Tutorial revamp 2026

[bo3z]

Initial restructuring of the tutorial, as discussed in the developer meetings.

The idea is to have 6 parts, each of which will have multiple notebooks. And users can select what "trajectory" to follow; e.g., Keras -> QKeras/HGQ or PyTorch. The README would have a nice overview figure showing a logical progression one could follow depending on the frontend, backend etc. The suggested structure is:

Part 1 - Getting started, where users can either choose Part 1a) and train a model with Keras or Part 1b) for PyTorch. Both are followed by Part 1c) for hls4ml conversion, synthesis. The idea is to eventually also support oneAPI in 1c). It should be reasonably straightforward - simply change the backend and the report parsing function.

Part 2 - Quantization; currently with QKeras v3 (which will hopefully be merged upstream soon) or Brevitas via QONNX (still has some bugs). HGQ needs to be updated to HGQ2 and added to this section. Once PQuantML is merged, a new notebook should be added. This part was bumped from Part 3, since quantization seems to be more important than profiling & tracing

Part 3 - Advanced config, showing reuse factor (part a) and profiling & tracing (part b). The idea is to split these up, as they are rather independent and users can then choose what interests them more (and it's also less confusion in live tutorials).  Eventually, we could maybe add another notebook on hardware-specific optimizations (e.g., reomving Softmax, da4ml etc.)

Part 4 - These would be "advanced" models. Initially, this would be the CNN example. However, we should also add the models from the Imperial tutorials (HGQ CNN, GNN, more advanced MLPs), as well as any others that could be interesting. This could be a mix of notebooks for users wanting to learn about specific architectures. Not done yet.

Part 5 - FPGA inference; these will be the accelerator backends, once merged

Part 6 - Other models; which, basically, means non-"neural network" architectures (BDT, SR)

As this is a draft, feedback is welcome.