Release Release v1.4.0 · openvinotoolkit/training_extensions

Support encrypted dataset training (#2209)
Add custom max iou assigner to prevent CPU OOM when large annotations are used (#2228)
Auto train type detection for Semi-SL, Self-SL and Incremental: "--train-type" now is optional (#2195)
Add per-class XAI saliency maps for Mask R-CNN model (#2227)
Add new object detector Deformable DETR (#2249)
Add new object detector DINO (#2266)
Add new visual prompting task (#2203, #2274, #2311, #2354, #2318)
Add new object detector ResNeXt101-ATSS (#2309)

Introduce channel_last parameter to improve the performance (#2205)
Decrease time for making a workspace (#2223)
Set persistent_workers and pin_memory as True in detection task (#2224)
New algorithm for Semi-SL semantic segmentation based on metric learning via class prototypes (#2156)
Self-SL for classification now can recieve just folder with any images to start contrastive pretraining (#2219)
Update OpenVINO version to 2023.0, and NNCF verion to 2.5 (#2090)
Improve XAI saliency map generation for tiling detection and tiling instance segmentation (#2240)
Remove CenterCrop from Classification test pipeline and editing missing docs link(#2375)
Switch to PTQ for sseg (#2374)

Fix the bug that auto adapt batch size is unavailable with IterBasedRunner (#2182)
Fix the bug that learning rate isn't scaled when multi-GPU trianing is enabled(#2254)
Fix the bug that label order is misaligned when model is deployed from Geti (#2369)
Fix NNCF training on CPU (#2373)
Fix H-label classification (#2377)
Fix invalid import structures in otx.api (#2383)
Add for async inference calculating saliency maps from predictions (Mask RCNN IR) (#2395)

OpenVINO(==2023.0) IR inference is not working well on 2-stage models (e.g. Mask-RCNN) exported from torch==1.13.1

New Contributors

Full Changelog: v1.0.0...1.4.0