Skip to content

v0.3.0

Choose a tag to compare

@github-actions github-actions released this 05 Oct 14:39
· 9 commits to main since this release

Train a detector for the OhmLab Neuro N6 from any labelled dataset and run
it on the board.

New

  • Training for the Neuro N6. The Train card trains YOLOX-Nano or
    YOLOX-Tiny (128 to 416 px square inputs) and exports an Arduino sketch
    folder: the int8 model, trained quantisation-aware so it scores like the
    float model, a labels header with a confidence threshold calibrated on
    the validation images, and a report. Validation is a per-class split or
    a second dataset; generated augmentations join the training set; a run
    resumes after an interruption or starts from an earlier run. With ST
    Edge AI Core installed, every export is compiled for the NPU and its fit
    reported.
  • On the board. The sketch streams its boxes to Neuro Studio, where
    the threshold is a live slider. It needs the Neuro N6 core 0.1.3 or
    later and says so on an older one. Measured on a Neuro N6: Nano runs in
    7.4 ms at 256x256 and Tiny in 31 ms at 320x320, and the boxes match
    PixelKit's int8 predictions on the same pictures to within a few pixels.
  • How the model fails. Per-class precision and recall, a confusion
    matrix, a gallery of the worst validation images, and every validation
    image beside its ground truth.
  • Label with your own model. A finished run pre-labels the dataset's
    unlabelled images; hand-edited images are never touched.
  • Background images. Pictures with nothing to label count as
    labelled and train as negatives.
  • Neuro Studio recordings. An .mjpeg recording from the board
    imports as its frames, trimmed and sampled.
  • More import formats. COCO and YOLO datasets import as editable
    projects, as Pascal VOC did.
  • Headless. pixelkit train and pixelkit label from a terminal,
    and the local API behind the Train card (docs/API.md).

Fixed

  • On a machine with one GPU, SAM 3 clicks and labelling jobs no longer
    wait for a training run to finish: the run pauses between steps for
    them.
  • Box-to-mask and click-to-detect find the object they are pointed at
    (99% of objects in our tests, from 55%): a faint whole-picture mask used
    to filter it out.
  • Augmented copies keep their boxes on the objects through rotation,
    perspective and lens distortion, and keep every box inside the
    picture.
  • The editor never loses edits or masks, and a painted mask keeps its
    size through save and reopen.
  • Imports and exports keep what they were given, and storage cannot lose
    or revert data.
  • The desktop app keeps a busy engine running, cleans up on quit, and
    updates its AI runtime when a new version needs new packages.
  • A finished training run gives its GPU memory back, and DINOv2 runs in
    half precision on any GPU chosen in Settings.

Notes

  • The installers are still unsigned, so SmartScreen and Gatekeeper warn
    on first launch.
  • An exported model is fine-tuned from Megvii's COCO-pretrained YOLOX
    checkpoint: its NOTICE.txt carries Megvii's Apache-2.0 notice, to keep
    with the model when you share it.

Full Changelog: v0.2.1...v0.3.0