Hailo Model Zoo v2.13

Changelog

v2.13

• Update to use Dataflow Compiler v3.29.0 (developer-zone <https://hailo.ai/developer-zone/>_)

• Update to use HailoRT 4.19.0 (developer-zone <https://hailo.ai/developer-zone/>_)

• Using jit_compile which reduces dramatically the emulation inference time of the Hailo Model Zoo models.

• New tasks:

  • BEV: Multi-View 3D Object Detection

    • Added support for NuScenes dataset

    • Added PETRv2 with the following configuration:

      1. Backbone: RepVGG-B0 (800x320 input resolution)

      2. Transformer: 3 decoder layers, detection queries=304, replaced LN with UN

• New Models:

  • CAS-ViT <https://arxiv.org/pdf/2408.03703>_ - S, M, T - Convolutional-Attention based classification model
  • YOLOv10 <https://arxiv.org/pdf/2405.14458>_ - base, x-large - Latest YOLO detectors
  • CLIP <https://arxiv.org/pdf/2103.00020>_ Text Encoders - ResNet50x4, ViT-Large

• New retraining Docker containers for:

  • PETR - Multi-View 3D Object Detection

• Introduced new flags for hailomz CLI:

  • --ap-per-class for measuring average-precision per-class. Relevant for object detection and instance segmentation tasks.

• Bug fixes

Hailo Model Zoo v2.11

Changelog

v2.11

• Update to use Dataflow Compiler v3.27.0 (developer-zone <https://hailo.ai/developer-zone/>_)

• Update to use HailoRT 4.17.0 (developer-zone <https://hailo.ai/developer-zone/>_)

• New Models:

  • FastSAM-s - Zero-shot Instance Segmentation
  • Yolov9c - Latest Object Detection model of the YOLO family

• Using HailoRT-pp for postprocessing of the following variants:

  • nanodet

  Postprocessing JSON configurations are now part of the cfg directory.

• Introduced new flags for hailomz CLI:

  • --start-node-names and --end-node-names for customizing parsing behavior.
  • --classes for adjusting the number of classes in post-processing configuration.

  The --performance flag, previously utilized for compiling models with their enhanced model script if available, now offers an additional functionality.
  In instances where a model lacks an optimized model script, this flag triggers the compiler's Performance Mode to achieve the best performance

  These flags simplify the process of compiling models generated from our retrain dockers.

• Bug fixes

Hailo Model Zoo v2.10

v2.10

• Update to use Dataflow Compiler v3.26.0 (developer-zone)

• Update to use HailoRT 4.16.0 (developer-zone)

• Using HailoRT-pp for postprocessing of the following variants:

  • yolov8

• Porfiler change:

  • Removal of --mode flag from hailomz profile command, which generates a report according to provided HAR state.

• CLI change:

  • hailo8 target is deprecated in favor of hardware

• Support KITTI Stereo Dataset

• New Models:

  • vit_pose_small - encoder based transformer with layernorm for pose estimation

  • segformer_b0_bn - encoder based transformer with batchnorm for semantic segmentation

• Bug fixes

hotfix to v2.9

added KITTI Stereo dataset and fixed a few broken links

Hailo Model Zoo v2.9

• Update to use Dataflow Compiler v3.25.0 (developer-zone <https://hailo.ai/developer-zone/>_)

• Update to use HailoRT 4.15.0 (developer-zone <https://hailo.ai/developer-zone/>_)

• A new CLI-compatible API that allows users to incorporate format conversion and reshaping capabilities into the input:

• code-block:

  hailomz compile yolov5s --resize 1080 1920 --input-conversion nv12_to_rgb

• New transformer models added:

  • vit_pose_small_bn - encoder based transformer with batchnorm for pose estimation

  • clip_resnet_50x4 - Contrastive Language-Image Pre-Training for zero-shot classification

• New retraining dockers for vit variants using unified normalization.

