Releases#

v1.6.1 (2Q24)#

Replace the default model for rotated_det/ins_seg task from resnet50_maskrcnn to efficientnetb2b_maskrcnn
Update pymongo version to 4.6.3 for resolving CVE-2024-21506
Use torchvision in MRCNN on CUDA
Update IPEX version in installation guide documentation
Update benchmark
Bump idan version to 3.7
Support benchmark history summary
Upgrade MAPI
Add NMS iou threshold configurable parameter
Remedy some medium/low severity bandit issues
Update documentations
Add perf benchmark test cases for action and visual prompting
Explicitly cast incorrect output type in OV model
Update QAT configs for rotated detection
Hotfix :wrench: Bypass ClsIncrSampler for tiling
[NNCF] Dynamic shape datasets WA
[Hotfix] :fire: Fixing detection oriented OV inferencer
Revert adaptive batch size
Fix e2e tests for XPU
Remove torch.xpu.optimize for semantic_segmentation task

Enable configurable confidence threshold for otx eval and export
Add YOLOX variants as new object detector models
Enable FeatureVectorHook to support action tasks
Add ONNX metadata to detection, instance segmantation, and segmentation models
Add a new feature to configure input size
Introduce the OTXSampler and AdaptiveRepeatDataHook to achieve faster training at the small data regime
Add a new object detector Lite-DINO
Add Semi-SL Mean Teacher algorithm for Instance Segmentation task
Official supports for YOLOX-X, YOLOX-L, YOLOX-S, ResNeXt101-ATSS
Add new argument to track resource usage in train command
Add Self-SL for semantic segmentation of SegNext families
Adapt input size automatically based on dataset statistics
Refine input data in-memory caching
Adapt timeout value of initialization for distributed training
Optimize data loading by merging load & resize operations w/ caching support for cls/det/iseg/sseg
Support torch==2.0.1
Set “Auto” as default input size mode

Support encrypted dataset training
Add custom max iou assigner to prevent CPU OOM when large annotations are used
Auto train type detection for Semi-SL, Self-SL and Incremental: “–train-type” now is optional
Add per-class XAI saliency maps for Mask R-CNN model
Add new object detector Deformable DETR
Add new object detector DINO
Add new visual prompting task
Add new object detector ResNeXt101-ATSS
Introduce channel_last parameter to improve the performance
Decrease time for making a workspace
Set persistent_workers and pin_memory as True in detection task
New algorithm for Semi-SL semantic segmentation based on metric learning via class prototypes
Self-SL for classification now can recieve just folder with any images to start contrastive pretraining
Update OpenVINO version to 2023.0, and NNCF verion to 2.5
Improve XAI saliency map generation for tiling detection and tiling instance segmentation
Remove CenterCrop from Classification test pipeline and editing missing docs link
Switch to PTQ for sseg
Minor bug fixes

Support direct annotation input for COCO format
Action task supports multi GPU training
Support storage cache in Apache Arrow using Datumaro for action tasks
Add a simplified greedy labels postprocessing for hierarchical classification
Support auto adapting batch size
Support auto adapting num_workers
Support noisy label detection for detection tasks
Make semantic segmentation OpenVINO models compatible with ModelAPI
Support label hierarchy through LabelTree in LabelSchema for classification task
Enhance exportable code file structure, video inference and default value for demo
Speedup OpenVINO inference in image classificaiton, semantic segmentation, object detection and instance segmentation tasks
Refactoring of ONNX export functionality
Minor bug fixes

Add generating feature cli_report.log in output for otx training
Support multiple python versions up to 3.10
Support export of onnx models
Add option to save images after inference in OTX CLI demo together with demo in exportable code
Support storage cache in Apache Arrow using Datumaro for cls, det, seg tasks
Add noisy label detection for multi-class classification task
Clean up and refactor the output of the OTX CLI
Enhance DetCon logic and SupCon for semantic segmentation
Detection task refactoring
Classification task refactoring
Extend OTX explain CLI
Segmentation task refactoring
Action task refactoring
Optimize data preprocessing time and enhance overall performance in semantic segmentation
Support automatic batch size decrease when there is no enough GPU memory
Minor bug fixes

Add FP16 IR export support
Add in-memory caching in dataloader
Add MoViNet template for action classification
Add Semi-SL multilabel classification algorithm
Integrate multi-gpu training for semi-supervised learning and self-supervised learning
Add train-type parameter to otx train
Add embedding of inference configuration to IR for classification
Enable VOC dataset in OTX
Add mmcls.VisionTransformer backbone support
Parametrize saliency maps dumping in export
Bring mmdeploy to action recognition model export & Test optimization of action tasks
Update backbone lists
Add explanation for XAI & minor doc fixes
Refactor phase#1: MPA modules

Installation through PyPI - Package will be renamed as OpenVINO™ Training Extensions
CLI update - Update otx find command to find configurations of tasks/algorithms - Introduce otx build command to customize task or model configurations - Automatic algorithm selection for the otx train command using the given input dataset
Adaptation of Datumaro component as a dataset interface