Milestone to track future developments that are not planned yet

1. Project maintenance

1.1 CI

• Automatization of release pipeline (gitlab_shared_files#4)
• Automatization of fix pipeline (gitlab_shared_files#9)
• Refactor CI list_dependencies script (gitlab_shared_files#1)
• Selective run of the CI (gitlab_shared_files#10)

1.2 Unit test

• Better code coverage on aidge_core -> 60% (Python & C++)
• Use benchmarking tools for CI

2. Documentation

2.1 User documentation

• Create user guide pages:
  • Graph Representation (IR)
  • ONNX
  • Data/tensor
  • Registrar / Implementation
  • Scheduler
  • Learning
  • Benchmark
• Refactor home page
  • Add more figures with key points

2.2 Dev/wiki

• CheatSheet dev aidge

3 Features

3.1 Aidge Core

• Add a Aidge models zoo written with aidge API
• Extend benchmark with models, instead of only nodes

3.2 Quantization

• Mix precision
• LLM quantization
• Channel-Wise quantization for the PTQ

3.3 ONNX simplifier

• Benchmark vs Optimum, onnx_sim and onnx_slim
• Create Hugging face space

3.4 Hardware model

• Add first heterogeneous export using HWM

3.5 Exports

3.5.1 General

• Tag element wise operator and add a mecanism to "streamline" these operators
• Enable mixing export lib (for example arm + cpp)

3.5.2 Compilation

• Export compilation
• Use hardware model to constraint exploration
• Add user constraints

3.5.3 TensorRT

• Add support for Aidge quantization
• Refactor export for more recent version of TRT

3.5.4 Arm

• Support (back) for sub int8 kernels
• Export CMSISNN for model zoo

3.6 Interrop PyTorch

• Compat with cuda tensor
• Add unit test
• Enable multi I/O
• Update deprecated code

3.7 Model explorer

• Create a plugin to visualize aidge models

3.8 GGUF

• Add support to load an Aidge model from a GGUF file (aidge#297)
• Add support to save an Aidge model from a GGUF file (aidge#297)

4. Performances

4.1 DataLoading

• Add a custom ImageNet dataloader
• Experiment with python GIL to try to load in a thread data from a python custom database