• New Models:

  • yolov8s_pose / yolov8m_pose - pose estimation

  • scdepthv3 - depth-estimation

  • dncnn3 / dncnn_color_blind - image denoising

  • zero_dce_pp - low-light enhancement

  • stereonet - stereo depth estimation

• Using HailoRT-pp for postprocessing of the following models:

  • efficientdet_lite0 / efficientdet_lite1 / efficientdet_lite

Hailo Model Zoo v2.8

• Update to use Dataflow Compiler v3.24.0 (developer-zone)
• Update to use HailoRT 4.14.0 (developer-zone)
• The Hailo Model Zoo now supports the following vision transformers models:
  • vit_tiny / vit_small / vit_base - encoder based transformer with batchnorm for classification
  • detr_resnet_v1_18_bn - encoder/decoder transformer for object detection
  • clip_resnet_50 - Contrastive Language-Image Pre-Training for zero-shot classification
  • yolov5s_c3tr - object detection model with a MHSA block
• Using HailoRT-pp for postprocessing of the following variants:
  • yolov5
  • yolox
  • ssd
  • efficientdet
  • yolov7
• New Models:
  • repvgg_a1 / repvgg_a2 - classification
  • yolov8_seg: yolov8n_seg / yolov8s_seg / yolov8m_seg - instance segmentation
  • yolov6n_0.2.1 - object detection
  • zero_dce - low-light enhancement
• New retraining dockers for:
  • yolov8
  • yolov8_seg
• Enable compilation for hailo15h device
• Enable evaluation of models with RGBX / NV12 input format
• Bug fixes

Hailo Model Zoo v2.7

• Update to use Dataflow Compiler v3.23.0 (developer-zone)
• Updated to use HailoRT 4.13.0 (developer-zone)
• Inference flow was moved to new high-level APIs
• New object detection variants:
  • yolov8: yolov8n / yolov8s / yolov8m / yolov8l / yolov8x
  • damoyolo: damoyolo_tinynasL20_T / damoyolo_tinynasL25_S / damoyolo_tinynasL35_M
• New transformers based models:
  • vit_base - classification model
  • yolov5s_c3tr - object detection model with a self-attention block
• Examples for using HailoRT-pp - support for seamless integration of models and their corresponding postprocessing
  • yolov5m_hpp
• Configuration YAMLs and model-scripts for networks with YUY2 input format
• DAMO-YOLO retraining docker
• Bug fixes

Hailo Model Zoo v2.6

• Update to use Dataflow Compiler v3.22.0
• Updated to use HailoRT 4.12.0
• ViT (Vision Transformer) - new classification network with transformers-encoder based architecture
• New instance segmentation variants:
  • yolov5n_seg
  • yolov5s_seg
  • yolov5m_seg
• New object detecion variants for high resolution images:
  • yolov7e6
  • yolov5n6_6.1
  • yolov5s6_6.1
  • yolov5m6_6.1
• New flag --performance to reproduce highest performance for a subset of networks
• Hailo model-zoo log is now written into sdk_virtualenv/etc/hailo/modelzoo/hailo_examples.log
• Bug fixes

Hailo Model Zoo v2.5

• Update to use Dataflow Compiler v3.20.1 (developer-zone)
• Model scripts use new bgr to rgb conversion
• New Yolact variants - with all COCO classes:
  • yolact_regnetx_800mf
  • yolact_regnetx_1.6gf
• Bug fixes

Hailo Model Zoo v2.4

• Updated to use Dataflow Compiler v3.20 (developer-zone)
• Required FPS was moved from models YAML into the models scripts
• Model scripts use new change activation syntax
• New models:
  • Face Detection - scrfd_500m / scrfd_2.5g / scrfd_10g
• New tasks:
  • Super-Resolution
    • Added support for BSD100 dataset
    • The following models were added: espcn_x2 / espcn_x3 / espcn_x4
  • Face Recognition
    • Support for LFW dataset
    • The following models were added:
      • arcface_r50
      • arcface_mobilefacenet
    • Retraining docker for arcface architecture
• Added support for new hw-arch - hailo8l