Changes and new features in JeVois 1.22

JeVois 1.22 brings the following new features:

• New post-processors for YOLOv8 / v9/ v10 / 11, for: detection, semantic segmentation, pose/skeleton, and oriented bounding boxes. Also new post-processor for YOLOX networks.

• JeVois-Pro NPU: support for KSNN network conversion framework, which eliminates the need to specify input and output tensor shapes in the JeVois model zoo.

• JeVois-Pro NPU: one-line script to convert yolo-family networks for NPU:

  jevoispro-npu-convert.sh yolov8n-pose-1024x576-custom.onnx /path/to/sample-dataset.txt

  parses the model name to determine how to convert it and what post-processor to use, then converts and quantizes the model, and creates a weight file for NPU, a dynamic library to load the weights, and a JeVois zoo yaml file to describe the model. Just copy these 3 files to /jevoispro/share/snn/custom/ on microSD and you can run your model without further editing or configuring.

• JeVois-Pro: New docs and tutorials to convert YOLOv8-compatible networks for NPU and SPU (see Converting and running neural networks for JeVois-Pro NPU, Converting and running neural networks for Hailo-8 SPU, and http://jevois.org/tutorials/UserYolo.html )
• Faster dequantization and image resizing using parallel code.
• The Network class can now apply optional transforms to the output tensors:
  • dequant: dequantize the output tensors
  • outtransform: apply a semicolon-separated sequence (from left to right) of transforms to the outputs. Outputs are numbered starting at 0. Available transforms are:
    • shape(outnum, AxBxC...) to reshape an output to a given new shape. Does not change or move any tensor data. Total number of data elements must remain the same. Useful to squeeze/unsqueeze tensors.
    • transpose(outnum, oldaxisA, oldaxisB, ...) where transposed new axis 0 (the outermost dimension, typically batch size) will be from oldaxisA, new axis 1 from oldaxisB, etc. If outnum is *, transpose all outputs.
    • order(oldidx0, oldidx1, ...) where the new ordering of the output tensors will be: new tensor 0: old tensor oldidx0 (which is zero-based); new tensor 1: old tensor oldidx1, etc. It is ok to have duplicated or missing entries.
    • split(outnum, axis, newsize1, ..., newsizeN) where axis 0 is the outtermost dimension (typically, batch size), and newsize1 + ... + newsizeN must be equal to the original size of that axis. If outnum is *, split all outputs.
    • merge(axis, outnum1, ..., outnumN) where axis 0 is the outermost dimension (typically, batch size) and outnum1, ..., outnumN are the outputs to merge along that axis. All the outputs to be merged must have matching number of dimensions, and matching sizes on all other axes. The merged tensor will replace the first output listed in the merge, and the other listed will be removed. Outputs to merge must be listed in ascending order (use an order() transform first if needed) See the model zoo files in /jevoispro/share/dnn/ for examples. For instance:

      outtransform: "split(*,1,80,64); order(1,0,3,2,5,4); transpose(*,0,2,3,1)"

• JeVois-Pro: Updated Hailo libraries to 4.19, ImGui 1.91.2 which adds more graphics primitives for JeVois-Pro.
• JeVois-Pro: Updated to Ollama 0.3.14 with new qwen2.5:0.5b which is quite reasonable in our tests.
• JeVois-Pro: new Pipeline parameter to skip models marked as 'extramodel', to try to keep the list of models shorter.
• JeVois-Pro: slow SDL_Init() on host fixed by disabling game controllers
• JeVois-Pro: added access functions to latest postproc results, except yunet
• JeVois-Pro: Added fixes to deal with bogus always-NCHW tensors produced by the Khadas KSNN conversion tool for NPU
• Added log level '-1' for Timer, added timers for network extra inputs and dequantize (npu only for now)
• Miscellaneous bug fixes and performance improvements.
• Many doc updates.