers

[0.2] - 2020-05-22

Added

  • ALBERT training scripts.
  • Draft of computer vision framework.

Changed

  • ResNet training scripts moved to /legacy.

[0.1] - 2020-05-01

Added

  • ResNet training scripts.